To see the other types of publications on this topic, follow the link: Neural network: performance.
Dissertations / Theses on the topic 'Neural network: performance'
Create a spot-on reference in APA, MLA, Chicago, Harvard, and other styles
Consult the top 50 dissertations / theses for your research on the topic 'Neural network: performance.'
Next to every source in the list of references, there is an 'Add to bibliography' button. Press on it, and we will generate automatically the bibliographic reference to the chosen work in the citation style you need: APA, MLA, Harvard, Chicago, Vancouver, etc.
You can also download the full text of the academic publication as pdf and read online its abstract whenever available in the metadata.
Browse dissertations / theses on a wide variety of disciplines and organise your bibliography correctly.
1
Tupas, Ronald-Ray Tiñana. "Artificial neural network modelling of filtration performance." Thesis, National Library of Canada = Bibliothèque nationale du Canada, 2000. http://www.collectionscanada.ca/obj/s4/f2/dsk1/tape4/PQDD_0011/MQ59890.pdf.
Full text
2
Bataineh, Mohammad Hindi. "Artificial neural network for studying human performance." Thesis, University of Iowa, 2012. https://ir.uiowa.edu/etd/3259.
Full textAbstract:
The vast majority of products and processes in industry and academia require human interaction. Thus, digital human models (DHMs) are becoming critical for improved designs, injury prevention, and a better understanding of human behavior. Although many capabilities in the DHM field continue to mature, there are still many opportunities for improvement, especially with respect to posture- and motion-prediction. Thus, this thesis investigates the use of artificial neural network (ANN) for improving predictive capabilities and for better understanding how and why human behave the way they do.
With respect to motion prediction, one of the most challenging opportunities for improvement concerns computation speed. Especially, when considering dynamic motion prediction, the underlying optimization problems can be large and computationally complex. Even though the current optimization-based tools for predicting human posture are relatively fast and accurate and thus do not require as much improvement, posture prediction in general is a more tractable problem than motion prediction and can provide a test bead that can shed light on potential issues with motion prediction. Thus, we investigate the use of ANN with posture prediction in order to discover potential issues. In addition, directly using ANN with posture prediction provides a preliminary step towards using ANN to predict the most appropriate combination of performance measures (PMs) - what drives human behavior. The PMs, which are the cost functions that are minimized in the posture prediction problem, are typically selected manually depending on the task. This is perhaps the most significant impediment when using posture prediction. How does the user know which PMs should be used? Neural networks provide tools for solving this problem.
This thesis hypothesizes that the ANN can be trained to predict human motion quickly and accurately, to predict human posture (while considering external forces), and to determine the most appropriate combination of PM(s) for posture prediction. Such capabilities will in turn provide a new tool for studying human behavior. Based on initial experimentation, the general regression neural network (GRNN) was found to be the most effective type of ANN for DHM applications. A semi-automated methodology was developed to ease network construction, training and testing processes, and network parameters. This in turn facilitates use with DHM applications.
With regards to motion prediction, use of ANN was successful. The results showed that the calculation time was reduced from 1 to 40 minutes, to a fraction of a second without reducing accuracy. With regards to posture prediction, ANN was again found to be effective. However, potential issues with certain motion-prediction tasks were discovered and shed light on necessary future development with ANNs. Finally, a decision engine was developed using GRNN for automatically selecting four human PMs, and was shown to be very effective. In order to train this new approach, a novel optimization formulation was used to extract PM weights from pre-existing motion-capture data. Eventually, this work will lead to automatically and realistically driving predictive DHMs in a general virtual environment.
3
Alrumah, Muhammad K. "Neural networks predict well inflow performance." Texas A&M University, 2003. http://hdl.handle.net/1969.1/349.
Full textAbstract:
Predicting well inflow performance relationship accurately is very important for production engineers. From these predictions, future plans for handling and improving well performance can be established. One method of predicting well inflow performance is to use artificial neural networks.
Vogel's reference curve, which is produced from a series of simulation runs for a reservoir model proposed by Weller, is typically used to predict inflow performance relationship for solution-gas-drive reservoirs. In this study, I reproduced Vogel's work, but instead of producing one curve by conventional regression, I built three neural network models. Two models predict the IPR efficiently with higher overall accuracy than Vogel's reference curve.
4
Chen, Dong. "Neural network model for predicting performance of projects." Thesis, National Library of Canada = Bibliothèque nationale du Canada, 1999. http://www.collectionscanada.ca/obj/s4/f2/dsk1/tape9/PQDD_0021/MQ48059.pdf.
Full text
5
Schilling, Glenn D. "Modeling Aircraft Fuel Consumption with a Neural Network." Thesis, Virginia Tech, 1997. http://hdl.handle.net/10919/36533.
Full textAbstract:
This research involves the development of an aircraft fuel consumption model to simplify Bela Collins of the MITRE Corporation aircraft fuelburn model in terms of level of computation and level of capability. MATLAB and its accompanying Neural Network Toolbox, has been applied to data from the base model to predict fuel consumption. The approach to the base model and neural network is detailed in this paper. It derives from the basic concepts of energy balance. Multivariate curve fitting techniques used in conjunction with aircraft performance data derive the aircraft specific constants. Aircraft performance limits are represented by empirical relationships that also utilize aircraft specific constants. It is based on generally known assumptions and approximations for commercial jet operations. It will simulate fuel consumption by adaptation of a specific aircraft using constants that represent the relationship of lift-to-drag and thrust-to-fuel flow.
The neural network model invokes the output from MITRE1s algorithm and provides: (1) a comparison to the polynomial fuelburn function in the fuelburn post- processor of the FAA Airport and Airspace Simulation Model (SIMMOD), (2) an established sensitivity of system performance for a range of variables that effect fuel consumption, (3) a comparison of post fuel burn (fuel consumption algorithms) techniques to new techniques, and (4) the development of a trained demo neural network.
With the powerful features of optimization, graphics, and hierarchical modeling, the MATLAB toolboxes proved to be effective in this modeling process.Master of Science
6
Rosenfeld, Jonathan S. (Jonathan Shmuel). "On the relation between neural network size and performance." Thesis, Massachusetts Institute of Technology, 2019. https://hdl.handle.net/1721.1/122703.
Full textAbstract:
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.Thesis: S.M., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2019
Cataloged from student-submitted PDF version of thesis.
Includes bibliographical references (pages 57-58).
Artificial Neural Networks (NN) are notorious for their size requirements and for the effort involved in developing well performing network models. This thesis uncovers a fundamental relationship that ties model size and performance in a predictable manner. This relationship enables a well-founded development of networks at small scale while producing insight into their large-scale behavior.
by Jonathan S. Rosenfeld.
S.M.
S.M. Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science
7
Mitchell, David. "Classification by Neural Network and Statistical Models in Tandem: Does Integration Enhance Performance?" Thesis, University of North Texas, 1998. https://digital.library.unt.edu/ark:/67531/metadc278874/.
Full textAbstract:
The major purposes of the current research are twofold. The first purpose is to present a composite approach to the general classification problem by using outputs from various parametric statistical procedures and neural networks. The second purpose is to compare several parametric and neural network models on a transportation planning related classification problem and five simulated classification problems.
8
Mamidanna, Pranav. "Optimizing Neural Source Extraction Algorithms: A Performance Measure Based on Neuronal Network Properties." Thesis, KTH, Numerisk analys, NA, 2017. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-210052.
Full textAbstract:
Extracting neural activity from electrophysiological and calcium All existing automated algorithms for this purpose, however, rely heavily on manual intervention and parameter tuning. In this thesis, we introduce a novel performance measure based on well-founded notions of neuronal network organization. This enables us to systematically tune parameters, using techniques from statistical design of experiments and response surface methods. We implement this framework on an algorithm used to extract neural activity from microendoscopic calcium imaging datasets, and demonstrate that this greatly reduces manual intervention.Extraktion av neuronal aktivitet från elektrofysiologiska och kalciumavbildningsmätningar utgör ett viktigt problem inom neurovetenskapen. Alla existerande automatiska algoritmer för detta ändamål beror dock i dagsläget på manuell handpåläggning och parameterinställning. I detta examensarbete presenterar vi ett nytt prestandamått baserat på välgrundade begrepp rörande organisationen av neuronala nätverk. Detta möjliggör en systematisk parameterinställning genom att använda tekniker från statistisk experimentdesign och response surface-metoder. Vi har implementerat detta ramverk för en algoritm som används för att extrahera neuronal aktivitet från mikroendoskopisk kalciumavbildningsdata och visar att detta förfarande avsevärt minskar behovet av manuell inblandning.
9
Nichols, Roger Alan. "A performance baseline for machinery condition classification by neural network." Master's thesis, This resource online, 1993. http://scholar.lib.vt.edu/theses/available/etd-03172010-020117/.
Full text
10
Lin, Yu Chu. "E-government website performance evaluation based on BP neural network." Thesis, University of Macau, 2017. http://umaclib3.umac.mo/record=b3691489.
Full text
11
Gil, Ferrer Alejandro. "A neural network performance analysis with three different model structures." Thesis, KTH, Skolan för elektroteknik och datavetenskap (EECS), 2021. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-302144.
Full textAbstract:
This report analyzes three neural network structures: dense, convolutional and recurrent. One data set example and problem has been chosen for each type of structure: a multi-class classification problem, an image classifier and a time sequence prediction, respectively. This report also aims at understanding which structure performs better for different problem statements, and how the different parameters they depend on affect their performance. The most common parameters that have been analyzed are the following: the number of intermediate layers, the number of neurons, the number of epochs, the batch size, the activation function, the loss function and the optimizer. The results showed that a dense structure has high dependencies between the values of its internal operations. Hence, the average execution time for CPU and GPU are similar. However, accelerated algorithms for GPUs made a substantial difference for convolutional and recurrent structures in comparison to CPU launches. Furthermore, the results of each model showed that most of their attributes vary the performance of the model during training, obtaining a combination of values that are suitable for each structure.Detta arbete analyserar tre strukturtyper av neurala nätverk: dense, convolutional och recurrent. För varje strukturtyp har en problemtyp valts och ett dataset tilldelats dem. Dessa är: flerklass-klassificering, bildklassificering, och förutsägelse för tidssekvenser. Arbetet ämnar även att ta reda på vilken av dessa strukturer som ger bäst resultat för olika problemformuleringar, och hur förändringar av de parametrar som de beror på påverkar resultatet. De parametrar som analyserats är: antalet mellanlager, antalet neuroner, antalet träningscykler, urvalsstorleken, aktiveringsfuntionen, förlustfunktionen, och optimering. Resultaten visade att det i typen dense är hög beroendegrad mellan värdena av dess interna beräkningar. Som följd av detta är den genomsnittliga beräkningstiden ungefär densamma på både CPU och GPU. Däremot gjorde det stor skillnad på beräkningstiden för både convolutional och recurrent när acceleratorer för GPU användes på dem, då detta var signifikant snabbare än det var på CPU. Utöver detta framgick även att för varje modell så hade majoriteten av deras attribut påverkan på prestandan under träningsmomentet. Detta visade på en kombination av värden som var lämpliga att använda för respektive struktur.
12
Kopel, Ariel. "NEURAL NETWORKS PERFORMANCE AND STRUCTURE OPTIMIZATION USING GENETIC ALGORITHMS." DigitalCommons@CalPoly, 2012. https://digitalcommons.calpoly.edu/theses/840.
Full textAbstract:
Artificial Neural networks have found many applications in various fields such as function approximation, time-series prediction, and adaptive control. The performance of a neural network depends on many factors, including the network structure, the selection of activation functions, the learning rate of the training algorithm, and initial synaptic weight values, etc.
Genetic algorithms are inspired by Charles Darwin’s theory of natural selection (“survival of the fittest”). They are heuristic search techniques that are based on aspects of natural evolution, such as inheritance, mutation, selection, and crossover.
This research utilizes a genetic algorithm to optimize multi-layer feedforward neural network performance and structure. The goal is to minimize both the function of output errors and the number of connections of network. The algorithm is modeled in C++ and tested on several different data sets. Computer simulation results show that the proposed algorithm can successfully determine the appropriate network size for optimal performance. This research also includes studies of the effects of population size, crossover type, probability of bit mutation, and the error scaling factor.
13
Ekpenyong, Frank Udo. "An investigation into automatic road network update using trajectory data and performance-guided neural network." Thesis, University of East London, 2010. http://roar.uel.ac.uk/2603/.
Full textAbstract:
This research aims to categorise road network recorded trajectory data using Artificial Neural Network (ANN) such that the travelled road class can be revealed. This would inform on the feasibility of implementing an automated road update system that would rely on user recorded trajectory data to automate the discovery, classification, and update of candidate road network segments to existing road network database. End-users of digital GIS road network database are increasingly the major source of road change error reports. At present, vendors of digital road network database only provide web forms for user to report road errors. To investigate these errors they travel such roads and analyse satellite images to register changes. However, the major limitations to this method are that it is time consuming and logistically challenging to visit all locations of reported road error. Also the accuracy of road user road error report depends on the user's interpretation of the road network representation offered on the device in relation to the road in the real world, and the user's geographic knowledge and familiarity of the area. In the literature, different solutions have been proposed to deal with the key road update functions road change detection, representation and update. But most of these approaches are exclusively tied to remote sensing images. While these methods of road updating have been successfully used to extract roads from images, their accuracy is directly tied to the quality of the images and object model used for road extraction. Hence, existing solutions are image-specific and cannot be applied to other image type obtained from another sensor without significant adjustments of the parameters. An alternative approach investigated in this thesis uses the trajectory of moving vehicles to automate the detection of new roads and thus update a road network database. GPS recorded trajectory data were collected during field tests from a range of road types. The trajectory data are an abstraction of the road segments travelled and this study assumes for the sake of experimentation that these road segments are not present in the GIS road coverage and seeks to group the GPS-based trajectory data using an ANN to reveal the presence and class of public thoroughfares. This will establish the extent to which drive characteristics naturally fall into road feature classes. The results suggests that from the ANNs investigated, the unsupervised Snap-Drift Neural Network (SDNN) and the supervised Snap-Drift Adaptive Function Neural Network (SADFUNN) have the potential to support vehicle trajectory similarity grouping (classification) that can inform whether the road feature travelled is a new road feature that needs to be added to existing road database. The Probabilistic Neural Network (PNN) and Radial Basis Function (RBF) neural network also offered good classification performance.
14
Watcharapichat, Pijika. "Improving the performance of dataflow systems for deep neural network training." Thesis, Imperial College London, 2017. http://hdl.handle.net/10044/1/57955.
Full textAbstract:
Deep neural networks (DNNs) have led to significant advancements in machine learning. With deep structure and flexible model parameterisation, they exhibit state-of-the-art accuracies for many complex tasks e.g. image recognition. To achieve this, models are trained iteratively over large datasets. This process involves expensive matrix operations, making it time-consuming to obtain converged models. To accelerate training, dataflow systems parallelise computation. A scalable approach is to use parameter server framework: it has workers that train model replicas in parallel and parameter servers that synchronise the replicas to ensure the convergence. With distributed DNN systems, there are three challenges that determine the training completion time. In this thesis, we propose practical and effective techniques to address each of these challenges. Since frequent model synchronisation results in high network utilisation, the parameter server approach can suffer from network bottlenecks, thus requiring decisions on resource allocation. Our idea is to use all available network bandwidth and synchronise subject to the available bandwidth. We present Ako, a DNN system that uses partial gradient exchange for synchronising replicas in a peer-to-peer fashion. We show that our technique exhibits a 25% lower convergence time than a hand-tuned parameter-server deployments. For a long training, the compute efficiency of worker nodes is important. We argue that processing hardware should be fully utilised for the best speed-up. The key observation is it is possible to overlap the execution of several matrix operations with other workloads. We describe Crossbow, a GPU-based system that maximises hardware utilisation. By using a multi-streaming scheduler, multiple models are trained in parallel on GPU and achieve a 2.3x speed-up compared to a state-of-the-art system. The choice of model configuration for replicas also directly determines convergence quality. Dataflow systems are used for exploring the promising configurations but provide little support for efficient exploratory workflows. We present Meta-dataflow (MDF), a dataflow model that expresses complex workflows. By taking into account all configurations as a unified workflow, MDFs efficiently reduce time spent on configuration exploration.
15
林楠林 and Nanlin Lin. "A neural-network approach to high-performance adaptive control for robot manipulators." Thesis, The University of Hong Kong (Pokfulam, Hong Kong), 1998. http://hub.hku.hk/bib/B31237411.
Full text
16
Lin, Nanlin. "A neural-network approach to high-performance adaptive control for robot manipulators /." Hong Kong : University of Hong Kong, 1998. http://sunzi.lib.hku.hk/hkuto/record.jsp?B19852265.
Full text
17
Ibrani, Lavdrus. "High performance dynamic control of two-axes system." Thesis, Leeds Beckett University, 1999. http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.285927.
Full text
18
Kaster, Joshua M. "Training Convolutional Neural Network Classifiers Using Simultaneous Scaled Supercomputing." University of Dayton / OhioLINK, 2020. http://rave.ohiolink.edu/etdc/view?acc_num=dayton1588973772607826.
Full text
19
Oheda, Hakim. "Artificial neural network control strategies for fuel cell hybrid system." Thesis, Cranfield University, 2013. http://dspace.lib.cranfield.ac.uk/handle/1826/7964.
Full textAbstract:
The greening of air transport is the driver for developing technologies to reduce the
environmental impact of aviation with the aim of halving the amount of carbon dioxide
(COଶ) emitted by air transport, cutting specific emissions of nitrogen oxides (NO୶) by 80%
and halving perceived noise by the year 2020. Fuel Cells (FC) play an important role in the
new power generation field as inherently clean, efficient and reliable source of power
especially when comparing with the traditional fossil-fuel based technologies.
The project investigates the feasibility of using an electric hybrid system consisting of a fuel
cell and battery to power a small model aircraft (PiperCub J3). In order to meet the desired
power requirements at different phases of flight efficiently, a simulation model of the
complete system was first developed, consisting of a Proton Exchange Membrane hybrid fuel
cell system, 6DoF aircraft model and neural network based controller. The system was then
integrated in one simulation environment to run in real-time and finally was also tested in
hardware-in-the-loop with real-time control.
The control strategy developed is based on a neural network model identification technique;
specifically Model Reference Control (MRC), since neural network is well suited to nonlinear
systems. To meet the power demands at different phases of flight, the controller controls the
battery current and rate of charging/discharging.
Three case studies were used to validate and assess the performance of the hybrid system:
battery fully charged (high SOC), worst case scenario and taking into account the external
factors such as wind speeds and wind direction. In addition, the performance of the Artificial
Neural Network Controller was compared to that of a Fuzzy Logic controller. In all cases the
fuel cell act as the main power source for the PiperCub J3 aircraft. The tests were carried-out
in both simulation and hardware-in-the-loop.
20
Thissen-Roe, Anne. "Adaptive selection of personality items to inform a neural network predicting job performance /." Thesis, Connect to this title online; UW restricted, 2005. http://hdl.handle.net/1773/9138.
Full text
21
Hansen, Vedal Amund. "Comparing performance of convolutional neural network models on a novel car classification task." Thesis, KTH, Medieteknik och interaktionsdesign, MID, 2017. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-213468.
Full textAbstract:
Recent neural network advances have lead to models that can be used for a variety of image classification tasks, useful for many of today’s media technology applications. In this paper, I train hallmark neural network architectures on a newly collected vehicle image dataset to do both coarse- and fine-grained classification of vehicle type. The results show that the neural networks can learn to distinguish both between many very different and between a few very similar classes, reaching accuracies of 50.8% accuracy on 28 classes and 61.5% in the most challenging 5, despite noisy images and labeling of the dataset.Nya neurala nätverksframsteg har lett till modeller som kan användas för en mängd olika bildklasseringsuppgifter, och är därför användbara många av dagens medietekniska applikationer. I detta projektet tränar jag moderna neurala nätverksarkitekturer på en nyuppsamlad bilbild-datasats för att göra både grov- och finkornad klassificering av fordonstyp. Resultaten visar att neurala nätverk kan lära sig att skilja mellan många mycket olika bilklasser, och även mellan några mycket liknande klasser. Mina bästa modeller nådde 50,8% träffsäkerhet vid 28 klasser och 61,5% på de mest utmanande 5, trots brusiga bilder och manuell klassificering av datasetet.
22
Samarnggoon, Keattikorn. "Modelling of human control and performance evaluation using artificial neural network and brainwave." Thesis, Staffordshire University, 2016. http://eprints.staffs.ac.uk/2389/.
Full textAbstract:
Conventionally, a human has to learn to operate a machine by himself/herself. Human Adaptive Mechatronics (HAM) aims to investigate a machine that has the capability to learn its operator skills in order to provide assistance and guidance appropriately. Therefore, the understanding of human behaviour during the human-machine interaction (HMI) from the machine’s side is essential. The focus of this research is to propose a model of human-machine control strategy and performance evaluation from the machine’s point of view. Various HAM simulation scenarios are developed for the investigations of the HMI. The first case study that utilises the classic pendulum-driven capsule system reveals that a human can learn to control the unfamiliar system and summarise the control strategy as a set of rules. Further investigation of the case study is conducted with nine participants to explore the performance differences and control characteristics among them. High performers tend to control the pendulum at high frequency in the right portion of the angle range while the low performers perform inconsistent control behaviour. This control information is used to develop a human-machine control model by adopting an Artificial Neural Network (ANN) and 10-time- 10-fold cross-validation. Two models of capsule direction and position predictions are obtained with 88.3% and 79.1% accuracies, respectively. An Electroencephalogram (EEG) headset is integrated into the platform for monitoring brain activity during HMI. A number of preliminary studies reveal that the brain has a specific response pattern to particular stimuli compared to normal brainwaves. A novel human-machine performance evaluation based on the EEG brainwaves is developed by utilising a classical target hitting task as a case study of HMI. Six models are obtained for the evaluation of the corresponding performance aspects including the Fitts index of performance. The averaged evaluation accuracy of the models is 72.35%. However, the accuracy drops to 65.81% when the models are applied to unseen data. In general, it can be claimed that the accuracy is satisfactory since it is very challenging to evaluate the HMI performance based only on the EEG brainwave activity.
23
Song, Philip, and André Brogärd. "Performance Analysis of Various Activation Functions Using LSTM Neural Network For Movie Recommendation Systems." Thesis, KTH, Skolan för elektroteknik och datavetenskap (EECS), 2020. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-280451.
Full textAbstract:
The growth of importance and popularity of recommendations system has increased in many various areas. This thesis focuses on recommendation systems for movies. Recurrent neural networks using LSTM blocks have shown some success for movie recommendation systems. Research has indicated that by changing activation functions in LSTM blocks, the performance, measured as accuracy in predictions, can be improved. In this study we compare four different activation functions (hyperbolic tangent, sigmoid, ELU and SELU activation functions) used in LSTM blocks, and how they impact the prediction accuracy of the neural networks. Specifically, they are applied to the block input and the block output of the LSTM blocks. Our results indicate that the hyperbolic tangent, which is the default, and sigmoid function perform about the same, whereas the ELU and SELU functions perform worse. Further research is needed to identify other activation functions that could improve the prediction accuracy and improve certain aspects of our methodology.Rekommendationssystem har ökat i betydelse och popularitet i många olika områden. Denna avhandling fokuserar på rekommendationssystem för filmer. Recurrent neurala nätverk med LSTM blocks har visat viss framgång för rekommendationssystem för filmer. Tidigare forskning har indikerat att en ändring av aktiverings funktioner har resulterat i förbättrad prediktering. I denna studie jämför vi fyra olika aktiveringsfunktioner (hyperbolic tangent, sigmoid, ELU and SELU) som appliceras i LSTM blocks och hur de påverkar predikteringen i det neurala nätverket. De appliceras specifikt på block input och block output av LSTM blocken. Våra resultat indikerar att den hyperboliska tangentfunktionen, som är standardvalet, och sigmoid funktionen presterar lika, men ELU och SELU presterar båda sämre. Ytterligare forskning krävs för att indentifiera andra aktiveringsfunktioner och för att förbättra flera delar av metodologin.
24
Chen, Weiliang. "The performance of associative memory models with biologically inspired connectivity." Thesis, University of Hertfordshire, 2009. http://hdl.handle.net/2299/3102.
Full textAbstract:
This thesis is concerned with one important question in artificial neural networks, that is, how biologically inspired connectivity of a network affects its associative memory performance. In recent years, research on the mammalian cerebral cortex, which has the main responsibility for the associative memory function in the brains, suggests that the connectivity of this cortical network is far from fully connected, which is commonly assumed in traditional associative memory models. It is found to be a sparse network with interesting connectivity characteristics such as the “small world network” characteristics, represented by short Mean Path Length, high Clustering Coefficient, and high Global and Local Efficiency. Most of the networks in this thesis are therefore sparsely connected. There is, however, no conclusive evidence of how these different connectivity characteristics affect the associative memory performance of a network. This thesis addresses this question using networks with different types of connectivity, which are inspired from biological evidences. The findings of this programme are unexpected and important. Results show that the performance of a non-spiking associative memory model is found to be predicted by its linear correlation with the Clustering Coefficient of the network, regardless of the detailed connectivity patterns. This is particularly important because the Clustering Coefficient is a static measure of one aspect of connectivity, whilst the associative memory performance reflects the result of a complex dynamic process. On the other hand, this research reveals that improvements in the performance of a network do not necessarily directly rely on an increase in the network’s wiring cost. Therefore it is possible to construct networks with high associative memory performance but relatively low wiring cost. Particularly, Gaussian distributed connectivity in a network is found to achieve the best performance with the lowest wiring cost, in all examined connectivity models. Our results from this programme also suggest that a modular network with an appropriate configuration of Gaussian distributed connectivity, both internal to each module and across modules, can perform nearly as well as the Gaussian distributed non-modular network. Finally, a comparison between non-spiking and spiking associative memory models suggests that in terms of associative memory performance, the implication of connectivity seems to transcend the details of the actual neural models, that is, whether they are spiking or non-spiking neurons.
25
Caplan, Jonathan Stuart. "Temperature Scaling in Pyloric Networks| A Computational Study of a Small Neural Network Oscillator and the Effects of Ion Channel Temperature Dependences on Network Performance." Thesis, Brandeis University, 2013. http://pqdtopen.proquest.com/#viewpdf?dispub=3596761.
Full textAbstract:
Neurons in poikilotherms must operate over the animal's natural temperature range if they are to survive. The effects of temperature on various cellular processes can vary dramatically, which suggests that it may be difficult to design a circuit that behaves consistently over a temperature range. Previous work in the crab Cancer borealis (Tang et al., 2010, 2012) showed that the pyloric rhythm of the stomatogastric ganglion (STG) maintains its bursting duty cycle and phase relationships over a temperature range of 7 to 23 °C. Rinberg et al., 2013 also observed this phase invariance over a temperature range in the three cell pyloric pacemaker kernel.
To explore the effects of temperature on this system, we implemented a computational model of the STG pacemaker kernel (Soto-Treviño et al., 2005), that simulates two electrically coupled cells and includes temperature dependences, represented as Q10's. Separate Q10's were assigned for maximal conductance, rate of activation and inactivation. We also assigned a Q10 for the buffering rate of intracellular Ca2+. All Q10's were selected randomly from 1 to 4, except the maximal conductance Q10's that were set to 1.6. Maximal conductance values at the reference temperature of 11 °C were initially set to the values selected by Soto-Treviño et al., 2005. Each model was run over a range of 7 to 23 °C.
While some Q10 values, such as those for mKCa, mKd and Ca2+ buffering are critical for appropriate temperature scaling, the system is only moderately sensitive to others such as hNa, CaT and CaS and largely insensitive to Q10 values for slower conductances such as A, Nap, IMI and leak.
Overall, we find that robust neuronal behavior can be achieved over a temperature range within a subset of Q10-space. Within our model, certain Q10's are tightly constrained while others can be chosen over a relatively wide range. This provides insight into the relative contribution of different ionic conductances to high-level neuronal dynamics.
26
Munipalli, Sirish Kumar. "An FPGA Implementation of a High Performance AER Packet Network." PDXScholar, 2013. https://pdxscholar.library.pdx.edu/open_access_etds/639.
Full textAbstract:
This thesis presents a design to route the spikes in a cognitive computing project called Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE). SyNAPSE is a DARPA-funded program to develop electronic neuromorphic ma- chine technology that scales to biological levels. The basic computational block in the SyNAPSE system is the asynchronous spike processor (ASP) chip. This analog core contains the neurons and synapses in a neural fabric and performs the neural and synaptic computations.An ASP takes asynchronous pulses (spikes) as inputs and after some small delay produces asyn- chronous pulses as outputs.The ASP chips are organized in a nxn (where n [approximately equal to] 10) 2-dimensional grid with a dedicated node for each chip. This interconnected network is called Digital Fabric(DF) and the node is called Digital Fabric Node (DFN). The DF is a packet network that routes pulse (AER - Address event rep- resentation) packets between ASP's. This thesis also presents a technique for design implementation on a FPGA, perfor- mance testing of the network and validation of the network using various tools.
27
Fischer, Manfred M., and Sucharita Gopal. "Neural Network Models and Interregional Telephone Traffic. Comparative Performance Comparisons between Multilayer Feedforward Networks and the Conventional Spatial Interaction Model." WU Vienna University of Economics and Business, 1992. http://epub.wu.ac.at/4206/1/WSG_DP_2792.pdf.
Full text
28
Rönnholm, Niklas. "A study of limitations and performance in scalable hosting using mobile devices." Thesis, KTH, Skolan för elektroteknik och datavetenskap (EECS), 2018. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-224646.
Full textAbstract:
At present day, distributed computing is a widely used technique, where volunteers support different computing power needs organizations might have. This thesis sought to benchmark distributed computing performance limited to mobile device support since this type of support is seldom done with mobile devices. This thesis proposes two approaches to harnessing computational power and infrastructure of a group of mobile devices. The problems used for benchmarking are small instances of deep learning training. One requirement posed by the mobile devices’ non-static nature was that this should be possible without any significant prior configuration. The protocol used for communication was HTTP. The reason deep-learning was chosen as the benchmarking problem is due to its versatility and variability. The results showed that this technique can be applied successfully to some types of problem instances, and that the two proposed approaches also favour different problem instances. The highest request rate found for the prototype with a 99% response rate was a 2100% increase in efficiency compared to a regular server. This was under the premise that it was provided just below 2000 mobile devices for only particular problem instances.För närvarande är distribuerad databehandling en utbredd teknik, där frivilliga individer stödjer olika organisationers behov av datorkraft. Denna rapport försökte jämföra prestandan för distribuerad databehandling begränsad till enbart stöd av mobila enheter då denna typ av stöd sällan görs med mobila enheter. Rapporten föreslår två sätt att utnyttja beräkningskraft och infrastruktur för en grupp mobila enheter. De problem som används för benchmarking är små exempel på deep-learning. Ett krav som ställdes av mobilenheternas icke-statiska natur var att detta skulle vara möjligt utan några betydande konfigureringar. Protokollet som användes för kommunikation var HTTP. Anledningen till att deeplearning valdes som referensproblem beror på dess mångsidighet och variation. Resultaten visade att denna teknik kan tillämpas framgångsrikt på vissa typer av probleminstanser, och att de två föreslagna tillvägagångssätten också gynnar olika probleminstanser. Den högsta requesthastigheten hittad för prototypen med 99% svarsfrekvens var en 2100% ökning av effektiviteten jämfört med en vanlig server. Detta givet strax under 2000 mobila enheter för vissa speciella probleminstanser.
29
Björk, Gustav, and Alexander Wester. "A Deep Neural Network Approach for Intersection Testing of Two 3D Meshes." Thesis, Blekinge Tekniska Högskola, Institutionen för datavetenskap, 2020. http://urn.kb.se/resolve?urn=urn:nbn:se:bth-19623.
Full textAbstract:
Background. Neural Networks have mainly been used in behavior and gameplayrelated areas in games, but they have not yet been used specifically for intersection testing. This thesis explores the possibility to use deep neural networks for intersection testing of two 3D meshes. Objectives. The main goal of the thesis is to train a Deep Neural Network that can be used to replace traditional intersection test algorithms by having similar accuracy and a faster execution time. Methods. The research methods used in this thesis are implementation and experimentation. The deep neural network is trained using TensorFlow. Two different mesh generation techniques are implemented, one generating heightmaps and one generating planets. The two mesh types are combined to test all combinations of generated meshes. Attempts to make the network as general as possible are done through importance sampling to expose the network to tricky situations. A test application is developed where the intersection testing can be performed and compared to the Separating Axis Theorem (SAT). Heatmaps are also created to see how accurate the network is. Results. The results show that the network is accurate at classifying intersection between meshes similar to the ones it trained on. However, the network lacks generality and has bad accuracy if new meshes are introduced. The measured execution times show that the trained Deep Neural Network is 15.6 times as fast as a singlethreaded implementation of the SAT and 2.3 times as fast as the multi-threaded SAT. Conclusions. The trained network can be used as an early exit intersection test before using more expensive algorithms. The faster intersection testing can be useful in game physics by allowing faster classification of which meshes need to be tested for collisions. However, the main outcome is the shown potential for future work in the area including training a more general network, allowing variable mesh sizes, and providing information for solving collision responses.Bakgrund. Neurala Nätverk har främst använts för beteende- och spelmekanikrelaterade områden inom spel, men de har ännu inte använts för genomskärningstester. Det här examensarbetet utförskar möjligheten att använda djupinlärning för att utföra genomskärningstester mellan två tredimensionella spelobjekt. Syfte. Huvudmålet med det här examensarbetet är att träna ett djupinlärt neuralt nätverk som kan ersätta traditionella genomskärningstestalgoritmer genom likvärdig precision och snabbare exekveringstid. Metod. Forskningsmetoderna som användes under examensarbetet är implementation och experimentation. Det djupinlärda neurala nätverket tränas med TensorFlow. Två olika spelobjektsgenereringsmetoder implementeras, där den ena genererar heightmaps och den andra genererar planeter. De två objekttyperna kombineras så att alla kombinationer av spelobjekt kan testas. För att göra nätverket så generellt som möjligt används importance sampling som utsätter nätverket för svåra situationer. Ett testprogram utvecklas där genomskärningstester kan utföras och jämföras mot Separating Axis Theorem (SAT). Grafer av typen heatmaps skapas också för att visa hur hög precision nätverket har. Resultat. Resultaten visar att nätverket har hög precision vid klassificering av spelobjekt liknande de som den tidigare har tränat på. Nätverket har sämre precision när nya spelobjekt introduceras. De uppmätta exekveringstiderna visar att det neurala nätverket är 15.6 gånger så snabbt som singeltrådade implementationen av SAT och 2.3 gånger så snabbt som den flertrådade SAT-implementationen. Slutsatser. Det tränade nätverket kan användas som ett tidig avbrott innan en dyrare algoritm används. Den snabbare genomskärningstestningen kan vara användbar i spelfysik eftersom den tillåter snabbare klassificering av vilka spelobjekt som behöver testas för kollision. Det huvudsakliga utfallet är den visade potentialen för vidare forskning inom området vilket inkluderar träning av ett mer generellt nätverk, möjlighet att variera spelobjektens storlek samt ge information för att kunna lösa kollisioner.
30
Ericsson, Andreas, and Kana Filip Döringer. "Convolutional Neural Networks for Classification of Metastatic Tissue in Lymph Nodes : How Does Cutout Affect the Performance of Convolutional Neural Networks for Biomedical Image Classification?" Thesis, KTH, Skolan för elektroteknik och datavetenskap (EECS), 2021. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-302529.
Full textAbstract:
One of every eight women will in their lifetime suffer from breast cancer, making it the most common type of cancer for women. A successful treatment is very much dependent on identifying metastatic tissue which is cancer found beyond the initial tumour. Using deep learning within biomedical analysis has become an effective approach. However, its success is very dependent on large datasets. Data augmentation is a way to enhance datasets without requiring more annotated data. One way of doing this is using the cutout method which masks parts of an input image. Our research focused on investigating how the cutout method could improve the performance of Convolutional Neural Networks for classifying metastatic tissue on the Patch Camelyon dataset. Our research showed that improvements in performance can be achieved by using the cutout method. Further, our research suggests that using a non label- preserving version of cutout is better than a label- preserving version. The most improvement in accuracy was seen when we used a randomly sized cutout mask. The experiment resulted in an increase in accuracy by 3.6%, from the baseline of 82,3% to 85.9%. The cutout method was also compared- and used in conjunction with other well- established data augmentation techniques. Our conclusion is that cutout can be a competitive form of data augmentation that can be used both with and without other data augmentation techniques.Var åttonde kvinna drabbas under sin livstid av bröstcancer. Detta gör det till den vanligaste formen av cancer för kvinnor. En framgångsrik behandling är beroende av att kunna identifiera metastatisk vävnad, vilket är cancer som spridit sig bortom den ursprungliga tumören. Att använda djupinlärning inom biomedicinsk analys har blivit en effektiv metod. Dock är dess framgång väldigt beroende av stora datamängder. Dataförstärkning är olika sätt att förbättra en mängd data som inte innebär att addera ytterligare annoterad data. Ett sätt att göra detta är genom den en metod som kallas Cutout som maskar en del av en bild. Vår studie undersöker hur Cutout påverkar resultatet när Convolutional Neural Networks klassificerar huruvida bilder från datasetet Patch Camelyon innehåler metastaser eller inte. Vår studie visar att användandet av Cutout kan innebära förbättringar i resultatet. Dessutom tyder vår studie på att resultatet förbättras än mer om även delen av bilden som kan innehålla metastaser kan maskas ut. Den största förbättringen i resultatet var när maskningen var av varierande storlek från bild till bild. Resultatet förbättrades från 82.3% korrekta klassifikationer utan någon dataförstärkning till 85.9% med den bästa versionen av Cutout. Cutout jämfördes också, och användas tillsammans med, andra väletablerade dataförstärkningsmetoder. Vår slutsats är att Cutout är en dataförstärkningsmetod med potentital att vara användbar såväl med som utan andra dataförstärkningsmetoder.
31
Chamberlain, Matthew. "Novel control of a high performance rotary wood planing machine." Thesis, Loughborough University, 2013. https://dspace.lboro.ac.uk/2134/12261.
Full textAbstract:
Rotary planing, and moulding, machining operations have been employed within the woodworking industry for a number of years. Due to the rotational nature of the machining process, cuttermarks, in the form of waves, are created on the machined timber surface. It is the nature of these cuttermarks that determine the surface quality of the machined timber. It has been established that cutting tool inaccuracies and vibrations are a prime factor in the form of the cuttermarks on the timber surface. A principal aim of this thesis is to create a control architecture that is suitable for the adaptive operation of a wood planing machine in order to improve the surface quality of the machined timber. In order to improve the surface quality, a thorough understanding of the principals of wood planing is required. These principals are stated within this thesis and the ability to manipulate the rotary wood planing process, in order to achieve a higher surface quality, is shown. An existing test rig facility is utilised within this thesis, however upgrades to facilitate higher cutting and feed speeds, as well as possible future implementations such as extended cutting regimes, the test rig has been modified and enlarged. This test rig allows for the dynamic positioning of the centre of rotation of the cutterhead during a cutting operation through the use of piezo electric actuators, with a displacement range of ±15μm. A new controller for the system has been generated. Within this controller are a number of tuneable parameters. It was found that these parameters were dependant on a high number external factors, such as operating speeds and run‐out of the cutting knives. A novel approach to the generation of these parameters has been developed and implemented within the overall system. Both cutterhead inaccuracies and vibrations can be overcome, to some degree, by the vertical displacement of the cutterhead. However a crucial information element is not known, the particular displacement profile. Therefore a novel approach, consisting of a subtle change to the displacement profile and then a pattern matching approach, has been implemented onto the test rig. Within the pattern matching approach the surface profiles are simplified to a basic form. This basic form allows for a much simplified approach to the pattern matching whilst producing a result suitable for the subtle change approach. In order to compress the data levels a Principal Component Analysis was performed on the measured surface data. Patterns were found to be present in the resultant data matrix and so investigations into defect classification techniques have been carried out using both K‐Nearest Neighbour techniques and Neural Networks. The application of these novel approaches has yielded a higher system performance, for no additional cost to the mechanical components of the wood planing machine, both in terms of wood throughput and machined timber surface quality.
32
Fischer, Manfred M., and Petra Staufer-Steinnocher. "Optimization in an Error Backpropagation Neural Network Environment with a Performance Test on a Pattern Classification Problem." WU Vienna University of Economics and Business, 1998. http://epub.wu.ac.at/4150/1/WSG_DP_6298.pdf.
Full textAbstract:
Various techniques of optimizing the multiple class cross-entropy error function
to train single hidden layer neural network classifiers with softmax output transfer
functions are investigated on a real-world multispectral pixel-by-pixel classification
problem that is of fundamental importance in remote sensing. These techniques
include epoch-based and batch versions of backpropagation of gradient descent,
PR-conjugate gradient and BFGS quasi-Newton errors. The method of choice
depends upon the nature of the learning task and whether one wants to optimize
learning for speed or generalization performance. It was found that, comparatively
considered, gradient descent error backpropagation provided the best and most stable
out-of-sample performance results across batch and epoch-based modes of operation.
If the goal is to maximize learning speed and a sacrifice in generalisation is acceptable,
then PR-conjugate gradient error backpropagation tends to be superior. If the
training set is very large, stochastic epoch-based versions of local optimizers should
be chosen utilizing a larger rather than a smaller epoch size to avoid inacceptable
instabilities in the generalization results. (authors' abstract)Series: Discussion Papers of the Institute for Economic Geography and GIScience
33
Baudin, Lastra Tomas. "Performance based diagnostics of a twin shaft aeroderivative gas turbine: water wash scheduling." Thesis, Cranfield University, 2015. http://dspace.lib.cranfield.ac.uk/handle/1826/10003.
Full textAbstract:
Aeroderivative gas turbines are used all over the world for different applications
as Combined Heat and Power (CHP), Oil and Gas, ship propulsion and others.
They combine flexibility with high efficiencies, low weight and small footprint,
making them attractive where power density is paramount as off shore Oil and
Gas or ship propulsion. In Western Europe they are widely used in CHP small
and medium applications thanks to their maintainability and efficiency. Reliability,
Availability and Performance are key parameters when considering plant
operation and maintenance. The accurate diagnose of Performance is
fundamental for the plant economics and maintenance planning. There has been
a lot of work around units like the LM2500® , a gas generator with an
aerodynamically coupled gas turbine, but nothing has been found by the author
for the LM6000® .
Water wash, both on line or off line, is an important maintenance practice
impacting Reliability, Availability and Performance. This Thesis aims to select and
apply a suitable diagnostic technique to help establishing the schedule for off line
water wash on a specific model of this engine type. After a revision of Diagnostic
Methods Artificial Neural Network (ANN) has been chosen as diagnostic tool.
There was no WebEngine model available of the unit under study so the first step
of setting the tool has been creating it. The last step has been testing of ANN as
a suitable diagnostic tool. Several have been configured, trained and tested and
one has been chosen based on its slightly better response. Finally, conclusions
are discussed and recommendations for further work laid out.
34
Malmgren, Henrik. "Revision of an artificial neural network enabling industrial sorting." Thesis, Uppsala universitet, Institutionen för teknikvetenskaper, 2019. http://urn.kb.se/resolve?urn=urn:nbn:se:uu:diva-392690.
Full textAbstract:
Convolutional artificial neural networks can be applied for image-based object classification to inform automated actions, such as handling of objects on a production line. The present thesis describes theoretical background for creating a classifier and explores the effects of introducing a set of relatively recent techniques to an existing ensemble of classifiers in use for an industrial sorting system.The findings indicate that it's important to use spatial variety dropout regularization for high resolution image inputs, and use an optimizer configuration with good convergence properties. The findings also demonstrate examples of ensemble classifiers being effectively consolidated into unified models using the distillation technique. An analogue arrangement with optimization against multiple output targets, incorporating additional information, showed accuracy gains comparable to ensembling. For use of the classifier on test data with statistics different than those of the dataset, results indicate that augmentation of the input data during classifier creation helps performance, but would, in the current case, likely need to be guided by information about the distribution shift to have sufficiently positive impact to enable a practical application. I suggest, for future development, updated architectures, automated hyperparameter search and leveraging the bountiful unlabeled data potentially available from production lines.
35
Sidebo, Edvin. "Charged particle distributions and robustness of the neural network pixel clustering in ATLAS." Licentiate thesis, KTH, Partikel- och astropartikelfysik, 2016. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-190858.
Full textAbstract:
This thesis contains a study of the robustness of the artificial neural network used in the ATLAS track reconstruction algorithm as a tool to recover tracks in dense environments. Different variations, motivated by potential discrepancies between data and simulation, are performed to the neural network’s input while monitoring the corresponding change in the output. Within reasonable variation magnitudes, the neural networks prove to be robust to most variations. In addition, a measurement of charged particle distributions is summarised. This is one of the first such measurements carried out for proton-proton colli- sions at √s = 13 TeV, limited to a phase space defined by transverse momentum pT > 100 MeV and absolute pseudorapidity |η| < 2.5. Tracks are corrected for de- tector inefficiencies and unfolded to particle-level. The result is compared to the prediction of different models. Overall, the EPOS and Pythia 8 A2 models show the best agreement with the data.Spår från elektriskt laddade partiklar rekonstrueras i ATLAS genom att kombinera mätningar från de innersta subdetektorerna. I de extrema miljöer som skapas i proton-proton-kollisionerna i Large Hadron Collider vid CERN är det av yttersta vikt att algoritmen för att rekonstruera spår är högpresterande. Uppgiften är särskilt svår i partikelrika miljöer där flera partiklar färdas nära varandra, åtskilda av avstånd jämförbara med storleken på detektorns utläsningselement. Ett artificiellt neuralt nätverk används i algoritmen för att klassificera mätdata från pixeldetektorn, belägen närmast interaktionspunkten, för att lyckas identifiera spår i partikelrika miljöer som annars hade gått förlorade. I denna avhandling utreds det neurala nätverkets stabilitet. Dess känslighet studeras genom att manuellt manipulera dess indata och därefter utvärdera dess resultat. Nätverket tränas med simulerad data. Variationerna i indata är utformade för att undersöka skillnader mellan data och simulering, orsakade av osäkerheter i simuleringsmodellen eller osäkerheter i pixeldetektorns kalibrering. Av de undersökta variationerna har en osäkerhet i skalan eller utläsningströskeln för pixeldetektorns kalibrering den största effekten på nätverkets resultat. Andra variationer har en betydligt mindre påverkan. Avhandlingen presenterar också en studie av distributioner av elektriskt laddade partiklar producerade i proton-proton-kollisioner. Det är en av de första studierna av partikeldistributioner för Large Hadron Colliders andra körning med mass-centrum-energi √s = 13 TeV. Mätningen är begränsad till fasrymden definierad av en transversell rörelsemängd pT > 100 MeV, och absolut rapiditet |η| < 2.5. Spår av partiklar rekonstrueras och korrigeras för detektorns ineffektiviteter för att presenteras på partikelnivå. Dessa jämförs sedan med förutsägelser från olika modeller. Modellerna EPOS och Pythia 8 A2 är generellt de som bäst överensstämmer med data. Författaren har undersökt partiklar som migrerar in och ut ur fasrymden. Andelen spår associerade till partiklar som migrerat utifrån uppskattas med simulerad data, till som mest 10% nära fasrymdens gränser. Osäkerheten på denna andel uppskattas till att vara som mest 4.5%, huvudsakligen orsakad av osäkerheten på mängden material i de innersta subdetektorerna.
QC 20160817
36
Wang, Boqian. "High-Performance Network-on-Chip Design for Many-Core Processors." Licentiate thesis, KTH, Elektronik och inbyggda system, 2020. http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-283517.
Full textAbstract:
With the development of on-chip manufacturing technologies and the requirements of high-performance computing, the core count is growing quickly in Chip Multi/Many-core Processors (CMPs) and Multiprocessor System-on-Chip (MPSoC) to support larger scale parallel execution. Network-on-Chip (NoC) has become the de facto solution for CMPs and MPSoCs in addressing the communication challenge. In the thesis, we tackle a few key problems facing high-performance NoC designs. For general-purpose CMPs, we encompass a full system perspective to design high-performance NoC for multi-threaded programs. By exploring the cache coherence under the whole system scenario, we present a smart communication service called Advance Virtual Channel Reservation (AVCR) to provide a highway to target packets, which can greatly reduce their contention delay in NoC. AVCR takes advantage of the fact that we can know or predict the destination of some packets ahead of their arrival at the Network Interface (NI). Exploiting the time interval before a packet is ready, AVCR establishes an end-to-end highway from the source NI to the destination NI. This highway is built up by reserving the Virtual Channel (VC) resources ahead of the target packet transmission and offering priority service to flits in the reserved VC in the wormhole router, which can avoid the target packets’ VC allocation and switch arbitration delay. Besides, we also propose an admission control method in NoC with a centralized Artificial Neural Network (ANN) admission controller, which can improve system performance by predicting the most appropriate injection rate of each node using the network performance information. In the online control process, a data preprocessing unit is applied to simplify the ANN architecture and make the prediction results more accurate. Based on the preprocessed information, the ANN predictor determines the control strategy and broadcasts it to each node where the admission control will be applied. For application-specific MPSoCs, we focus on developing high-performance NoC and NI compatible with the common AMBA AXI4 interconnect protocol. To offer the possibility of utilizing the AXI4 based processors and peripherals in the on-chip network based system, we propose a whole system architecture solution to make the AXI4 protocol compatible with the NoC based communication interconnect in the many-core system. Due to possible out-of-order transmission in the NoC interconnect, which conflicts with the ordering requirements specified by the AXI4 protocol, in the first place, we especially focus on the design of the transaction ordering units, realizing a high-performance and low cost solution to the ordering requirements. The microarchitectures and the functionalities of the transaction ordering units are also described and explained in detail for ease of implementation. Then, we focus on the NI and the Quality of Service (QoS) support in NoC. In our design, the NI is proposed to make the NoC architecture independent from the AXI4 protocol via message format conversion between the AXI4 signal format and the packet format, offering high flexibility to the NoC design. The NoC based communication architecture is designed to support high-performance multiple QoS schemes. The NoC system contains Time Division Multiplexing (TDM) and VC subnetworks to apply multiple QoS schemes to AXI4 signals with different QoS tags and the NI is responsible for traffic distribution between two subnetworks. Besides, a QoS inheritance mechanism is applied in the slave-side NI to support QoS during packets’ round-trip transfer in NoC.Med utvecklingen av tillverkningsteknologi av on-chip och kraven på högpresterande da-toranläggning växer kärnantalet snabbt i Chip Multi/Many-core Processors (CMPs) ochMultiprocessor Systems-on-Chip (MPSoCs) för att stödja större parallellkörning. Network-on-Chip (NoC) har blivit den de facto lösningen för CMP:er och MPSoC:er för att mötakommunikationsutmaningen. I uppsatsen tar vi upp några viktiga problem med hög-presterande NoC-konstruktioner.Allmänna CMP:er omfattas ett fullständigt systemperspektiv för att design högprester-ande NoC för flertrådad program. Genom att utforska cachekoherensen under hela system-scenariot presenterar vi en smart kommunikationstjänst, AVCR (Advance Virtual ChannelReservation) för att tillhandahålla en motorväg till målpaket, vilket i hög grad kan min-ska deras förseningar i NoC. AVCR utnyttjar det faktum att vi kan veta eller förutsägadestinationen för vissa paket före deras ankomst till nätverksgränssnittet (Network inter-face, NI). Genom att utnyttja tidsintervallet innan ett paket är klart, etablerar AVCRen ände till ände motorväg från källan NI till destinationen NI. Denna motorväg byggsupp genom att reservera virtuell kanal (Virtual Channel, VC) resurser före målpaket-söverföringen och erbjuda prioriterade tjänster till flisar i den reserverade VC i wormholerouter. Dessutom föreslår vi också en tillträdeskontrollmetod i NoC med en centraliseradartificiellt neuronät (Artificial Neural Network, ANN) tillträdeskontroll, som kan förbättrasystemets prestanda genom att förutsäga den mest lämpliga injektionshastigheten för varjenod via nätverksprestationsinformationen. I onlinekontrollprocessen används en förbehan-dlingsenhet på data för att förenkla ANN-arkitekturen och göra förutsägningsresultatenmer korrekta. Baserat på den förbehandlade informationen bestämmer ANN-prediktornkontrollstrategin och sänder den till varje nod där tillträdeskontrollen kommer att tilläm-pas.För applikationsspecifika MPSoC:er fokuserar vi på att utveckla högpresterande NoCoch NI kompatibla med det gemensamma AMBA AXI4 protokoll. För att erbjuda möj-ligheten att använda AXI4-baserade processorer och kringutrustning i det on-chip baseradenätverkssystemet föreslår vi en hel systemarkitekturlösning för att göra AXI4 protokolletkompatibelt med den NoC-baserade kommunikation i det multikärnsystemet. På grundav den out-of-order överföring i NoC, som strider mot ordningskraven som anges i AXI4-protokollet, fokuserar vi i första hand på utformningen av transaktionsordningsenheterna,för att förverkliga en hög prestanda och låg kostnad-lösning på ordningskraven. Sedanfokuserar vi på NI och Quality of Service (QoS)-stödet i NoC. I vår design föreslås NI attgöra NoC-arkitekturen oberoende av AXI4-protokollet via meddelandeformatkonverteringmellan AXI4 signalformatet och paketformatet, vilket erbjuder NoC-designen hög flexi-bilitet. Den NoC-baserade kommunikationsarkitekturen är utformad för att stödja fleraQoS-schema med hög prestanda. NoC-systemet innehåller Time-Division Multiplexing(TDM) och VC-subnät för att tillämpa flera QoS-scheman på AXI4-signaler med olikaQoS-taggar och NI ansvarar för trafikdistribution mellan två subnät. Dessutom tillämpasen QoS-arvsmekanism i slav-sidan NI för att stödja QoS under paketets tur-returöverföringiNoC
QC 20201008
37
Nilsson, Kristian, and Hans-Eric Jönsson. "A comparison of image and object level annotation performance of image recognition cloud services and custom Convolutional Neural Network models." Thesis, Blekinge Tekniska Högskola, Institutionen för programvaruteknik, 2019. http://urn.kb.se/resolve?urn=urn:nbn:se:bth-18074.
Full textAbstract:
Recent advancements in machine learning has contributed to an explosive growth of the image recognition field. Simultaneously, multiple Information Technology (IT) service providers such as Google and Amazon have embraced cloud solutions and software as a service. These factors have helped mature many computer vision tasks from scientific curiosity to practical applications. As image recognition is now accessible to the general developer community, a need arises for a comparison of its capabilities, and what can be gained from choosing a cloud service over a custom implementation. This thesis empirically studies the performance of five general image recognition services (Google Cloud Vision, Microsoft Computer Vision, IBM Watson, Clarifai and Amazon Rekognition) and image recognition models of the Convolutional Neural Network (CNN) architecture that we ourselves have configured and trained. Image and object level annotations of images extracted from different datasets were tested, both in their original state and after being subjected to one of the following six types of distortions: brightness, color, compression, contrast, blurriness and rotation. The output labels and confidence scores were compared to the ground truth of multiple levels of concepts, such as food, soup and clam chowder. The results show that out of the services tested, there is currently no clear top performer over all categories and they all have some variations and similarities in their output, but on average Google Cloud Vision performs the best by a small margin. The services are all adept at identifying high level concepts such as food and most mid-level ones such as soup. However, in terms of further specifics, such as clam chowder, they start to vary, some performing better than others in different categories. Amazon was found to be the most capable at identifying multiple unique objects within the same image, on the chosen dataset. Additionally, it was found that by using synonyms of the ground truth labels, performance increased as the semantic gap between our expectations and the actual output from the services was narrowed. The services all showed vulnerability to image distortions, especially compression, blurriness and rotation. The custom models all performed noticeably worse, around half as well compared to the cloud services, possibly due to the difference in training data standards. The best model, configured with three convolutional layers, 128 nodes and a layer density of two, reached an average performance of almost 0.2 or 20%. In conclusion, if one is limited by a lack of experience with machine learning, computational resources and time, it is recommended to make use of one of the cloud services to reach a more acceptable performance level. Which to choose depends on the intended application, as the services perform differently in certain categories. The services are all vulnerable to multiple image distortions, potentially allowing adversarial attacks. Finally, there is definitely room for improvement in regards to the performance of these services and the computer vision field as a whole.
38
Geng, Tong. "FPGA-based high-performance neural network acceleration." Thesis, 2021. https://hdl.handle.net/2144/41887.
Full textAbstract:
In the last ten years, Artificial Intelligence through Deep Neural Networks (DNNs) has penetrated virtually every aspect of science, technology, and business. Advances are rapid with thousands of papers being published annually. Many types of DNNs have been and continue to be developed -- in this thesis, we address Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Graph Neural Networks (GNNs) -- each with a different set of target applications and implementation challenges. The overall problem for all of these Neural Networks (NNs) is that their target applications generally pose stringent constraints on latency and throughput, but also have strict accuracy requirements. Much research has therefore gone into all aspects of improving NN quality and performance: algorithms, code optimization, acceleration with GPUs, and acceleration with hardware, both dedicated ASICs and off-the-shelf FPGAs. In this thesis, we concentrate on the last of these approaches.
There have been many previous efforts in creating hardware to accelerate NNs. The problem designers face is that optimal NN models typically have significant irregularities, making them hardware unfriendly. One commonly used approach is to train NN models to follow regular computation and data patterns. This approach, however, can hurt the models' accuracy or lead to models with non-negligible redundancies. This dissertation takes a different approach. Instead of regularizing the model, we create architectures friendly to irregular models. Our thesis is that high-accuracy and high-performance NN inference and training can be achieved by creating a series of novel irregularity-aware architectures for Field-Programmable Gate Arrays (FPGAs). In four different studies on four different NN types, we find that this approach results in speedups of 2.1x to 3255x compared with carefully selected prior art; for inference, there is no change in accuracy.
The bulk of this dissertation revolves around these studies, the various workload balancing techniques, and the resulting NN acceleration architectures. In particular, we propose four different architectures to handle, respectively, data structure level, operation level, bit level, and model level irregularities.
At the data structure level, we propose AWB-GCN, which uses runtime workload rebalancing to handle Sparse Matrices Multiplications (SpMM) on extremely sparse and unbalanced input. With GNN inference as a case study, AWB-GCN achieves over 90% system efficiency, guarantees efficient off-chip memory access, and provides considerable speedups over CPUs (3255x), GPUs (80x), and a prior ASIC accelerator (5.1x).
At the operation level, we propose O3BNN-R, which can detect redundant operations and prune them at run time. This works even for those that are highly data-dependent and unpredictable. With Binarized NNs (BNNs) as a case study, O3BNN-R can prune over 30% of the operations, without any accuracy loss, yielding speedups over state-of-the-art implementations on CPUs (1122x), GPUs (2.3x), and FPGAs (2.1x).
At the bit level, we propose CQNN. CQNN embeds a Coarse-Grained Reconfigurable Architecture (CGRA) which can be programmed at runtime to support NN functions with various data-width requirements. Results show that CQNN can deliver us-level Quantized NN (QNN) inference.
At the model level, we propose FPDeep, especially for training. In order to address model-level irregularity, FPDeep uses a novel model partitioning schemes to balance workload and storage among nodes. By using a hybrid of model and layer parallelism to train DNNs, FPDeep avoids the large gap that commonly occurs between training and testing accuracy due to the improper convergence to sharp minimizers (caused by large training batches). Results show that FPDeep provides scalable, fast, and accurate training and leads to 6.6x higher energy efficiency than GPUs.
39
Tryphonas, Marinos. "Modeling Hedge Fund Performance Using Neural Network Models." Thesis, 2012. http://hdl.handle.net/1807/32497.
Full textAbstract:
Hedge fund performance is modeled from publically available data using feed-forward neural networks trained using a resilient backpropagation algorithm. The neural network’s performance is then compared with linear regression models. Additionally, a stepwise factor regression approach is introduced to reduce the number of inputs supplied to the models in order to increase precision.
Three main conclusions are drawn: (1) neural networks effectively model hedge fund returns, illustrating the strong non-linear relationships between the economic risk factors and hedge fund performance, (2) while the group of 25risk factors we draw variables from are used to explain hedge fund performance, the best model performance is achieved using different subsets of the 25 risk factors, and, (3) out-of-sample model performance degrades across the time during the recent (and still on-going) financial crisis compared to less volatile time periods, indicating the models’ inability to predict severely volatile economic scenarios such as economic crises.
40
Lin, Yu-Cchen, and 林昱辰. "Performance Forecasting of Heat Pipe Using Neural network." Thesis, 2010. http://ndltd.ncl.edu.tw/handle/71589682570022402965.
Full textAbstract:
碩士聖約翰科技大學
自動化及機電整合研究所
98
In this study, silver nano-fluid was utilized to be the working fluid on grooved heat pipes. The silver nano-particles in two different particle sizes (10nm and 35nm) were used in this experiment. Also, combined with nano-fluid in five different kinds of concentration—1ppm, 5ppm, 10ppm, 50ppm, and 100ppm, the 39 sets of thermal resistance values in this experiment were completed under different heating power—30W, 40W, 50W, and 60W. Besides, the suite program “Neural Network Toolbox” embedded in METLAB was utilized to perform the Back-Propagation Artificial Neural Network Learning. The heating wattages were divided into 7 groups: the whole wattage (30W, 40W, 50W, and 60W), low wattage (30W, and 40W), high wattage (50W, and 60W), and respective single wattage, for performing cross reference, so as to predict the thermal resistance values of the heat pipes having non-linear relations. The findings indicated: When the preset functions and hidden layer formulas in this experiment were utilized, the thermal resistance values of the grooved heat pipes could be predicted accurately and effectively. Also, the verification precision rate of random sampling could reach 100%.
41
Lin, Kuan-Yu, and 林冠佑. "Performance Analysis of Probabilistic Neural Network Data Fusion Algorithms." Thesis, 2005. http://ndltd.ncl.edu.tw/handle/08209130999429651632.
Full textAbstract:
碩士育達商業技術學院
資訊管理所
93
The desired improvements of multi-sensor network tracking system rely on more accurate state estimates and less computation loads. An algorithm is presented to the problem of a distributed multi-sensor network track to track data fusion. For sensor level, to reduce the computational loads involved in physical implementation, the method is essentially based on the decoupling technique that Kalman filter gain formulations are recursively computed. For central level, an approach called Probabilistic Neural Network algorithm is utilized to process state estimation using track data transmitted from sensor level. Performance results for the proposed algorithm are compared with that of the sensor level, using computer simulations of typical target maneuvering scenarios.
42
Tsai, Jun-Cheng, and 蔡俊成. "Implementation of High Performance Hardware Based Toroidal Neural Network." Thesis, 2007. http://ndltd.ncl.edu.tw/handle/6k6ymd.
Full textAbstract:
碩士國立臺北科技大學
自動化科技研究所
95
Neural networks play an important role in artificial intelligence application domains. In most of applications, neural networks are often implemented in software form. Although the software implementation of neural networks provides flexibility, the operating speed is limited due to the sequential machine architecture. In most applications, the learning procedure is carried off-line. A large amount of mathematics operations are needed when learning task of neural networks is performed. The neural network systems implemented using software can only work well in high speed computers. The performance is not adequate when it is implemented on embedded systems. Following the development of modern technologies, people attempt to realize the neural networks by hardware in order to improve the performance. Designs utilizing special architectures and parameters to achieve the performance were proposed in the past in order to provide higher performance. This thesis proposes a high efficiency and generic neural network hardware architecture. The architecture uses the toroidal series multiple data stream to process the back propagation neural network operations, which has the full function of recall and learning capabilities. Users can adjust the number of processor unit in the system based on the requirement of the applications. Since the proposed system is developed in hardware, it can be integrated into embedded systems. The experimental results show that the system can reach higher performance by using fewer logical elements while maintaining flexibility.
43
"Radial basis function of neural network in performance attribution." 2003. http://library.cuhk.edu.hk/record=b5891681.
Full textAbstract:
Wong Hing-Kwok.Thesis (M.Phil.)--Chinese University of Hong Kong, 2003.
Includes bibliographical references (leaves 34-35).
Abstracts in English and Chinese.
Abstract --- p.i
Acknowledgement --- p.iii
Chapter 1 --- Introduction --- p.1
Chapter 2 --- Radial Basis Function (RBF) of Neural Network --- p.5
Chapter 2.1 --- Neural Network --- p.6
Chapter 2.2 --- Radial Basis Function (RBF) Network --- p.8
Chapter 2.3 --- Model Specification --- p.10
Chapter 2.4 --- Estimation --- p.12
Chapter 3 --- RBF in Performance Attribution --- p.17
Chapter 3.1 --- Background of Data Set --- p.18
Chapter 3.2 --- Portfolio Construction --- p.20
Chapter 3.3 --- Portfolio Rebalance --- p.22
Chapter 3.4 --- Result --- p.23
Chapter 4 --- Comparison --- p.26
Chapter 4.1 --- Standard Linear Model --- p.27
Chapter 4.2 --- Fixed Additive Model --- p.28
Chapter 4.3 --- Refined Additive Model --- p.29
Chapter 4.4 --- Result --- p.30
Chapter 5 --- Conclusion --- p.32
Bibliography --- p.34
44
JU, LEE HUI, and 李惠如. "The Call Warrants Evaluation Performance by Using Neural Network." Thesis, 2003. http://ndltd.ncl.edu.tw/handle/68874859539232572923.
Full textAbstract:
碩士國立臺北大學
企業管理學系
91
In the recent years, the theory and practice of derivatives develop rapidly. To discover the reason, we can trace it back from the European call pricing formula derived from Fischer Black and Myron Scholes. The B-S options pricing formulas not only becomes the pricing foundation of many financial derivatives, but also greatly contribute to the development of option pricing model and application of financial engineering. Although many investors broadly use the B-S pricing model, many studies show that they’re still many unreasonable problems exiting in it. In later studies, researcher developed other model to price the derivatives. For example: Monte Carlo Simulation Among later studies, the most mentionable method is the financial quantitative methodology based on computational intelligence, e.g. Neural Network、Genetic Algorithm、Chaos theory etc. This financial pricing methodology based on the machine learning techniques is the most complicated and difficult way of the financial application. This study tries to use two neural network models as the evaluated model of Taiwan warrants market. They are RBFN and BPN. We will compare the evaluation performance of neural network pricing model and traditional B-S pricing model, to see which one is the best pricing model of Taiwan warrants market. Our study shows that: 1. In the estimation of stock volatility, implied volatility is better than historical volatility. 2. Between the two neural network-pricing models, the BPN is better. 3. The estimated warrants market prices of neural network pricing models are the closest to the real market price.
45
Liu, Jun. "Performance investigation of artificial neural network models of associative memories." 1992. http://hdl.handle.net/1993/18531.
Full text
46
Yu, Kuan-Lun, and 游冠倫. "A Neural Network Model for Executive Compensation and Firm Performance." Thesis, 2010. http://ndltd.ncl.edu.tw/handle/39195410727056303466.
Full textAbstract:
碩士臺灣大學
會計學研究所
98
The conflict of interest between the shareholders and the executives is known as the principal-agent problem. If the shareholders have complete information, they can easily design a contract (or incentive plan) that encourages the actions they want. However, the literature suggests weak or statistically insignificant relation between executive compensation and firm performance. In order to overcome the limitation in prior empirical or analytical studies, this paper investigates the association between executive incentive plans and firm performance by using an artificial neural network. Our results show that, overall, we can accurately associate the executives'' incentive plan with the firm''s performance 63% of the time. For the best and the worst performing firms, the accuracy rate is about 70%. Our findings also suggest that (1) the importance of the component of the incentive plan changes over time, (2) accounting-based performance measure is associated with EPS while market-based performance measure is associated with the market-to-book ratio, and (3) when firms with higher uncertainty, they rely less on stock/option incentives. Finally, the simplicity of the model can help firms better design or change the compensation scheme of the executives.
47
Wang, Zu-Chung, and 王子銓. "Performance Evaluation of Fund Portfolios Constucted by Artificial Neural Network." Thesis, 2008. http://ndltd.ncl.edu.tw/handle/10645166925458239651.
Full textAbstract:
碩士雲林科技大學
財務金融系碩士班
96
Nowdays it is not easy for fund investors to make investment decisions when cofronting wide varieties of fund catalogs.However, as pointed out by plenty of empirical researches in finance that artificial neural networks can have competent capacity in constructing investment portfolios when comparing to more traditional portfolio construction.This article mainly adopts more flexible artificial neural networks to construct fund of funds in Taiwan mutual fund market. The fund of funds construction of this study consists of a two- step procedure. The first step takes Arnott (2004) fundamental index to select the components of the funnd.Then the historical returns of funds using AR 4 form are fed into the feedforward three layers artifical neural network to forecaste the returns of componet funds.Specifically, AR 4 integrated with neural newtowrks is trained in thirty-six months retruns data to forecaste the thirty-seven monthly return of the underlying fund.The procedure is moving forward in one-month window through thirty-six months (three years) forecasting period. The empirical results drawn from this study are as follows. (1).Four types of funds of funds constructed in this study are superior to TAIEX and Taiwan 50 Index. (2)The size of the four funds of funds in this study does not significantly affect the relevant returns. (3) However, the size of the four funds of funds does exhibit inverse relation assicaited with their volatility. (4)In short, the neural networks construction funds of this study in general can not render any excess returns after risk-adjusting.
48
陳國彰. "Comparing Multivariate Analysis and Neural Network Analysis on Performance Evaluation." Thesis, 2003. http://ndltd.ncl.edu.tw/handle/92678916996190099657.
Full textAbstract:
碩士國立交通大學
科技管理學程碩士班
91
In the past two decades, Taiwan electronics and information industries have played important roles in the global market. In order to remain competitive, each company has to know its positioning and competitiveness. This research analyzed the performance of companies based on their financial indexes. Multivariable analysis models such as Factor Analysis, Cluster Analysis, and Discriminant Analysis were used to formulate a model to forecast a company’s future performance. The results were compared with those generated by a Neural Network Analysis Model. This thesis used 243 Taiwanese companies in the electronic and information industries as research subjects. Financial information from 70% of these companies were used for model building, while the rest 30% of companies were used to verify the forecast ability of the two models. Our analysis demonstrated that the Neural Network Analysis Model outperformed the multivariable models.
49
LO, WEN-CHUNG, and 羅文忠. "Performance Evaluation of Dimming Electronic Ballast by Using Artificial Neural Network." Thesis, 2000. http://ndltd.ncl.edu.tw/handle/18272648728914264313.
Full textAbstract:
碩士國立臺灣科技大學
電機工程系
88
High lighting efficiency and stable luminous output of the light sources are the major concern in lighting engineering. The dimming electronic ballast for fluorescent lamp (CFL) developed due to energy consideration becomes the most favorable product in indoor lighting design right now and in coming 20 years. However, the variation and performance changes in electric characteristics and luminous output during dimming process should be measured and evaluated from macroscopic overall point of view in order to certify the dimming stability. The research is devoted to discuss the measuring and evaluating techniques for dimming electronic ballast in order to estimate the electric and luminous performances systematically, especially including the luminous efficiency, current harmonic distortion content, power factor and light flicker. For a long time, researches about dimming fluorescent lamp had focused on the luminous output, power factor and current harmonics distortion(THDA) individually. The weakness of the researches is lacking evaluating flicker effect and the whole performance. The research of this thesis is devoted to plan an effective measuring technique to justify and certify the overall performances of the dimming electronic ballasts by using neural network techniques. The most valuable contribution of the research is that one could develop and modify more powerful and effective dimming electronic ballast to produce much more comfortable light quality accompanied by higher luminous efficiency.
50
Chiang, Yen-Cheng, and 姜彥丞. "Recurrent Neural Network Applied to Performance Analysis of Air Conditioning Systems." Thesis, 2018. http://ndltd.ncl.edu.tw/handle/577zjt.
Full textAbstract:
碩士國立臺灣大學
機械工程學研究所
106
The purpose of this study is to set up an analysis system for air conditioning systems. The first part of the system was construct based on physics. For this purpose; the temperature of refrigerant, temperature and flow rate of water, and input power of compressor have been measured for calculating the coefficient of efficiency(COP), pump efficiency, and irreversibility. Another aim was to develop a method for predicting the performance of chiller. The models for predicting the outlet temperature, input power of compressor, and COP were constructed based on a large dataset obtained from the experiments. The multiple regressions were compared with long short-term memory(LSTM) based recurrent neural network(RNN) for the prediction error. The objective for the analysis system is to make diagnosis and long term monitoring for air conditioning systems simpler and feasible in the industry. The experiments were conducted including water cooled chiller and packaged air conditioner which are located at the industrial area of the collaborate company Dragon Steel Co.,Ltd., which is subsidiary of China Steel Co.,Ltd.. To verify the reliability and validity of the two main idea of the study, the examination processes were carried out with these two cases. The results indicate that the analysis system for detecting the performance of chiller and heat pump is practicable. The trend of the COP, pump efficiency, and irreversibility simulated from the analysis system are fit to the theory and the references. On the other hand, the study shows that the LSTM neural network provides the best results due to the strong ability to model the temporal relationship between time series. For each output parameters, LSTM structure performs more accurate and stable than multiple regression. The analysis system only needs to measure the temperature of refrigerant and water which is more easily than pressure drop and flow rate to obtain. The analysis system could be proposed as an alternative method for engineers to diagnosis or monitor air conditioning systems.