Journal articles on the topic 'Speed Invariant Learned Corners'

Abstract. Indoor localization for pedestrians, which relies solely on inertial odometry, has been a topic of great interest. Its significance lies in its ability to provide positioning solutions independently, without the need for external data. Although traditional strap-down inertial navigation shows rapid drift, the introduction of pedestrian dead reckoning (PDR), and artificial intelligence (AI) has enhanced the applicability of inertial odometry for indoor localization. However, inertial odometry continues to be affected by drift, inherent to the nature of dead reckoning. This implies that even a slight error at a given moment can lead to a significant decrease in accuracy after continuous integration operations. In this paper, we propose a novel approach aimed at enhancing the positioning accuracy of inertial odometry. Firstly, we derive a learning-based forward speed using inertial measurements from a smartphone. Unlike mainstream methods where the learned speed is directly used to determine the position, we use the forward speed combined with non-holonomic constraint (NHC) as a measurement to update the state predicted within a strap-down inertial navigation framework. Secondly, we employ an invariant extended Kalman filter (IEKF)-based state estimation to facilitate fusion to cope with the nonlinearity arising from the system and measurement model. Experimental tests are carried out in different scenarios using an iPhone 12, and traditional methods, including PDR, robust neural inertial navigation (RONIN), and the EKF-based method, are compared. The results suggest that the method we propose surpasses these traditional methods in performance.

Object recognition is central to perception and cognition. Yet relatively little is known about the environmental factors that cause invariant object recognition to emerge in the newborn brain. Is this ability a hardwired property of vision? Or does the development of invariant object recognition require experience with a particular kind of visual environment? Here, we used a high-throughput controlled-rearing method to examine whether newborn chicks ( Gallus gallus ) require visual experience with slowly changing objects to develop invariant object recognition abilities. When newborn chicks were raised with a slowly rotating virtual object, the chicks built invariant object representations that generalized across novel viewpoints and rotation speeds. In contrast, when newborn chicks were raised with a virtual object that rotated more quickly, the chicks built viewpoint-specific object representations that failed to generalize to novel viewpoints and rotation speeds. Moreover, there was a direct relationship between the speed of the object and the amount of invariance in the chick's object representation. Thus, visual experience with slowly changing objects plays a critical role in the development of invariant object recognition. These results indicate that invariant object recognition is not a hardwired property of vision, but is learned rapidly when newborns encounter a slowly changing visual world.

One of the ways to describe objects on images is to identify some of their characteristic points or points of attention. Areas of neighborhoods of attention points are described by descriptors (lots of signs) in such way that they can be identified and compared. These signs are used to search for identical points in other images. The article investigates and establishes the possibility of searching for arbitrary local image regions by descriptors constructed with using invariant moments. A feature of the proposed method is that the calculation of the invariant moments of local areas is carried out with using the integral representation of the geometric moments of the image. Integral representation is a matrix with the same size as the image. The elements of the matrix is the sums of the geometric moments of individual pixels, which are located above and to the left with respect to the coordinates of this element. The number of matrices depends on the order of the geometric moments. For moments up to the second order (inclusively), there will be six such matrices. Calculation of one of six geometric moments of an arbitrary rectangular area of the image comes down up to 3 operations such as summation or subtraction of elements of the corresponding matrix located in the corners of this area. The invariant moments are calculated on base of six geometric moments. The search is performed by scanning the image coordinate grid with a window of a given size. In this case, the invariant moments and additional parameters are calculated and compared with similar parameters of the neighborhoods of the reference point of different size (taking into account the possible change in the image scale). The best option is selected according to a given condition. Almost all mass operations of the procedures for calculating the parameters of standards and searching of identical points make it possible explicitly perform parallel computations in the SIMD mode. As a result, the integral representation of geometric moments and the possibility of using parallel computations at all stages will significantly speed up the calculations and allow you to get good indicators of the search efficiency for identical points and the speed of work

Recently, the event camera has become a popular and promising vision sensor in the research of simultaneous localization and mapping and computer vision owing to its advantages: low latency, high dynamic range, and high temporal resolution. As a basic part of the feature-based SLAM system, the feature tracking method using event cameras is still an open question. In this article, we present a novel asynchronous event feature generation and tracking algorithm operating directly on event-streams to fully utilize the natural asynchronism of event cameras. The proposed algorithm consists of an event-corner detection unit, a descriptor construction unit, and an event feature tracking unit. The event-corner detection unit addresses a fast and asynchronous corner detector to extract event-corners from event-streams. For the descriptor construction unit, we propose a novel asynchronous gradient descriptor inspired by the scale-invariant feature transform descriptor, which helps to achieve quantitative measurement of similarity between event feature pairs. The construction of the gradient descriptor can be decomposed into three stages: speed-invariant time surface maintenance and extraction, principal orientation calculation, and descriptor generation. The event feature tracking unit combines the constructed gradient descriptor and an event feature matching method to achieve asynchronous feature tracking. We implement the proposed algorithm in C++ and evaluate it on a public event dataset. The experimental results show that our proposed method achieves improvement in terms of tracking accuracy and real-time performance when compared with the state-of-the-art asynchronous event-corner tracker and with no compromise on the feature tracking lifetime.

This study proposes a novel receiver structure for underwater vertical acoustic communication in which the bias in the correlation-based estimation for the timing offset is learned and then estimated by a deep neural network (DNN) to an accuracy that renders subsequent use of equalizers unnecessary. For a duration of 7 s, 15 timing offsets of the linear frequency modulation (LFM) signals obtained by the correlation were fed into the DNN. The model was based on the Pierson–Moskowitz (PM) random surface height model with a moderate wind speed and was further verified under various wind speeds and experimental waveforms. This receiver, embedded with the DNN model, demonstrated lower complexity and better performance than the adaptive equalizer-based receiver. The 5000 m depth deep-sea experimental data show the superiority of the proposed combination of DNN-based synchronization and the time-invariant equalizer.

The public pays large sums of money to watch skilled motor performance. Notably, however, in recent decades motor skill learning (performance improvement beyond baseline levels) has received less experimental attention than motor adaptation (return to baseline performance in the setting of an external perturbation). Motor skill can be assessed at the levels of task success and movement quality, but the link between these levels remains poorly understood. We devised a motor skill task that required visually guided curved movements of the wrist without a perturbation, and we defined skill learning at the task level as a change in the speed–accuracy trade-off function (SAF). Practice in restricted speed ranges led to a global shift of the SAF. We asked how the SAF shift maps onto changes in trajectory kinematics, to establish a link between task-level performance and fine motor control. Although there were small changes in mean trajectory, improved performance largely consisted of reduction in trial-to-trial variability and increase in movement smoothness. We found evidence for improved feedback control, which could explain the reduction in variability but does not preclude other explanations such as an increased signal-to-noise ratio in cortical representations. Interestingly, submovement structure remained learning invariant. The global generalization of the SAF across a wide range of difficulty suggests that skill for this task is represented in a temporally scalable network. We propose that motor skill acquisition can be characterized as a slow reduction in movement variability, which is distinct from faster model-based learning that reduces systematic error in adaptation paradigms.

In this study, novel deep learning models based on high-speed flame images are proposed to diagnose the combustion instability of a gas turbine. Two different network layers that can be combined with any existing backbone network are established—(1) An early-fusion layer that can learn to extract the power spectral density of subsequent image frames, which is time-invariant under certain conditions. (2) A late-fusion layer which combines the outputs of a backbone network at different time steps to predict the current combustion state. The performance of the proposed models is validated by the dataset of high speed flame images, which have been obtained in a gas turbine combustor during the transient process from stable condition to unstable condition and vice versa. Excellent performance is achieved for all test cases with high accuracy of 95.1–98.6% and a short processing time of 5.2–12.2 ms. Interestingly, simply increasing the number of input images is as competitive as combining the proposed early-fusion layer to a backbone network. In addition, using handcrafted weights for the late-fusion layer is shown to be more effective than using learned weights. From the results, the best combination is selected as the ResNet-18 model combined with our proposed fusion layers over 16 time-steps. The proposed deep learning method is proven as a potential tool for combustion instability identification and expected to be a promising tool for combustion instability prediction as well.

High-speed trains operate under varying conditions, leading to different distributions of vibration data collected from the wheel bearings. To detect bearing faults in situations where the source and target domains exhibit differing data distributions, the technique of transfer learning can be applied to move the distribution of features gleaned from unlabeled data in the source domain. However, traditional deep transfer learning techniques do not take into account the relationships between subdomains within the same class of different domains, resulting in suboptimal transfer learning performance and limiting the use of intelligent fault diagnosis for wheel bearings under various conditions. In order to tackle this problem, we have developed the Deep Subdomain Transfer Learning Network (DSTLN). This innovative approach transfers the distribution of features by harmonizing the subdomain distributions of layer activations specific to each domain through the implementation of the Local Maximum Mean Discrepancy (LMMD) method. The DSTLN consists of three modules: a feature extractor, fault category recognition, and domain adaptation. The feature extractor is constructed using a newly proposed SA-ConvLSTM model and CNNs, which aim to automatically learn features. The fault category recognition module is a classifier that categorizes the samples based on the extracted features. The domain adaptation module includes an adversarial domain classifier and subdomain distribution discrepancy metrics, making the learned features domain-invariant across both the global domain and subdomains. Through 210 transfer fault diagnosis experiments with wheel bearing data under 15 different operating conditions, the proposed method demonstrates its effectiveness.

In recent years, intelligent fault diagnosis methods based on deep learning have developed rapidly. However, most of the existing work performs well under the assumption that training and testing samples are collected from the same distribution, and the performance drops sharply when the data distribution changes. For rolling bearings, the data distribution will change when the load and speed change. In this article, to improve fault diagnosis accuracy and anti-noise ability under different working loads, a transfer learning method based on multi-scale capsule attention network and joint distributed optimal transport (MSCAN-JDOT) is proposed for bearing fault diagnosis under different loads. Because multi-scale capsule attention networks can improve feature expression ability and anti-noise performance, the fault data can be better expressed. Using the domain adaptation ability of joint distribution optimal transport, the feature distribution of fault data under different loads is aligned, and domain-invariant features are learned. Through experiments that investigate bearings fault diagnosis under different loads, the effectiveness of MSCAN-JDOT is verified; the fault diagnosis accuracy is higher than that of other methods. In addition, fault diagnosis experiment is carried out in different noise environments to demonstrate MSCAN-JDOT, which achieves a better anti-noise ability than other transfer learning methods.

The design of airborne superconducting generators for intermittent duty requires the understanding of some unique free-convection processes in the spinning helium bath. Toward that end, some fundamental experiments on steady and transient free convection in rotating containers of representative geometries have been performed. Heat transfer data from heaters of various geometries mounted on the outer container surface to several fluids are reported. A correlation for steady-state Nusselt number is presented for a wide range of Rayleigh and Prandtl numbers. The heat transfer coefficient was found to be independent of heater size, geometry, and fluid viscosity. Heat transfer measurements during simultaneous thermal transients and sudden increases in rotational speed were also made. They show an enhancement of heat transfer due to the relative counterrotation of the fluid following the acceleration of the container. This persists for a period well below that for fluid spinup. A model based upon the submergence of the thermal boundary layer by the diffusive wave from the wall was successful in correlating this period. Quasi-steady flow visualization experiments indicate that the thermal plumes generate two-dimensional, axially invariant flow fields. Their trajectories are radial relative to the spinning container. Those observations are shown to be consistent with the fact that weak buoyant plumes in containers rotating at small Ekman numbers result in low Rossby number motions. Those are two dimensional according to the Taylor–Proudman theorem. It is shown that the Coriolis and pressure forces on such a thermal column are in azimuthal equilibrium, hence the radial trajectory. Flow visualization following impulsive acceleration in an off-axis, nonaxisymmetric container shows that the flow field is dominated by vortices expelled from corners. The fluid spinup time, however, was found to be the same as that for an on-axis circular cylinder of the same characteristic diameter.

An improved corner detection algorithm for medical image based on USAN is proposed in this paper. The algorithm firstly defines the concept of end line which through the core of the circular template, and analyses the number of end lines in the corner point, thus obtains an initial response function of the corner. Then puts forward a non-maxima suppression method by defining the concept of uniformity coefficient of the non-end line. Finally, compared with existing algorithms, experiments indicates that the improved algorithm has the fast and accurate advantages. I.IntroductionCorner detection based on grayscale images has been widely used in medical image registration. Medical image registration techniques can be divided into two categories, which is based on pixels (voxels) and the feature-based methods. Feature-based registration method is matched by extracting the structural features of different images (the common features including points, contours, curves and surfaces, etc.) , in which the registration based on feature points is the most widely applied method[1]. Corner points as the feature of image has rich information content and a little calculation, matching simple and rotation, translation, scaling invariant properties. In recent years, some improved methods for Harris algorithm[2,3] and USAN (Univalue Segment Assimilating Nucleus, with the value of the shrinking core) algorithm[4] method have been applied in image registration. In 2002, Zhou Peng, etc. developed a new corner detection method based on the model proposed in the USAN[5].Experimental results show that image registration based on the method without restrictions on the rotation of the image, it improves the registration accuracy and has less computation. In 2006, Rae etc. built a new multi-scale Harris corner detection algorithm[6]. The new corner-detection method can obtain corner points at different scales corner, improved the detection performance of detection operators, advanced the registration accuracy of corner-based image registration algorithm. In 2007, Huashun Gang, etc. absorbed and improved the SUSAN , proposed a dense image matching algorithm[7], using Sobel operator and the feature similarity to initially match the corner on the left and right image, achieved dense matching of all the pixels of two images which without calibration. At present, image corner detection based on gray, in particular the method based on template made great development. However, the accuracy of these methods are mostly poor, algorithm complexity, time-consuming, so have some limits in real-time applications. Based on the USAN, a new improved algorithm was proposed and has been applied to medical image corner detection, the new algorithm in terms of speed and accuracy has a larger increase.II.Susan Corner Detection PrincipleSUSAN algorithm [4] defines a circular template, as shown in Fig. 1, the center of circular template is called the core, if some pixel's brightness of the circular template is the same or similar to the core, the region composed with these pixels will be defined USAN region ,at the corner point and non-corner point, USAN region area is different, and the USAN area of the corner point is the smallest.Fig. 1 The circular template of SUSAN algorithmDoing the specific defection, a circular template with 37 pixel will be used to scan the entire image, and to calculate the gray intensity’s minus between each pixel and the core of the template, if the absolute value is less than the given threshold t ,which will be considered to belong to USAN ,the commonly used similarity comparison function is as follows: (1)in this formula,is the discriminate result，is the gray value of the core of circular template， is the gray value of the arbitrary pixel in addition to the core of the template.Apply the formula (1) to each pixel, then the size of the USAN area can be expressed as: (2)In this formula, is the circular template whose core is [8].After getting the USAN area of all the pixels, then according to the following corner response function to generate the initial corner response: (3)In this formula, g is the geometric threshold value of the noise suppression, which decides the maximum value of output corners’ USAN area.SUSAN algorithms do not need the beforehand edge detection, which avoid the calculation of the gradient and not depend on the results of image segmentation, with integral characteristics, good anti-noise performance, and is not influenced by the type of the corner point, so it is widely used in image processing. However, a fixed threshold value will also make the positioning accuracy is not satisfactory, easy to produce pseudo-response, or easy to lose the true corner point, the process of integration has also led to more time-consuming.III.Description of the Improved AlgorithmA. Corner Detection Based On Non-end LineUse the circular template in SUSAN algorithm ,in the circular template start a axle-box whose origin is the core, if the pixels at the axis positive direction are all located in the USAN outside, there is the pixels within the USAN in the negative direction, state the core as the end point at the direction of the axis ,state the axis as the end line of the core, otherwise state it as the non-end line, as shown in Fig. 2,is the end point at the direction of ,is the end line. Due to that there are no pixels at the negative direction of within the USAN, so is not the end point at the direction of ,is non-end line, as the same, is not the end point at the direction of , , , is non-end line ,is the end point at the direction of , but is not the end point at the direction of , is the end line, is the non-end line.By analysing of the fig.2, we can get the following laws: there are numerous end lines and non-end lines inner the circular template whose core is the corner point; there are numerous end lines inner the circular template whose core is the edge point; but there is only one non-end line, which is the axis down the direction of USAN area edge; there are numerous non-end lines, but there is no end line inner the circular template whose core is inner point. That is to say, the corner point has numerous end lines and non-end lines; the edge point has numerous end points, but there is only one non-end line; the inner point has numerous non-end line, but no end lines.Fig. 2 The end line and non-end line of the circular templateFig. 3 The corner inspection based on the number of the end linesAccording to the above rules, you can filter out some of the non-corner points by the way of judging the number of the end lines and non-end lines inner the circular mask, and generate an initial response of the corner points. In Fig. 3(a), eight straight lines which pass through the circle are distributed in the circular mask, it can be respectively stated as , it is the USAN area shown by the shaded area, Stated from the Fig., the number of the end lines and non-end lines are respectively five and three, according the above rules, we can judge O is the corner point . In actual detections, considering of a circular template including 37 pixels, shown in Fig.3(b), each edge pixel and the pixel on the symmetry of the core constitute a line straight passing through the core. for example, the two pixels numbered g constitute the straight line g-g, the same as Fig.3(a), all pixels were composed of eight straight lines. Using a circular template to scan the entire image, to calculate the number of the end lines and non-end lines in each template to produce a candidate corner points.In Fig. 3 (a) , the line constituted by two symmetrical edge pixels is (), the collection of the end lines within the round templates is and the collection of the non-end lines is , then the judging formula of the end line is: (4)In this formula, 、 respectively stands for the similarity values of the two symmetrical edge pixels with the core ,and is the non-end line rejection threshold. For easy to calculate, the definition of the degree of the end line is: (5)Then the number of end line within a circular template is: (6)So the initial response function of the corner point is: (7)In this formula, is the geometric threshold for restraining edge points.Only the relationship between the edge pixel and the core is considered in the introduction of the new algorithms, which will generate some pseudo-response at the corner points, a number of suppression methods should be adopted to filter, but based on that the calculation amount can be greatly reduced because of the edge points’ initial response, which makes it more applicable for real-time applications.B. Non-maxima Suppression Based on Non-end Line’s uniformity coefficientAs a result of adapting the similarity between the edge pixels and core to generate the initial response of the corner points, pseudo-response will be produced in a certain range near the corner, as shown in Fig. 4, in the circular template 1, A corner points can be got by, , in the circular template 2, B corner points can be got by, , apparently B is a pseudo-response. The following non-maxima suppression method based on the non-end line’s uniformity can be adopted to filter the false responses.Fig. 4 The uniformity degree of the non-end lineAccording to Section 3.1, non-end line has two forms, the one is the direction line which is not passing through the USAN area (assuming that a small neighbor area near the core of the USAN area is negligible), the other is the direction line which is passing through the USAN area but the two intersection with the circular template is located outside USAN area. If a non-end lines do not pass through USAN region, we call it is uniform, or else call it a non-uniform, and according to the length of the segment inner USAN area to define the degree of non-uniform, the longer is, the greater the degree of non-uniform is, shown in Fig. 4, the degree of non-uniform 、 is greater than 、(, is uniform), it is easy to prove, the degree of non-uniform at the corner point is the greatest, but the degree of non-uniform at the non-corner point is the smallest, so according to this feature, pseudo-response suppression can be operated.While actually use it, first use the judging formula of the end lines to get the non-end lines, then through calculating the length of the line segment in the USAN region to get the degree of non-uniform, suppose the degree of non-uniform of the non-end line as ,so (8) is the non-end line which is passing through the core of the circular template, thus, the degree of non-uniform of all the non-end lines that have passed through is :(9)represent all the non-end lines that passes through the core of the circular template. Finally, the non-maxima suppression function is as follows: (10) in this formula, is the geometric value of pseudo-response suppression, which determines the sharp degree of the output corner points.IV. Experimental VerificationTo test the effectiveness of the new algorithm, this paper compared the improved algorithm with the present algorithm in accuracy and computational complexity .First, through adopting artificial typical corner-point image to test, shown as Fig. 5, (a) is the initial image, (b)~(d) is the effective image which respectively adopted Harris algorithm, SUSAN algorithm and the improved algorithm, through analysing, Harris algorithm has positioning bias in some corner points, there is also a small amount of pseudo-response, while SUSAN and the improved algorithm has greater effect. (a) Initial image (b) Harris (c) SUSAN (d) Improved algorithmFig. 5 The detection results in artificial corner detection of the improved algorithm and the initial algorithmFig. 6 is the detection test results of the three kinds of algorithm in medical image, you can see, Harris has stronger ability to process actual image processing ,but has a large number of pseudo-response points and there are few undetected corner. SUSAN has not very good processing effect to process actual image and has a lot of pseudo-response points. The new algorithm also has a little pseudo-response points, but on the whole there is a better detection result than the SUSAN algorithm .TableⅠ shows the average detection rate of the three algorithms in a variety samples. (a) Initial image (b) Harris (c) SUSAN (d) Improved algorithmFig. 6 The test results of the improved algorithm and the initial algorithm in medical imageTABLEⅠTHE AVERAGE DETECTION RATE OF THE THERE ALGORITHMWe can know from the former analysis of the algorithm, the improved algorithm do the complex calculations only on the most likely place for the corner points, but do little calculations on the easily judge where is the non-corner points, and the new algorithm is not based on a very abstract theory, which makes that the improved algorithm simple and easy to understand, with low computational complexity. The author adapted a variety samples to add up the detection time of the three algorithm using own computer, shown in the TableⅡ, we can know, the detection time of the improved algorithm is significantly lower than other algorithm.TABLEⅡ THE AVERAGE DETECTION TIME OF THE THERE ALGORITHMV.ConclusionsIn order to improve detection speed, the improved algorithm uses a progressive layer by layer mechanism, the basic idea is: first, through the most simple calculation to exclude certain non-corner points, followed the possible areas for the corner point will be implement the more sophisticated detection algorithm, due to the proportion of the corner point in image is small, this mechanism greatly increased the detection rate. For the pre-processing of the exclusion of the non-corner point, this paper proposed a method based on the smallest circular template with small amount of calculation, and can retain all the characteristics of corner points. In the corner detection algorithm, using the number characteristics of the corner points’ end line, and according to the non-end lines’ uniformity coefficient to perform non-maxima suppression, experiments show that the improved algorithm is effective in the speed and accuracy of calculation is better than the SUSAN method .ReferencesVanden P, Evert Jan D P, Viergever M, ”Medical image matching-a review with classification,” IEEE Engineering in Medicine and Biology, vol.12, no.1, pp. 26-39, January 1993.Jianhui Hou, Yi Lin, ” Harris checkerboard corner detection algorithm with self-adapting,”Computer Engineering and Design, vol.30, no.20, pp.4741 -4743, October 2009Harris C, Stephens M. A Combined Corner and Edge Detector. In Proceedings of the 4thAlvey Vision Conference, pp. 147-151,1988.Smith S M, Brady M. SUSAN, ”A new approach to low level image processing,” International Journal of Computer Vision, vol.23, no.1, pp.45-48, January 1997.Peng Zhou, Yong Tan, Shoushi Xu, ”A new image registration algorithm based on the corner detection ,”College journal of China University of Science and Technology, vol. 32, no. 4, pp. 455-461, August 2002. Bo Li, Dan Yang, Xiaohong Zhang,”A new image registration algorithm based on Harris multi-scale corner detection , ”Computer Engineering and Applications, no. 35, pp. 37-40, December 2006.Gang Huashun,Yi Zeng ,”A new image dense registration algorithm based on corner detection ,” Computer Engineering and Design, vol. 28, no.5, pp. 1092-1095, March 2007.Haixia Guo, Kai Xie.,” An improved corner detection algorithm based on USAN,”Computer Engineering, vol. 33, no. 22, pp. 232-234, November 2007.

We propose a learning approach to corner detection for event-based cameras that is stable even under fast and abrupt motions. Event-based cameras offer high temporal resolution, power efficiency, and high dynamic range. However, the properties of event-based data are very different compared to standard intensity images, and simple extensions of corner detection methods designed for these images do not perform well on event-based data. We first introduce an efficient way to compute a time surface that is invariant to the speed of the objects. We then show that we can train a Random Forest to recognize events generated by a moving corner from our time surface. Random Forests are also extremely efficient, and therefore a good choice to deal with the high capture frequency of event-based cameras—our implementation processes up to 1.6Mev/s on a single CPU. Thanks to our time surface formulation and this learning approach, our method is significantly more robust to abrupt changes of direction of the corners compared to previous ones. Our method also naturally assigns a confidence score for the corners, which can be useful for postprocessing. Moreover, we introduce a highresolution dataset suitable for quantitative evaluation and comparison of corner detection methods for event-based cameras. We call our approach SILC, for Speed Invariant Learned Corners, and compare it to the state-of-the-art with extensive experiments, showing better performance.

AbstractRecent years have seen vast progress in the development of machine learned force fields (MLFFs) based on ab-initio reference calculations. Despite achieving low test errors, the reliability of MLFFs in molecular dynamics (MD) simulations is facing growing scrutiny due to concerns about instability over extended simulation timescales. Our findings suggest a potential connection between robustness to cumulative inaccuracies and the use of equivariant representations in MLFFs, but the computational cost associated with these representations can limit this advantage in practice. To address this, we propose a transformer architecture called SO3krates that combines sparse equivariant representations (Euclidean variables) with a self-attention mechanism that separates invariant and equivariant information, eliminating the need for expensive tensor products. SO3krates achieves a unique combination of accuracy, stability, and speed that enables insightful analysis of quantum properties of matter on extended time and system size scales. To showcase this capability, we generate stable MD trajectories for flexible peptides and supra-molecular structures with hundreds of atoms. Furthermore, we investigate the PES topology for medium-sized chainlike molecules (e.g., small peptides) by exploring thousands of minima. Remarkably, SO3krates demonstrates the ability to strike a balance between the conflicting demands of stability and the emergence of new minimum-energy conformations beyond the training data, which is crucial for realistic exploration tasks in the field of biochemistry.

Executing safe and precise flight maneuvers in dynamic high-speed winds is important for the ongoing commoditization of uninhabited aerial vehicles (UAVs). However, because the relationship between various wind conditions and its effect on aircraft maneuverability is not well understood, it is challenging to design effective robot controllers using traditional control design methods. We present Neural-Fly, a learning-based approach that allows rapid online adaptation by incorporating pretrained representations through deep learning. Neural-Fly builds on two key observations that aerodynamics in different wind conditions share a common representation and that the wind-specific part lies in a low-dimensional space. To that end, Neural-Fly uses a proposed learning algorithm, domain adversarially invariant meta-learning (DAIML), to learn the shared representation, only using 12 minutes of flight data. With the learned representation as a basis, Neural-Fly then uses a composite adaptation law to update a set of linear coefficients for mixing the basis elements. When evaluated under challenging wind conditions generated with the Caltech Real Weather Wind Tunnel, with wind speeds up to 43.6 kilometers/hour (12.1 meters/second), Neural-Fly achieves precise flight control with substantially smaller tracking error than stateof-the-art nonlinear and adaptive controllers. In addition to strong empirical performance, the exponential stability of Neural-Fly results in robustness guarantees. Last, our control design extrapolates to unseen wind conditions, is shown to be effective for outdoor flights with only onboard sensors, and can transfer across drones with minimal performance degradation.

Human history has witnessed varied surveillance and counter-surveillance activities from time immemorial. Human beings could not surveille others effectively and accurately without the technology of their era. Technology is a tool that can empower both people and governments. The outcomes are different based on the users’ intentions and aims. 2,500 years ago, Sun Tzu noted that ‘If you know both yourself and your enemy, you can win numerous (literally, "a hundred") battles without jeopardy’. His words still ring true. To be a good surveiller and counter-surveiller it is essential to know both sides, and in order to be good at these activities access to technology is vital. There is no doubt that knowledge is power, and without technology to access the information, it is impossible to be powerful. As we become more expert at technology, we will learn what makes surveillance and counter-surveillance more effective, and will be more powerful.“Surveillance” is one of the most important aspects of living in the convergent media environment. This essay illustrates government surveillance and counter-surveillance during the Iranian Green Movement (2009) on social and mobile media. The Green Movement refers to a non-violent movement that arose after the disputed presidential election on June 2009. After that Iran was facing its most serious political crisis since the 1979 revolution. Claims of vote fraud triggered massive street protests. Many took to the streets with “Green” signs, chanting slogans such as ‘the government lied’, and ‘where is my vote?’ There is no doubt that social and mobile media has played an important role in Iran’s contemporary politics. According to Internet World Stats (IWS) Internet users in 2009 account for approximately 48.5 per cent of the population of Iran. In 2009, Iran had 30.2 million mobile phone users (Freedom House), and 72 cellular subscriptions for every 100 people (World Bank). Today, while Iran has the 19th-largest population in the world, its blogosphere holds the third spot in terms of number of users, just behind the United States and China (Beth Elson et al.). In this essay the use of social and mobile media (technology) is not debated, but the extent of this use, and who, why and how it is used, is clearly scrutinised.Visibility and Surveillance There have been different kinds of surveillance for a very long time. However, all types of surveillance are based on the notion of “visibility”. Previous studies show that visibility is not a new term (Foucault Discipline). The new things in the new era, are its scale, scope and complicated ways to watch others without being watched, which are not limited to a specific time, space and group, and are completely different from previous instruments for watching (Andrejevic). As Meikle and Young (146) have mentioned ‘networked digital media bring with them a new kind of visibility’, based on different kinds of technology. Internet surveillance has important implications in politics to control, protect, and influence (Marx Ethics; Castells; Fuchs Critique). Surveillance has been improved during its long history, and evolved from very simple spying and watching to complicated methods of “iSpy” (Andrejevic). To understand the importance of visibility and its relationship with surveillance, it is essential to study visibility in conjunction with the notion of “panopticon” and its contradictory functions. Foucault uses Bentham's notion of panopticon that carries within itself visibility and transparency to control others. “Gaze” is a central term in Bentham’s view. ‘Bentham thinks of a visibility organised entirely around a dominating, overseeing gaze’ (Foucault Eye). Moreover, Thomson (Visibility 11) notes that we are living in the age of ‘normalizing the power of the gaze’ and it is clear that the influential gaze is based on powerful means to see others.Lyon (Surveillance 2) explains that ‘surveillance is any collection and processing of personal data, whether identifiable or not, for the purpose of influencing or managing those whose data have been granted…’. He mentions that today the most important means of surveillance reside in computer power which allows collected data to be sorted, matched, retrieved, processed, marketed and circulated.Nowadays, the Internet has become ubiquitous in many parts of the world. So, the changes in people’s interactions have influenced their lives. Fuchs (Introduction 15) argues that ‘information technology enables surveillance at a distance…in real time over networks at high transmission speed’. Therefore, visibility touches different aspects of people’s lives and living in a “glasshouse” has caused a lot of fear and anxiety about privacy.Iran’s Green Movement is one of many cases for studying surveillance and counter-surveillance technologies in social and mobile media. Government Surveillance on Social and Mobile Media in Iran, 2009 In 2009 the Iranian government controlled technology that allowed them to monitor, track, and limit access to the Internet, social media and mobiles communication, which has resulted in the surveillance of Green Movement’s activists. The Iranian government had improved its technical capabilities to monitor the people’s behavior on the Internet long before the 2009 election. The election led to an increase in online surveillance. Using social media the Iranian government became even more powerful than it was before the election. Social media was a significant factor in strengthening the government’s power. In the months after the election the virtual atmosphere became considerably more repressive. The intensified filtering of the Internet and implementation of more advanced surveillance systems strengthened the government’s position after the election. The Open Net Initiative revealed that the Internet censorship system in Iran is one of the most comprehensive and sophisticated censorship systems in the world. It emphasized that ‘Advances in domestic technical capacity have contributed to the implementation of a centralized filtering strategy and a reduced reliance on Western technologies’.On the other hand, the authorities attempted to block all access to political blogs (Jaras), either through cyber-security methods or through threats (Tusa). The Centre for Investigating Organized Cyber Crimes, which was founded in 2007 partly ‘to investigate and confront social and economic offenses on the Internet’ (Cyber Police), became increasingly important over the course of 2009 as the government combated the opposition’s online activities (Beth Elson et al. 16). Training of "senior Internet lieutenants" to confront Iran's "virtual enemies online" was another attempt that the Intelligence minister announced following the protests (Iran Media Program).In 2009 the Iranian government enacted the Computer Crime Law (Jaras). According to this law the Committee in Charge of Determining Unauthorized Websites is legally empowered to identify sites that carry forbidden content and report that information to TCI and other major ISPs for blocking (Freedom House). In the late fall of 2009, the government started sending threatening and warning text messages to protesters about their presence in the protests (BBC). Attacking, blocking, hacking and hijacking of the domain names of some opposition websites such as Jaras and Kaleme besides a number of non-Iranian sites such as Twitter were among the other attempts of the Iranian Cyber Army (Jaras).It is also said that the police and security forces arrested dissidents identified through photos and videos posted on the social media that many imagined had empowered them. Furthermore, the online photos of the active protesters were posted on different websites, asking people to identify them (Valizadeh).In late June 2009 the Iranian government was intentionally permitting Internet traffic to and from social networking sites such as Facebook and Twitter so that it could use a sophisticated practice called Deep Packet Inspection (DPI) to collect information about users. It was reportedly also applying the same technology to monitor mobile phone communications (Beth Elson et al. 15).On the other hand, to cut communication between Iranians inside and outside the country, Iran slowed down the Internet dramatically (Jaras). Iran also blocked access to Facebook, YouTube, Wikipedia, Twitter and many blogs before, during and after the protests. Moreover, in 2009, text message services were shut down for over 40 days, and mobile phone subscribers could not send or receive text messages regardless of their mobile carriers. Subsequently it was disrupted on a temporary basis immediately before and during key protests days.It was later discovered that the Nokia Siemens Network provided the government with surveillance technologies (Wagner; Iran Media Program). The Iranian government built a complicated system that enabled it to monitor, track and intercept what was said on mobile phones. Nokia Siemens Network confirmed it supplied Iran with the technology needed to monitor, control, and read local telephone calls [...] The product allowed authorities to monitor any communications across a network, including voice calls, text messaging, instant messages, and web traffic (Cellan-Jones). Media sources also reported that two Chinese companies, Huawei and ZTE, provided surveillance technologies to the government. The Nic Payamak and Saman Payamak websites, that provide mass text messaging services, also reported that operator Hamrah Aval commonly blocked texts with words such as meeting, location, rally, gathering, election and parliament (Iran Media Program). Visibility and Counter-Surveillance The panopticon is not limited to the watchers. Similarly, new kinds of panopticon and visibility are not confined to government surveillance. Foucault points out that ‘the seeing machine was once a sort of dark room into which individuals spied; it has become a transparent building in which the exercise of power may be supervised by society as a whole’ (Discipline 207). What is important is Foucault's recognition that transparency, not only of those who are being observed but also of those who are observing, is central to the notion of the panopticon (Allen) and ‘any member of society will have the right to come and see with his own eyes how schools, hospitals, factories, and prisons function’ (Foucault, Discipline 207). Counter-surveillance is the process of detecting and mitigating hostile surveillance (Burton). Therefore, while the Internet is a surveillance instrument that enables governments to watch people, it also improves the capacity to counter-surveille, and draws public attention to governments’ injustice. As Castells (185) notes the Internet could be used by citizens to watch their government as an instrument of control, information, participation, and even decision-making, from the bottom up.With regards to the role of citizens in counter-surveillance we can draw on Jay Rosen’s view of Internet users as ‘the people formerly known as the audience’. In counter-surveillance it can be said that passive citizens (formerly the audience) have turned into active citizens. And this change was becoming impossible without mobile and social media platforms. These new techniques and technologies have empowered people and given them the opportunity to have new identities. When Thompson wrote ‘the exercise of power in modern societies remains in many ways shrouded in secrecy and hidden from the public gaze’ (Media 125), perhaps he could not imagine that one day people can gaze at the politicians, security forces and the police through the use of the Internet and mobile devices.Furthermore, while access to mobile media allows people to hold authorities accountable for their uses and abuses of power (Breen 183), social media can be used as a means of representation, organization of collective action, mobilization, and drawing attention to police brutality and reasons for political action (Gerbaudo).There is no doubt that having creativity and using alternative platforms are important aspects in counter-surveillance. For example, images of Lt. Pike “Pepper Spray Cop” from the University of California became the symbol of the senselessness of police brutality during the Occupy Movement (Shaw). Iranians’ Counter-Surveillance on Social and Mobile Media, 2009 Iran’s Green movement (2009) triggered a lot of discussions about the role of technology in social movements. In this regard, there are two notable attitudes about the role of technology: techno-optimistic (Shriky and Castells) and techno-pessimistic (Morozov and Gladwell) views should be taken into account. While techno-optimists overrated the role of social media, techno-pessimists underestimated its role. However, there is no doubt that technology has played a great role as a counter-surveillance tool amongst Iranian people in Iran’s contemporary politics.Apart from the academic discussions between techno-optimists and techno-pessimists, there have been numerous debates about the role of new technologies in Iran during the Green Movement. This subject has received interest from different corners of the world, including Western countries, Iranian authorities, opposition groups, and also some NGOs. However, its role as a means of counter-surveillance has not received adequate attention.As the tools of counter-surveillance are more or less the tools of surveillance, protesters learned from the government to use the same techniques to challenge authority on social media.Establishing new websites (such as JARAS, RASA, Kalemeh, and Iran green voice) or strengthening some previous ones (such as Saham, Emrooz, Norooz), also activating different platforms such as Facebook, Twitter, and YouTube accounts to broadcast the voice of the Iranian Green Movement and neutralize the government’s propaganda were the most important ways to empower supporters of Iran’s Green Movement in counter-surveillance.‘Reporters Without Borders issued a statement, saying that ‘the new media, and particularly social networks, have given populations collaborative tools with which they can change the social order’. It is also mentioned that despite efforts by the Iranian government to prevent any reporting of the protests and due to considerable pressure placed on foreign journalists inside Iran, social media played a significant role in sending the messages and images of the movement to the outside world (Axworthy). However, at that moment, many thought that Twitter performed a liberating role for Iranian dissenters. For example, Western media heralded the Green Movement in Iran as a “Twitter revolution” fuelled by information and communication technologies (ICTs) and social media tools (Carrieri et al. 4). “The Revolution Will Be Twittered” was the first in a series of blog posts published by Andrew Sullivan a few hours after the news of the protests was released.According to the researcher’s observation the numbers of Twitter users inside Iran who tweeted was very limited in 2009 and social media was most useful in the dissemination of information, especially from those inside Iran to outsiders. Mobile phones were mostly influential as an instrument firstly used for producing contents (images and videos) and secondly for the organisation of protests. There were many photos and videos that were filmed by very simple mobile cell phones, uploaded by ordinary people onto YouTube and other platforms. The links were shared many times on Twitter and Facebook and released by mainstream media. The most frequently circulated story from the Iranian protests was a video of Neda Agha-Sultan. Her final moments were captured by some bystanders with mobile phone cameras and rapidly spread across the global media and the Internet. It showed that the camera-phone had provided citizens with a powerful means, allowing for the creation and instant sharing of persuasive personalised eyewitness records with mobile and globalised target populations (Anden-Papadopoulos).Protesters used another technique, DDOS (distributed denial of service attacks), for political protest in cyber space. Anonymous people used DDOS to overload a website with fake requests, making it unavailable for users and disrupting the sites set as targets (McMillan) in effect, shutting down the site. DDOS is an important counter-surveillance activity by grassroots activists or hackers. It was a cyber protest that knocked the main Iranian governmental websites off-line and caused crowdsourcing and false trafficking. Amongst them were Mahmoud Ahmadinejad, Iran's supreme leader’s websites and those which belong to or are close to the government or security forces, including news agencies (Fars, IRNA, Press TV…), the Ministry of Foreign Affairs, the Ministry of Justice, the Police, and the Ministry of the Interior.Moreover, as authorities uploaded the pictures of protesters onto different platforms to find and arrest them, in some cities people started to put the pictures, phone numbers and addresses of members of security forces and plain clothes police officers who attacked them during the protests and asked people to identify and report the others. They also wanted people to send information about suspects who infringed human rights. Conclusion To sum up, visibility, surveillance and counter-surveillance are not new phenomena. What is new is the technology, which increased their complexity. As Foucault (Discipline 200) mentioned ‘visibility is a trap’, so being visible would be the weakness of those who are being surveilled in the power struggle. In the convergent era, in order to be more powerful, both surveillance and counter-surveillance activities aim for more visibility. Although both attempt to use the same means (technology) to trap the other side, the differences are in their subjects, objects, goals and results.While in surveillance, visibility of the many by the few is mostly for the purpose of control and influence in undemocratic ways, in counter-surveillance, the visibility of the few by the many is mostly through democratic ways to secure more accountability and transparency from the governments.As mentioned in the case of Iran’s Green Movement, the scale and scope of visibility are different in surveillance and counter-surveillance. The importance of what Shaw wrote about Sydney occupy counter-surveillance, applies to other places, such as Iran. She has stressed that ‘protesters and police engaged in a dance of technology and surveillance with one another. Both had access to technology, but there were uncertainties about the extent of technology and its proficient use…’In Iran (2009), both sides (government and activists) used technology and benefited from digital networked platforms, but their levels of access and domains of influence were different, which was because the sources of power, information and wealth were divided asymmetrically between them. Creativity was important for both sides to make others more visible, and make themselves invisible. Also, sharing information to make the other side visible played an important role in these two areas. References Alen, David. “The Trouble with Transparency: The Challenge of Doing Journalism Ethics in a Surveillance Society.” Journalism Studies 9.3 (2008): 323-40. 8 Dec. 2013 ‹http://www.tandfonline.com/doi/full/10.1080/14616700801997224#.UqRFSuIZsqN›. Anden-Papadopoulos, Kari. “Citizen Camera-Witnessing: Embodied Political Dissent in the Age of ‘Mediated Mass Self-Communication.’” New Media & Society 16.5 (2014). 753-69. 9 Aug. 2014 ‹http://nms.sagepub.com/content/16/5/753.full.pdf+html›. Andrejevic, Mark. iSpy: Surveillance and Power in the Interactive Era. Lawrence, Kan: UP of Kansas, 2007. Axworthy, Micheal. Revolutionary Iran: A History of the Islamic Republic. London: Penguin Books, 2014. Bentham, Jeremy. Panopticon Postscript. London: T. Payne, 1791. Beth Elson, Sara, Douglas Yeung, Parisa Roshan, S.R. Bohandy, and Alireza Nader. Using Social Media to Gauge Iranian Public Opinion and Mood after the 2009 Election. Santa Monica: RAND Corporation, 2012. 1 Aug. 2014 ‹http://www.rand.org/content/dam/rand/pubs/technical_reports/2012/RAND_TR1161.pdf›. Breen, Marcus. Uprising: The Internet’s Unintended Consequences. Champaign, Ill: Common Ground Pub, 2011. Burton, Fred. “The Secrets of Counter-Surveillance.” Stratfor Global Intelligence. 2007. 19 April 2015 ‹https://www.stratfor.com/secrets_countersurveillance›. Carrieri, Matthew, Ali Karimzadeh Bangi, Saad Omar Khan, and Saffron Suud. After the Green Movement Internet Controls in Iran, 2009-2012. OpenNet Initiative, 2013. 17 Dec. 2013 ‹https://opennet.net/sites/opennet.net/files/iranreport.pdf›. Castells, Manuel. The Internet Galaxy: Reflections on the Internet, Business, and Society. Oxford: Oxford UP: 2001. Cellan-Jones, Rory. “Hi-Tech Helps Iranian Monitoring.” BBC, 2009. 26 July 2014 ‹http://news.bbc.co.uk/1/hi/technology/8112550.stm›. “Cyber Crimes’ List.” Iran: Cyber Police, 2009. 17 July 2014 ‹http://www.cyberpolice.ir/page/2551›. Foucault, Michel. Discipline and Punish: The Birth of the Prison. Trans. Alan Sheridan. Harmondsworth: Penguin, 1977. Foucault, Michel. “The Eye of Power.” 1980. 12 Dec. 2013 ‹https://nbrokaw.files.wordpress.com/2010/12/the-eye-of-power.doc›. Freedom House. “Special Report: Iran.” 2009. 14 June 2014 ‹http://www.sssup.it/UploadDocs/4661_8_A_Special_Report_Iran_Feedom_House_01.pdf›. Fuchs, Christian. “Introduction.” Internet and Surveillance: The Challenges of Web 2.0 and Social Media. Ed. Christian Fuchs. London: Routledge, 2012. 1-28. Fuchs, Christian. “Critique of the Political Economy of Web 2.0 Surveillance.” Internet and Surveillance: The Challenges of Web 2.0 and Social Media. Ed. Christian Fuchs. London: Routledge, 2012. 30-70. Gerbaudo, Paolo. Tweets and the Streets: Social Media and Contemporary Activism. London: Pluto, 2012. “Internet: Iran’s New Imaginary Enemy.” Jaras Mar. 2009. 28 June 2014 ‹http://www.rahesabz.net/print/12143›.Iran Media Program. “Text Messaging as Iran's New Filtering Frontier.” 2013. 25 July 2014 ‹http://www.iranmediaresearch.org/en/blog/227/13/04/25/136›. Internet World Stats News. The Internet Hits 1.5 Billion. 2009. 3 July 2014 ‹ http://www.internetworldstats.com/pr/edi038.htm›. Lyon, David. Surveillance Society: Monitoring Everyday Life. Buckingham: Open UP, 2001. Lyon, David. “9/11, Synopticon, and Scopophilia: Watching and Being Watched.” The New Politics of Surveillance and Visibility. Eds. Richard V. Ericson and Kevin D. Haggerty. Toronto: UP of Toronto, 2006. 35-54. Marx, Gary T. “What’s New about the ‘New Surveillance’? Classify for Change and Continuity.” Surveillance & Society 1.1 (2002): 9-29. McMillan, Robert. “With Unrest in Iran, Cyber-Attacks Begin.” PC World 2009. 17 Apr. 2015 ‹http://www.pcworld.com/article/166714/article.html›. Meikle, Graham, and Sherman Young. Media Convergence: Networked Digital Media in Everyday Life. London: Palgrave Macmillan, 2012. Morozov, Evgeny. “How Dictators Watch Us on the Web.” Prospect 2009. 15 June 2014 ‹http://www.prospectmagazine.co.uk/magazine/how-dictators-watch-us-on-the-web/#.U5wU6ZRdU00›.Open Net. “Iran.” 2009. 26 June 2014 ‹https://opennet.net/research/profiles/iran›. Reporters without Borders. “Web 2.0 versus Control 2.0.” 2010. 27 May 2014 ‹http://en.rsf.org/web-2-0-versus-control-2-0-18-03-2010,36697›.Rosen, Jay. The People Formerly Known as the Audience. 2006. 7 Dec. 2013 ‹http://www.huffingtonpost.com/jay-rosen/the-people-formerly-known_1_b_24113.html›. Shaw, Frances. “'Walls of Seeing': Protest Surveillance, Embodied Boundaries, and Counter-Surveillance at Occupy Sydney.” Transformation 23 (2013). 9 Dec. 2013 ‹http://www.transformationsjournal.org/journal/issue_23/article_04.shtml›. “The Warning of the Iranian Revolutionary Guard Corps (IRGC) to the Weblogs and Websites.” BBC, 2009. 27 July 2014 ‹http://www.bbc.co.uk/persian/iran/2009/06/090617_ka_ir88_sepah_internet.shtml›. Thompson, John B. The Media And Modernity: A Social Theory of the Media. Cambridge: Polity Press, 1995. Thompson, John B. “The New Visibility.” Theory, Culture & Society 22.6 (2005): 31-51. 10 Dec. 2013 ‹http://tcs.sagepub.com/content/22/6/31.full.pdf+html›. Tusa, Felix. “How Social Media Can Shape a Protest Movement: The Cases of Egypt in 2011 and Iran in 2009.” Arab Media and Society 17 (Winter 2013). 15 July 2014 ‹http://www.arabmediasociety.com/index.php?article=816&p=0›. Tzu, Sun. Sun Tzu: The Art of War. S.l.: Pax Librorum Pub. H, 2009. Valizadeh, Reza. “Invitation to the Public Shooting with the Camera.” RFI, 2011. 19 June 2014 ‹http://www.persian.rfi.fr/%D8%AF%D8%B9%D9%88%D8%AA-%D8%A8%D9%87-%D8%B4%D9%84%DB%8C%DA%A9-%D8%B9%D9%85%D9%88%D9%85%DB%8C-%D8%A8%D8%A7-%D8%AF%D9%88%D8%B1%D8%A8%DB%8C%D9%86-%D8%B9%DA%A9%D8%A7%D8%B3%DB%8C-20110307/%D8%A7%DB%8C%D8%B1%D8%A7%D9%86›. Wagner, Ben. Exporting Censorship and Surveillance Technology. Netherlands: Humanist Institute for Co-operation with Developing Countries (Hivos), 2012. 7 July 2014 ‹https://hivos.org/sites/default/files/exporting_censorship_and_surveillance_technology_by_ben_wagner.pdf›. World Bank. Mobile Cellular Subscriptions (per 100 People). The World Bank. N.d. 27 June 2014 ‹http://data.worldbank.org/indicator/IT.CEL.SETS.P2›.