Search

Video enhancement is a challenging problem, more than that of stills, mainly due to high computational cost, larger data volumes and the difficulty of achieving consistency in the spatio-temporal domain. In practice, these challenges are often coupled with the lack of example pairs, which inhibits the application of supervised learning strategies. To address these challenges, we propose an efficient adversarial video enhancement framework that learns directly from unpaired video examples. In particular, our framework introduces new recurrent cells that consist of interleaved local and global modules for implicit integration of spatial and temporal information. The proposed design allows our recurrent cells to efficiently propagate spatio-temporal information across frames and reduces the need for high complexity networks. Our setting enables learning from unpaired videos in a cyclic adversarial manner, where the proposed recurrent units are employed in all architectures. Efficient training is accomplished by introducing one single discriminator that learns the joint distribution of source and target domain simultaneously. The enhancement results demonstrate clear superiority of the proposed video enhancer over the state-of-the-art methods, in all terms of visual quality, quantitative metrics, and inference speed. Notably, our video enhancer is capable of enhancing over 35 frames per second of FullHD video (1080x1920)., International Journal of Computer Vision, 131 (4), ISSN:0920-5691, ISSN:1573-1405

Significance Sequential activity is a prominent feature of many neural systems, in multiple behavioral contexts. Here, we investigate how Hebbian rules lead to storage and recall of random sequences of inputs in both rate and spiking recurrent networks. In the case of the simplest (bilinear) rule, we characterize extensively the regions in parameter space that allow sequence retrieval and compute analytically the storage capacity of the network. We show that nonlinearities in the learning rule can lead to sparse sequences and find that sequences maintain robust decoding but display highly labile dynamics to continuous changes in the connectivity matrix, similar to recent observations in hippocampus and parietal cortex., Sequential activity has been observed in multiple neuronal circuits across species, neural structures, and behaviors. It has been hypothesized that sequences could arise from learning processes. However, it is still unclear whether biologically plausible synaptic plasticity rules can organize neuronal activity to form sequences whose statistics match experimental observations. Here, we investigate temporally asymmetric Hebbian rules in sparsely connected recurrent rate networks and develop a theory of the transient sequential activity observed after learning. These rules transform a sequence of random input patterns into synaptic weight updates. After learning, recalled sequential activity is reflected in the transient correlation of network activity with each of the stored input patterns. Using mean-field theory, we derive a low-dimensional description of the network dynamics and compute the storage capacity of these networks. Multiple temporal characteristics of the recalled sequential activity are consistent with experimental observations. We find that the degree of sparseness of the recalled sequences can be controlled by nonlinearities in the learning rule. Furthermore, sequences maintain robust decoding, but display highly labile dynamics, when synaptic connectivity is continuously modified due to noise or storage of other patterns, similar to recent observations in hippocampus and parietal cortex. Finally, we demonstrate that our results also hold in recurrent networks of spiking neurons with separate excitatory and inhibitory populations.

International audience; Reinforcement Learning (RL) agents are commonly thought of as adaptive decision procedures. They work on input/output data streams called "states", "actions" and "rewards". Most current research about RL adaptiveness to changes works under the assumption that the streams signatures (i.e. arity and types of inputs and outputs) remain the same throughout the agent lifetime. As a consequence, natural situations where the signatures vary (e.g. when new data streams become available, or when others become obsolete) are not studied. In this paper, we relax this assumption and consider that signature changes define a new learning situation called Protean Learning (PL). When they occur, traditional RL agents become undefined, so they need to restart learning. Can better methods be developed under the PL view? To investigate this, we first construct a stream-oriented formalism to properly define PL and signature changes. Then, we run experiments in an idealized PL situation where input addition and deletion occur during the learning process. Results show that a simple PL-oriented method enables graceful adaptation of these arity changes, and is more efficient than restarting the process.

Significance This review attempts to unite three hitherto rather unconnected concepts of basic functions of the cerebral cortex, taking the visual system as an example: 1) feed-forward processing in multilayer hierarchies (labeled line coding), 2) dynamic association of features (assembly coding), and 3) matching of sensory evidence with stored priors (predictive coding). The latter two functions are supposed to rely on the high-dimensional dynamics of delay-coupled recurrent networks. Discharge rates of neurons (rate code) and temporal relations among discharges (temporal code) are identified as conveying complementary information. Thus, the new concept accounts for the coexistence of feed-forward and recurrent processing, accommodates both rate and temporal codes, and assigns crucial functions to the complex dynamics emerging from recurrent interactions., Current concepts of sensory processing in the cerebral cortex emphasize serial extraction and recombination of features in hierarchically structured feed-forward networks in order to capture the relations among the components of perceptual objects. These concepts are implemented in convolutional deep learning networks and have been validated by the astounding similarities between the functional properties of artificial systems and their natural counterparts. However, cortical architectures also display an abundance of recurrent coupling within and between the layers of the processing hierarchy. This massive recurrence gives rise to highly complex dynamics whose putative function is poorly understood. Here a concept is proposed that assigns specific functions to the dynamics of cortical networks and combines, in a unifying approach, the respective advantages of recurrent and feed-forward processing. It is proposed that the priors about regularities of the world are stored in the weight distributions of feed-forward and recurrent connections and that the high-dimensional, dynamic space provided by recurrent interactions is exploited for computations. These comprise the ultrafast matching of sensory evidence with the priors covertly represented in the correlation structure of spontaneous activity and the context-dependent grouping of feature constellations characterizing natural objects. The concept posits that information is encoded not only in the discharge frequency of neurons but also in the precise timing relations among the discharges. Results of experiments designed to test the predictions derived from this concept support the hypothesis that cerebral cortex exploits the high-dimensional recurrent dynamics for computations serving predictive coding.

Multi-sensor distributed fusion estimation algorithms based on machine learning are proposed in this paper. Firstly, using local estimations as inputs and estimations of three classic distributed fusion (weighted by matrices, by diagonal matrices and by scalars) as the training sets, three distributed fusion algorithms based on BP network (BP net-based fusion weighted by matrices, by diagonal matrices and by scalar) are proposed and the selection basis of the number of nodes in hidden layer is given. Furthermore, by using local estimations as inputs and centralized fusion estimation as training set, another recurrent net-based distributed fusion algorithm is proposed, in the case that neither true states nor cross-covariance matrices is available. This method is not limited to the linear minimum variance (LMV) criterion, so its accuracy is higher than the classical three distributed fusion algorithms. A radar tracking simulation verifies the effectiveness of the proposed fusion networks.

International audience; Self-driving vehicles need to continuously analyse the driving scene, understand the behavior of other road users and predict their future trajectories in order to plan a safe motion and reduce their reaction time. Motivated by this idea, this paper addresses the problem of vehicle trajectory prediction over an extended horizon. On highways, human drivers continuously adapt their speed and paths according to the behavior of their neighboring vehicles. Therefore, vehicles' trajectories are very correlated and considering vehicle interactions makes motion prediction possible even before the start of a clear maneuver pattern. To this end, we introduce and analyze trajectory prediction methods based on how they model the vehicles interactions. Inspired by human reasoning, we use an attention mechanism that explicitly highlights the importance of neighboring vehicles with respect to their future states. We go beyond pairwise vehicle interactions and model higher order interactions. Moreover, the existence of different goals and driving behaviors induces multiple potential futures. We exploit a combination of global and partial attention paid to surrounding vehicles to generate different possible trajectory. Experiments on highway datasets show that the proposed model outperforms the state-of-the-art performances.

Neural-network-based image and video captioning can be substantially improved by utilizing architectures that make use of special features from the scene context, objects, and locations. A novel discriminatively trained evaluator network for choosing the best caption among those generated by an ensemble of caption generator networks further improves accuracy.

Exo-atmospheric infrared (IR) point target discrimination is an important research topic of space surveillance systems. It is difficult to describe the characteristic information of the shape and micro-motion states of the targets and to discriminate different targets effectively by the characteristic information. This paper has constructed the infrared signature model of spatial point targets and obtained the infrared radiation intensity sequences dataset of different types of targets. This paper aims to design an algorithm for the classification problem of infrared radiation intensity sequences of spatial point targets. Recurrent neural networks (RNNs) are widely used in time series classification tasks, but face several problems such as gradient vanishing and explosion, etc. In view of shortcomings of RNNs, this paper proposes an independent random recurrent neural network (IRRNN) model, which combines independent structure RNNs with randomly weighted RNNs. Without increasing the training complexity of network learning, our model solves the problem of gradient vanishing and explosion, improves the ability to process long sequences, and enhances the comprehensive classification performance of the algorithm effectively. Experiments show that the IRRNN algorithm performs well in classification tasks and is robust to noise.

Currently, many applications in Machine Learning are based on define new models to extract more information about data, In this case Deep Reinforcement Learning with the most common application in video games like Atari, Mario, and others causes an impact in how to computers can learning by himself with only information called rewards obtained from any action. There is a lot of algorithms modeled and implemented based on Deep Recurrent Q-Learning proposed by DeepMind used in AlphaZero and Go. In this document, We proposed Deep Recurrent Double Q-Learning that is an implementation of Deep Reinforcement Learning using Double Q-Learning algorithms and Recurrent Networks like LSTM and DRQN., Comment: Accepted paper on LatinXinAI Workshop co-located with the International Conference on Machine Learning (ICML) 2019

Action recognition plays an important role in various applications such as video monitoring, automatic video indexing, crowd analysis, human-machine interaction, smart homes and personal assistive robotics. In this paper, we propose improvements to some methods for human action recognition from videos that work with data represented in the form of skeleton poses. These methods are based on the most widely used techniques for this problem—Graph Convolutional Networks (GCNs), Temporal Convolutional Networks (TCNs) and Recurrent Neural Networks (RNNs). Initially, the paper explores and compares different ways to extract the most relevant spatial and temporal characteristics for a sequence of frames describing an action. Based on this comparative analysis, we show how a TCN type unit can be extended to work even on the characteristics extracted from the spatial domain. To validate our approach, we test it against a benchmark often used for human action recognition problems and we show that our solution obtains comparable results to the state-of-the-art, but with a significant increase in the inference speed.

In this work, the authors propose several techniques for accelerating a modern action recognition pipeline. This article reviewed several recent and popular action recognition works and selected two of them as part of the tools used for improving the aforementioned acceleration. Specifically, temporal segment networks (TSN), a convolutional neural network (CNN) framework that makes use of a small number of video frames for obtaining robust predictions which have allowed to win the first place in the 2016 ActivityNet challenge, and MotionNet, a convolutional-transposed CNN that is capable of inferring optical flow RGB frames. Together with the last proposal, this article integrated a new software for decoding videos that takes advantage of NVIDIA GPUs. This article shows a proof of concept for this approach by training the RGB stream of the TSN network in videos loaded with NVIDIA Video Loader (NVVL) of a subset of daily actions from the University of Central Florida 101 dataset.

Our decisions often depend on multiple sensory experiences separated by time delays. The brain can remember these experiences and, simultaneously, estimate the timing between events. To understand the mechanisms underlying working memory and time encoding, we analyze neural activity recorded during delays in four experiments on nonhuman primates. To disambiguate potential mechanisms, we propose two analyses, namely, decoding the passage of time from neural data and computing the cumulative dimensionality of the neural trajectory over time. Time can be decoded with high precision in tasks where timing information is relevant and with lower precision when irrelevant for performing the task. Neural trajectories are always observed to be low-dimensional. In addition, our results further constrain the mechanisms underlying time encoding as we find that the linear “ramping” component of each neuron’s firing rate strongly contributes to the slow timescale variations that make decoding time possible. These constraints rule out working memory models that rely on constant, sustained activity and neural networks with high-dimensional trajectories, like reservoir networks. Instead, recurrent networks trained with backpropagation capture the time-encoding properties and the dimensionality observed in the data., Research was supported by NSF Next Generation Network for Neuroscience Award DBI-1707398, the Gatsby Charitable Foundation, the Simons Foundation, the Swartz Foundation (C.J.C. and S.F.), NIH Training Grant 5T32NS064929 (to C.J.C.), and the Kavli Foundation (S.F.). M.N.S. was supported by National Institute of Neurological Disorders and Stroke Brain Initiative Grant R01NS113113. R.R. was supported by the Dirección General de Asuntos del Personal Académico de la Universidad Nacional Autónoma de México (PAPIIT-IN210819) and Consejo Nacional de Ciencia y Tecnología (CONACYT-240892). A.G. was supported by National Institute of Mental Health Division of Intramural Research Grant Z01MH-01092 and by Italian Fondo per gli investimenti della ricerca di base 2010 Grant RBFR10G5W9_001.

Classification of very high-resolution (VHR) satellite images has three major challenges: 1) inherent low intraclass and high interclass spectral similarities; 2) mismatching resolution of available bands; and 3) the need to regularize noisy classification maps. Conventional methods have addressed these challenges by adopting separate stages of image fusion, feature extraction, and postclassification map regularization. These processing stages, however, are not jointly optimizing the classification task at hand. In this paper, we propose a single-stage framework embedding the processing stages in a recurrent multiresolution convolutional network trained in an end-to-end manner. The feedforward version of the network, called FuseNet , aims to match the resolution of the panchromatic and multispectral bands in a VHR image using convolutional layers with corresponding downsampling and upsampling operations. Contextual label information is incorporated into FuseNet by means of a recurrent version called ReuseNet . We compared FuseNet and ReuseNet against the use of separate processing steps for both image fusions, e.g., pansharpening and resampling through interpolation and map regularization such as conditional random fields . We carried out our experiments on a land-cover classification task using a Worldview-03 image of Quezon City, Philippines, and the International Society for Photogrammetry and Remote Sensing 2-D semantic labeling benchmark data set of Vaihingen, Germany. FuseNet and ReuseNet surpass the baseline approaches in both the quantitative and qualitative results.

Earth observation (EO) sensors deliver data at daily or weekly intervals. Mostland use and land cover classification(LULC) approaches, however, are designed for cloud-free and mono-temporal observations. The increasing temporal capabilities of today’s sensors enable the use of temporal, along with spectral and spatial features. Domains such as speech recognition or neural machine translation, work with inherently temporal data and, today, achieve impressive results by using sequential encoder-decoder structures. Inspired by these sequence-to-sequence models, we adapt an encoder structure with convolutional recurrent layers in order to approximate a phenological model for vegetation classes based on a temporal sequence ofSentinel 2(S2) images. In our experiments, we visualize internal activations over a sequence of cloudy and non-cloudy images and find several recurrent cells that reduce the input activity for cloudy observations. Hence, we assume that our network has learned cloud-filtering schemes solely from input data, which could alleviate the need for tedious cloud-filtering as a preprocessing step for many EO approaches. Moreover, using unfiltered temporal series oftop-of-atmosphere(TOA) reflectance data, our experiments achieved state-of-the-art classification accuracies on a large number of crop classes with minimal preprocessing, compared to other classification approaches.

Feedback in Neuro-Evolution is explored and evaluated for its application in devising prediction models for foreign currency exchange rates. A novel approach to foreign currency exchange rates forecasting based on Recurrent Neuro-Evolution is introduced. Cartesian Genetic Programming (CGP) is the algorithm deployed for the forecasting model. Recurrent Cartesian Genetic Programming evolved Artificial Neural Network (RCGPANN) is demonstrated to produce computationally efficient and accurate model for forex prediction with an accuracy of as high as 98.872% for a period of 1000 days. The approach utilizes the trends that are being followed in historical data to predict five currency rates against Australian dollar. The model is evaluated using statistical metrics and compared. The computational method outperforms the other methods particularly due to its capability to select the best possible feature in real time and the flexibility that the system provides in feature selection, connectivity pattern and network.

The discovery of stimulus induced synchronization in the visual cortex suggested the possibility that the relations among low-level stimulus features are encoded by the temporal relationship between neuronal discharges. In this framework, temporal coherence is considered a signature of perceptual grouping. This insight triggered a large number of experimental studies which sought to investigate the relationship between temporal coordination and cognitive functions. While some core predictions derived from the initial hypothesis were confirmed, these studies, also revealed a rich dynamical landscape beyond simple coherence whose role in signal processing is still poorly understood. In this paper, a framework is presented which establishes links between the various manifestations of cortical dynamics by assigning specific coding functions to low-dimensional dynamic features such as synchronized oscillations and phase shifts on the one hand and high-dimensional non-linear, non-stationary dynamics on the other. The data serving as basis for this synthetic approach have been obtained with chronic multisite recordings from the visual cortex of anesthetized cats and from monkeys trained to solve cognitive tasks. It is proposed that the low-dimensional dynamics characterized by synchronized oscillations and large-scale correlations are substates that represent the results of computations performed in the high-dimensional state-space provided by recurrently coupled networks.

As we look around a scene, we perceive it as continuous and stable even though each saccadic eye movement changes the visual input to the retinas. How the brain achieves this perceptual stabilization is unknown, but a major hypothesis is that it relies on presaccadic remapping, a process in which neurons shift their visual sensitivity to a new location in the scene just before each saccade. This hypothesis is difficult to test in vivo because complete, selective inactivation of remapping is currently intractable. We tested it in silico with a hierarchical, sheet-based neural network model of the visual and oculomotor system. The model generated saccadic commands to move a video camera abruptly. Visual input from the camera and internal copies of the saccadic movement commands, or corollary discharge, converged at a map-level simulation of the frontal eye field (FEF), a primate brain area known to receive such inputs. FEF output was combined with eye position signals to yield a suitable coordinate frame for guiding arm movements of a robot. Our operational definition of perceptual stability was “useful stability,” quantified as continuously accurate pointing to a visual object despite camera saccades. During training, the emergence of useful stability was correlated tightly with the emergence of presaccadic remapping in the FEF. Remapping depended on corollary discharge but its timing was synchronized to the updating of eye position. When coupled to predictive eye position signals, remapping served to stabilize the target representation for continuously accurate pointing. Graded inactivations of pathways in the model replicated, and helped to interpret, previous in vivo experiments. The results support the hypothesis that visual stability requires presaccadic remapping, provide explanations for the function and timing of remapping, and offer testable hypotheses for in vivo studies. We conclude that remapping allows for seamless coordinate frame transformations and quick actions despite visual afferent lags. With visual remapping in place for behavior, it may be exploited for perceptual continuity.

Working memory (WM), including a ‘central executive’, is used to guide behavior by internal goals or intentions. We suggest that WM is best described as a set of three interdependent functions which are implemented in the prefrontal cortex (PFC). These functions are maintenance, control of attention and integration. A model for the maintenance function is presented, and we will argue that this model can be extended to incorporate the other functions as well. Maintenance is the capacity to briefly maintain information in the absence of corresponding input, and even in the face of distracting information. We will argue that maintenance is based on recurrent loops between PFC and posterior parts of the brain, and probably within PFC as well. In these loops information can be held temporarily in an active form. We show that a model based on these structural ideas is capable of maintaining a limited number of neural patterns. Not the size, but the coherence of patterns (i.e., a chunking principle based on synchronous firing of interconnected cell assemblies) determines the maintenance capacity. A mechanism that optimizes coherent pattern segregation, also poses a limit to the number of assemblies (about four) that can concurrently reverberate. Top-down attentional control (in perception, action and memory retrieval) can be modelled by the modulation and re-entry of top-down information to posterior parts of the brain. Hierarchically organized modules in PFC create the possibility for information integration. We argue that large-scale multimodal integration of information creates an ‘episodic buffer’, and may even suffice for implementing a central executive.

We present a simple autoassociative neural network model of cortical adaptation that replicates a high-level category aftereffect. We study in our model the effect of a priming stimulus on retrieval from a briefly presented partial cue. We also explore, in our general framework, which learning conditions permit correct pattern retrieval.

Walking animals, like stick insects, cockroaches or ants, demonstrate a fascinating range of locomotive abilities and complex behaviors. The locomotive behaviors can consist of a variety of walking patterns along with adaptation that allow the animals to deal with changes in environmental conditions, like uneven terrains, gaps, obstacles etc. Biological study has revealed that such complex behaviors are a result of a combination of biomechanics and neural mechanism thus representing the true nature of embodied interactions. While the biomechanics helps maintain flexibility and sustain a variety of movements, the neural mechanisms generate movements while making appropriate predictions crucial for achieving adaptation. Such predictions or planning ahead can be achieved by way of internal models that are grounded in the overall behavior of the animal. Inspired by these findings, we present here, an artificial bio-inspired walking system which effectively combines biomechanics (in terms of the body and leg structures) with the underlying neural mechanisms. The neural mechanisms consist of (1) central pattern generator based control for generating basic rhythmic patterns and coordinated movements, (2) distributed (at each leg) recurrent neural network based adaptive forward models with efference copies as internal models for sensory predictions and instantaneous state estimations, and (3) searching and elevation control for adapting the movement of an individual leg to deal with different environmental conditions. Using simulations we show that this bio-inspired approach with adaptive internal models allows the walking robot to perform complex locomotive behaviors as observed in insects, including walking on undulated terrains, crossing large gaps, leg damage adaptations, as well as climbing over high obstacles. Furthermore, we demonstrate that the newly developed recurrent network based approach to online forward models outperforms the adaptive neuron forward models, which have hitherto been the state of the art, to model a subset of similar walking behaviors in walking robots. peerReviewed

International audience; The BLSTM-CTC is a novel recurrent neural network architecture that has outperformed previous state of the art algorithms in tasks such as speech recognition or handwriting recognition. It has the ability to process long term dependencies in temporal signals in order to label unsegmented data. This paper describes different ways of combining features using a BLSTM-CTC architecture. Not only do we explore the low level combination (feature space combination) but we also explore high level combination (decoding combination) and mid-level (internal system representation combination). The results are compared on the RIMES word database. Our results show that the low level combination works best, thanks to the powerful data modeling of the LSTM neurons.

Short-term plasticity (STP) is a phenomenon that widely occurs in the neocortex with implications for learning and memory. Based on a widely used STP model, we develop an analytical characterization of the STP parameter space to determine the nature of each synapse (facilitating, depressing or both) in a spiking neural network based on presynaptic firing rate and the corresponding STP parameters. We demonstrate consistency with previous work by leveraging the power of our characterization to replicate the functional volumes that are integral for the previous network stabilization results. We then use our characterization to predict the precise transitional point from the facilitating regime to the depressing regime in a simulated synapse, suggesting 'in vitro' experiments to verify the underlying STP model. We conclude the work by integrating our characterization into a framework for finding suitable STP parameters for self-sustaining random, asynchronous activity in a prescribed recurrent spiking neural network. The systematic process resulting from our analytical characterization improves the success rate of finding the requisite parameters for such networks by three orders of magnitude over a random search.

In 1986, Rumelhart and McClelland took the cognitive science community by storm with the Parallel Distributed Processing (PDP) framework. Rather than abstracting from the biological substrate as was sought by the “information processing” paradigms of the 1970s, connectionism, as it has come to be called, embraced it. An immediate appeal of the connectionist agenda was its aim: to construct at the algorithmic level models of cognition that were compatible with their implementation in the biological substrate. The PDP group argued that this could be achieved by turning to networks of artificial neurons, originally introduced by McCulloch and Pitts (1943) which the group showed were able to provide insights into a wide range of psychological domains, from categorization, to perception, to memory, to language. This work built on an earlier formulation by Rosenblatt (1958) who introduced a simple type of feed-forward neural network called the perceptron. Perceptrons were limited to solving simple linearly-separable problems and although networks composed of perceptrons were known to be able to compute any Boolean function (including XOR, Minsky and Papert, 1969), there was no effective way of training such networks. In 1986, Rumelhart, Hinton and Williams introduced the back-propagation algorithm, providing an effective way of training multi-layered neural networks, which could easily learn non linearly-separable functions. In addition to providing the field with an effective learning algorithm, the PDP group published a series of demonstrations of how long standing questions in cognitive psychology could be elegantly solved using simple learning rules, distributed representations, and interactive processing. To take a classic example, consider the word-superiority effect, in which people can detect letters within a word faster than individual letters or letters within a non-word (Reicher, 1969). This result is difficult to square with serial “information-processing” theories of cognition that were dominant at the time (how could someone recognize “R” before “FRIEND” if recognizing the word required recognizing the letters?). Accounting for such findings demanded a framework which could naturally accommodate interactive processes within a bidirectional flow of information. The so-called “Interactive-activation model” (McClelland and Rumelhart, 1981) provided just such a framework. The connectionist paradigm was not without its critics. The principal critiques can be divided into three classes. First, some neuroscientists (Crick, 1989) questioned the biological plausibility of backpropagation, when they failed to observe experimentally complex and differentiated back-propagating signals that are required to learn in multi-layered neural networks. A second critique concerned stability-plasticity of the learned representations in these models. Some phenomena require the ability to rapidly learn new information, but sometimes newly learned knowledge overwrites previously learned information (catastrophic interference; McCloskey and Cohen, 1989). Third, representing spatial and temporal invariance—something that apparently came easily to people—was difficult for models, e.g., recognizing that the letter “T” in “TOM” was the “same” as the “T” in “POT.” This invariance problem was typically solved by multiplying a large number of hard-wired units that were space- or time-locked (see e.g., McClelland and Elman, 1986). Finally, critics pointed out that the networks were incapable of learning true rules on which a number of human behavioral, namely language-learning was thought to depend (e.g., Marcus, 2003; cf. Fodor and Pylyshyn, 1988; Seidenberg, 1999). The connectionist approach has embraced these challenges: Although some connectionist models continue to rely on backpropagation, others have moved to more biologically realistic learning rules (Giese and Poggio, 2003; Masquelier and Thorpe, 2007). Far from being a critical flaw of connectionism, the phenomenon of catastrophic interference (Mermillod et al., 2013) proved to be a feature that led to the development of complementary learning systems (McClelland et al., 1995). Progress has also been made on the invariance problem. For example, within the speech domain representing the similarity between similar speech sounds regardless of their location within a word has been addressed in the past by Grossberg and Myers (2000) and Norris (1994) and this issue presents a new more streamlined and computationally efficient model (Hannagan et al., 2013). An especially powerful approach to solving the location invariance problem in the visual domain is presented by Di Bono and Zorzi (2013), also in this issue. A key challenge for connectionism is to explain the learning of abstract structural representations. The use of recurrent networks (Elman, 1990; Dominey, 2013) and self-organizing maps, has captured important aspects of language learning (e.g., Mayor and Plunkett, 2010; Li and Zhao, 2013), while work on deep learning (Hinton and Salakhutdinov, 2006) has made it possible to model the emergence of structured and abstract representations within multi-layered hierarchical networks (Zorzi et al., 2013). The work on verbal analogies by Kollias and McClelland (2013) continues to address the challenges of modeling more abstract representations, but truly understanding how neural architectures give rise to symbolic cognition is a gap that remains. Although learning and representing formal language rules may not be completely outside of the abilities of neural networks (e.g., Chang, 2009), it seems clear that understanding human cognition requires understanding how we solve these symbolic problems (Clark and Karmiloff-Smith, 1993; Lupyan, 2013). Future generations of connectionist modelers may wish to fill this gap and in so doing provide a fuller picture of how neural networks give rise to intelligence of the sort enables us to ponder the very workings of our cognition.

Delay match to sample (DMS) experiments provide an important link between the theory of recurrent network models and behavior and neural recordings. We define a simple recurrent network of binary neurons with stochastic neural dynamics and Hebbian synaptic learning. Most DMS experiments involve heavily learned images, and in this setting we propose a readout mechanism for match occurrence based on a smaller increment in overall network activity when the matched pattern is already in working memory, and a reset mechanism to clear memory from stimuli of previous trials using random network activity. Simulations show that this model accounts for a wide range of variations on the original DMS tasks, including ABBA tasks with distractors, and more general repetition detection tasks with both learned and novel images. The differences in network settings required for different tasks derive from easily defined changes in the levels of noise and inhibition. The same models can also explain experiments involving repetition detection with novel images, although in this case the readout mechanism for match is based on higher overall network activity. The models give rise to interesting predictions that may be tested in neural recordings.

International audience; This paper deals with the application of temporal averaging methods to recurrent networks of noisy neurons undergoing a slow and unsupervised modification of their connectivity matrix called learning. Three time-scales arise for these models: (i) the fast neuronal dynamics, (ii) the intermediate external input to the system, and (iii) the slow learning mechanisms. Based on this time-scale separation, we apply an extension of the mathematical theory of stochastic averaging with periodic forcing in order to derive a reduced deterministic model for the connectivity dynamics. We focus on a class of models where the activity is linear to understand the specificity of several learning rules (Hebbian, trace or anti-symmetric learning). In a weakly connected regime, we study the equilibrium connectivity which gathers the entire 'knowledge' of the network about the inputs. We develop an asymptotic method to approximate this equilibrium. We show that the symmetric part of the connectivity post-learning encodes the correlation structure of the inputs, whereas the anti-symmetric part corresponds to the cross correlation between the inputs and their time derivative. Moreover, the time-scales ratio appears as an important parameter revealing temporal correlations.

Our ability to generate well-timed sequences of movements is critical to an array of behaviors, including the ability to play a musical instrument or a video game. Here we address two questions relating to timing with the goal of better understanding the neural mechanisms underlying temporal processing. First, how does accuracy and variance change over the course of learning of complex spatiotemporal patterns? Second, is the timing of sequential responses most consistent with starting and stopping an internal timer at each interval or with continuous timing? To address these questions we used a psychophysical task in which subjects learned to reproduce a sequence of finger taps in the correct order and at the correct times – much like playing a melody at the piano. This task allowed us to calculate the variance of the responses at different time points using data from the same trials. Our results show that while “standard” Weber’s law is clearly violated, variance does increase as a function of time squared, as expected according to the generalized form of Weber’s law – which separates the source of variance into time-dependent and time-independent components. Over the course of learning, both the time-independent variance and the coefficient of the time-dependent term decrease. Our analyses also suggest that timing of sequential events does not rely on the resetting of an internal timer at each event. We describe and interpret our results in the context of computer simulations that capture some of our psychophysical findings. Specifically, we show that continuous timing, as opposed to “reset” timing, is consistent with “population clock” models in which timing emerges from the internal dynamics of recurrent neural networks.

The recurrent network of Xia et al. (1996) was proposed for solving quadratic programming problems and was recently adapted to support vector machine (SVM) learning by Tan et al. (2000). We show that this formulation contains some unnecessary circuits which, furthermore, can fail to provide the correct value of one of the SVM parameters and suggest how to avoid these drawbacks.

We survey and summarize the literature on the computational aspects of neural network models by presenting a detailed taxonomy of the various models according to their complexity theoretic characteristics. The criteria of classification include the architecture of the network (feedforward versus recurrent), time model (discrete versus continuous), state type (binary versus analog), weight constraints (symmetric versus asymmetric), network size (finite nets versus infinite families), and computation type (deterministic versus probabilistic), among others. The underlying results concerning the computational power and complexity issues of perceptron, radial basis function, winner-take-all, and spiking neural networks are briefly surveyed, with pointers to the relevant literature. In our survey, we focus mainly on the digital computation whose inputs and outputs are binary in nature, although their values are quite often encoded as analog neuron states. We omit the important learning issues.

In the last few years several kinds of recurrent neural networks (RNNs) have been proposed for solving linear and nonlinear optimization problems. In this paper, we provide a survey of RNNs that can be used to solve both the constrained quadratic optimization problem related to support vector machine (SVM) learning, and the SVM model selection by automatic hyperparameter tuning. The appeal of this approach is the possibility of implementing such networks on analog VLSI systems with relative easiness. We review several proposals appeared so far in the literature and test their behavior when applied to solve a telecommunication application, where a special purpose adaptive hardware is of great interest.

Behavior may be controlled by reactive systems. In a reactive system the motor output is exclusively driven by actual sensory input. An alternative solution to control behavior is given by "cognitive" systems capable of planning ahead. To this end the system has to be equipped with some kind of internal world model. A sensible basis of an internal world model might be a model of the system's own body. I show that a reactive system with the ability to control a body of complex geometry requires only a slight reorganization to form a cognitive system. This implies that the assumption that the evolution of cognitive properties requires the introduction of new, additional modules, namely internal world models, is not justified. Rather, these modules may already have existed before the system obtained cognitive properties. Furthermore, I discuss whether the occurrence of such world models may lead to systems having internal perspective. (C) 2002 Cognitive Science Society, Inc. All rights reserved.

International audience; This paper presents an original link between neural networks theory and mechanical modeling networks. The problem is to find the parameters characterizing mechanical structures in order to reproduce given mechanical behaviors. Replacing "neural" units with mechanically based units and applying classical learning algorithms dedicated to supervised dynamic networks to these mechanical networks allows us to find the parameters for a physical model. Some new variants of real-time recurrent learning (RTRL) are also introduced, based on mechanical principles. The notion of interaction during learning is discussed at length and the results of tests are presented. Instead of the classical {machine learning system, environment} pair, we propose to study the {machine learning system, human operator, environment} triplet. Experiments have been carried out in simulated scenarios and some original experiments with a force-feedback device are also described.

Recurrent neural networks can simulate any finite state automata as well as any multi-stack Turing machine. When constraining the network architecture, however, this computational power may no longer hold. For example, recurrent cascade-correlation cannot simulate any finite state automata. Thus, it is important to assess the computational power of a given network architecture, since this characterizes the class of functions which, in principle, can be computed by it. We discuss the computational power of neural networks for structures. Elman-style networks, cascade-correlation networks and neural trees for structures are introduced. We show that Elman-style networks can simulate any frontier-to-root tree automation, while neither cascade-correlation networks nor neural trees can. As a special case of the latter result, we obtain that neural trees for sequences cannot simulate any finite state machine. © 1997 Elsevier Science Ltd. All Rights Reserved.

The biomedical field is characterized by an ever-increasing production of sequential data, which often come in the form of biosignals capturing the time-evolution of physiological processes, such as blood pressure and brain activity. This has motivated a large body of research dealing with the development of machine learning techniques for the predictive analysis of such biosignals. Unfortunately, in high-stakes decision making, such as clinical diagnosis, the opacity of machine learning models becomes a crucial aspect to be addressed in order to increase the trust and adoption of AI technology. In this paper, we propose a model agnostic explanation method, based on occlusion, that enables the learning of the input’s influence on the model predictions. We specifically target problems involving the predictive analysis of time-series data and the models that are typically used to deal with data of such nature, i.e., recurrent neural networks. Our approach is able to provide two different kinds of explanations: one suitable for technical experts, who need to verify the quality and correctness of machine learning models, and one suited to physicians, who need to understand the rationale underlying the prediction to make aware decisions. A wide experimentation on different physiological data demonstrates the effectiveness of our approach both in classification and regression tasks.