Search

Inspired by more detailed modeling of biological neurons, Spiking neural networks (SNNs) have been investigated both as more biologically plausible and potentially more powerful models of neural computation, and also with the aim of extracting biological neurons' energy efficiency; the performance of such networks however has remained lacking compared to classical artificial neural networks (ANNs). Here, we demonstrate how a novel surrogate gradient combined with recurrent networks of tunable and adaptive spiking neurons yields state-of-the-art for SNNs on challenging benchmarks in the time-domain, like speech and gesture recognition. This also exceeds the performance of standard classical recurrent neural networks (RNNs) and approaches that of the best modern ANNs. As these SNNs exhibit sparse spiking, we show that they theoretically are one to three orders of magnitude more computationally efficient compared to RNNs with comparable performance. Together, this positions SNNs as an attractive solution for AI hardware implementations., Comment: 11 pages

Neural computation has evolved to optimize the behaviors that enable our survival. Although much previous work in neuroscience has focused on constrained task behaviors, recent advances in computer vision are fueling a trend toward the study of naturalistic behaviors. Automated tracking of fine-scale behaviors is generating rich datasets for animal models including rodents, fruit flies, zebrafish, and worms. However, extracting meaning from these large and complex data often requires sophisticated computational techniques. Here we review the latest methods and modeling approaches providing new insights into the brain from behavior. We focus on unsupervised methods for identifying stereotyped behaviors and for resolving details of the structure and dynamics of behavioral sequences.

Spike trains (STs) induced by external stimuli are complex and challenging to decode the embedded spatiotemporal information. Although various methods have been developed to characterize STs, few applications have been reported in tactile discrimination applications. In this paper, a neurocomputing method based on reference spike train (RST) is proposed to establish a neural computation scheme, on which existing ST metrics could be fed into various traditional models directly. Moreover, existing metrics in the field of statistics and vector measurement are introduced together to extract more discriminative features. With the binning technique and feature selection algorithm applied, the neural computation scheme is aimed at taking advantage of as maximal as possible information contained in tactile signals. Based on our designed artificial fingertip, the effect is validated by improving the recognition accuracy from 77.6% to 83.4% when it is applied to the discrimination of eight roughness surfaces. Furthermore, properties of RST, such as spike intervals and distributions, are evaluated and it is found that RSTs with uniform distribution perform the best for tactile discrimination.

Uncertainty is a fundamental property of neural computation that becomes amplified when sensory information does not match a person’s expectations of the world. Uncertainty and hesitation are often early indicators of potential disruption, and the ability to rapidly measure uncertainty would have implications for future educational and training efforts by targeting reflective discussions about past actions, supporting in-progress corrections, and generating forecasts about future disruptions. An approach is described combining neurodynamics and machine learning to provide quantitative measures of uncertainty. Models of neurodynamic information derived from electroencephalogram (EEG) brainwaves have provided detailed neurodynamic histories of US Navy submarine navigation team members. Persistent periods (25–30 s) of neurodynamic information were seen as discrete peaks when establishing the submarine’s position and were identified as periods of uncertainty by an artificial intelligence (AI) system previously trained to recognize the frequency, magnitude, and duration of different patterns of uncertainty in healthcare and student teams. Transition matrices of neural network states closely predicted the future uncertainty of the navigation team during the three minutes prior to a grounding event. These studies suggest that the dynamics of uncertainty may have common characteristics across teams and tasks and that forecasts of their short-term evolution can be estimated.

Though the temporal precision of neural computation has been studied intensively, a data-driven determination of this precision remains a fundamental challenge. Reproducible spike time patterns may be obscured on single trials by uncontrolled temporal variability in behavior and cognition, or may not even be time locked to measurable signatures in either behavior or local field potentials (LFP). To overcome these challenges, we describe a general-purpose time warping framework that reveals precise spike-time patterns in an unsupervised manner, even when spiking is decoupled from behavior or is temporally stretched across single trials. We demonstrate this method across diverse systems: cued reaching in nonhuman primates, motor sequence production in rats, and olfaction in mice. This approach flexibly uncovers diverse dynamical firing patterns, including pulsatile responses to behavioral events, LFP-aligned oscillatory spiking, and even unanticipated patterns, like 7 Hz oscillations in rat motor cortex that are not time-locked to measured behaviors or LFP.

Understanding the mechanisms of neural computation and learning will require knowledge of the underlying circuitry. Because it is difficult to directly measure the wiring diagrams of neural circuits, there has long been an interest in estimating them algorithmically from multicell activity recordings. We show that even sophisticated methods, applied to unlimited data from every cell in the circuit, are biased toward inferring connections between unconnected but highly correlated neurons. This failure to ‘explain away’ connections occurs when there is a mismatch between the true network dynamics and the model used for inference, which is inevitable when modeling the real world. Thus, causal inference suffers when variables are highly correlated, and activity-based estimates of connectivity should be treated with special caution in strongly connected networks. Finally, performing inference on the activity of circuits pushed far out of equilibrium by a simple low-dimensional suppressive drive might ameliorate inference bias. The authors demonstrate that strongly recurrent circuits inferred from neural activity, even with unlimited data from every neuron, are biased. Synapses are inferred between unconnected but correlated neurons. Inference based on non-equilibrium activity may help remedy this.

Spiking neural networks (SNNs) are believed to be a powerful neural computation framework inspired by the vivo neurons. As a class of recurrent SNNs, liquid state machines (LSMs) are biologically more plausible models imitating the architecture and functions of the human brain for information processing. However, few LSM models can outperform conventional analogue neural networks for solving real-world classification or regression problems, which can mainly be attributed to the sensitivity of the training performance to the architecture of the reservoir and the parameters in the spiking neuron models. Most recently, many algorithms have been proposed for automated machine learning that aims to automatically design the architecture and parameters of deep neural networks without much human intervention. Although automated machine learning and neural architecture search have been extremely successful in conventional neural networks, little research on search for an optimal architecture and hyperparameters of LSMs has been reported. This work proposes on a surrogate-assisted evolutionary search method for optimization of the hyperparameters and neural architecture of the reservoir of LSMs using the covariance matrix adaptation evolution strategy (CMA-ES). For reducing the search space, the architecture of the LSM is encoded by a connectivity probability together with the hyperparameters in the spiking neuron models. To enhance the computational efficiency, a Gaussian process is adopted as the surrogate to assist the CMA-ES. The proposed GP-assisted CMA-ES is compared with the canonical CMA-ES and a Bayesian optimization algorithm on two popular datasets including image and action recognition. Our results confirm that the proposed algorithm is efficient and effective in optimizing the parameters and architecture of LSMs.

Toward the long-standing dream of artificial intelligence, two successful solution paths have been paved: 1) neuromorphic computing and 2) deep learning. Recently, they tend to interact for simultaneously achieving biological plausibility and powerful accuracy. However, models from these two domains have to run on distinct substrates, i.e., neuromorphic platforms and deep learning accelerators, respectively. This architectural incompatibility greatly compromises the modeling flexibility and hinders promising interdisciplinary research. To address this issue, we build a unified model description framework and a unified processing architecture (Tianjic), which covers the full stack from software to hardware. By implementing a set of integration and transformation operations, Tianjic is able to support spiking neural networks, biological dynamic neural networks, multilayered perceptron, convolutional neural networks, recurrent neural networks, and so on. A compatible routing infrastructure enables homogeneous and heterogeneous scalability on a decentralized many-core network. Several optimization methods are incorporated, such as resource and data sharing, near-memory processing, compute/access skipping, and intra-/inter-core pipeline, to improve performance and efficiency. We further design streaming mapping schemes for efficient network deployment with a flexible tradeoff between execution throughput and resource overhead. A 28-nm prototype chip is fabricated with >610-GB/s internal memory bandwidth. A variety of benchmarks are evaluated and compared with GPUs and several existing specialized platforms. In summary, the fully unfolded mapping can achieve significantly higher throughput and power efficiency; the semi-folded mapping can save 30x resources while still presenting comparable performance on average. Finally, two hybrid-paradigm examples, a multimodal unmanned bicycle and a hybrid neural network, are demonstrated to show the potential of our unified architecture. This article paves a new way to explore neural computing.

Many animals use an internal sense of direction to guide their movements through the world. Neurons selective to head direction are thought to support this directional sense and have been found in a diverse range of species, from insects to primates, highlighting their evolutionary importance. Across species, most head-direction networks share four key properties: a unique representation of direction at all times, persistent activity in the absence of movement, integration of angular velocity to update the representation, and the use of directional cues to correct drift. The dynamics of theorized network structures called ring attractors elegantly account for these properties, but their relationship to brain circuits is unclear. Here, we review experiments in rodents and flies that offer insights into potential neural implementations of ring attractor networks. We suggest that a theory-guided search across model systems for biological mechanisms that enable such dynamics would uncover general principles underlying head-direction circuit function.

Since the first quantum neural network based on quantum computing was proposed by famous scholar Kak, much attention has been taken focus on designing new quantum neural network models. In this paper, a novel efficient quantum feed-forward neural network based on quantum computing is established, which adopts genetic algorithm to improve the traditional back propagation algorithm in parameters learning process. We clearly show the mathematical process of the new proposed quantum network model and improved algorithm. The experimental results of MATLAB simulations show that the new network model which makes the best use of fast quantum neural computation does a better job in function approximation and prediction of educational short video’s spreading capacity than traditional back propagation neural network, and the improved algorithm is more efficient than common back propagation algorithm in the proposed quantum network model. Our model can be widely used in weather prediction, handwriting recognition, speech recognition, and other aspects.

1.AbstractAt ultra-high field, fMRI voxels can span the sub-millimeter range, allowing the recording of blood oxygenation level dependent (BOLD) responses at the level of fundamental units of neural computation, such as cortical columns and layers. This sub-millimeter resolution, however, is only nominal in nature as a number of factors limit the spatial acuity of functional voxels. Multivoxel Pattern Analysis (MVPA) may provide a means to detect information at finer spatial scales that may otherwise not be visible at the single voxel level due to limitations in sensitivity and specificity. Here, we evaluate the spatial scale of stimuli specific BOLD responses in multivoxel patterns exploited by linear Support Vector Machine, Linear Discriminant Analysis and Naïve Bayesian classifiers across cortical depths in V1. To this end, we artificially misaligned the testing relative to the training portion of the data in increasing spatial steps, then investigated the breakdown of the classifiers’ performances. A one voxel shift led to a significant decrease in decoding accuracy (p

Efficient data collection is an important goal in cognitive neuroimaging studies because of the high cost of data acquisition. One method of improving efficiency is to maximize the informativeness of the data collected on each trial. We propose an Adaptive Design Optimization (Cavagnaro et al. Neural Computation 22, 887–905 2010; Myung et al. Journal of Mathematical Psychology 57, 53–67 2013) procedure to optimize the sequencing of stimuli for model-based functional neuroimaging studies. Our method uses the Joint Modeling Framework (Turner et al. NeuroImage 72, 193–206 2013, 2019) to maximize the information learned about how the brain produces a behavior by integrating over neural and behavioral data simultaneously. We validate our method in simulation and real-world experiments by showing how Adaptive Design Optimization proposes the optimal stimulus sequence to reduce uncertainty and improve accuracy from a Bayesian perspective.

According to the critical brain hypothesis, the brain is considered to operate near criticality and realize efficient neural computations. Despite the prior theoretical and empirical evidence in favor of the hypothesis, no direct link has been provided between human cognitive performance and the neural criticality. Here we provide such a key link by analyzing resting-state dynamics of functional magnetic resonance imaging (fMRI) networks at a whole-brain level. We develop a data-driven analysis method, inspired from statistical physics theory of spin systems, to map out the whole-brain neural dynamics onto a phase diagram. Using this tool, we show evidence that neural dynamics of human participants with higher fluid intelligence quotient scores are closer to a critical state, i.e., the boundary between the paramagnetic phase and the spin-glass (SG) phase. The present results are consistent with the notion of “edge-of-chaos” neural computation., Ezaki et al. develop a computational tool to analyze neural resting-state dynamics of functional magnetic resonance imaging data. Their data from adult humans suggest that the ability to think logically and find solutions improves with the brain located closer to criticality.

Learning to select appropriate actions based on their values is fundamental to adaptive behavior. This form of learning is supported by fronto-striatal systems. The dorsal-lateral prefrontal cortex (dlPFC) and the dorsal striatum (dSTR), which are strongly interconnected, are key nodes in this circuitry. Substantial experimental evidence, including neurophysiological recordings, have shown that neurons in these structures represent key aspects of learning. The computational mechanisms that shape the neurophysiological responses, however, are not clear. To examine this, we developed a recurrent neural network (RNN) model of the dlPFC-dSTR circuit and trained it on an oculomotor sequence learning task. We compared the activity generated by the model to activity recorded from monkey dlPFC and dSTR in the same task. This network consisted of a striatal component which encoded action values, and a prefrontal component which selected appropriate actions. After training, this system was able to autonomously represent and update action values and select actions, thus being able to closely approximate the representational structure in corticostriatal recordings. We found that learning to select the correct actions drove action-sequence representations further apart in activity space, both in the model and in the neural data. The model revealed that learning proceeds by increasing the distance between sequence-specific representations. This makes it more likely that the model will select the appropriate action sequence as learning develops. Our model thus supports the hypothesis that learning in networks drives the neural representations of actions further apart, increasing the probability that the network generates correct actions as learning proceeds. Altogether, this study advances our understanding of how neural circuit dynamics are involved in neural computation, showing how dynamics in the corticostriatal system support task learning.

Neural coding, including encoding and decoding, is one of the key problems in neuroscience for understanding how the brain uses neural signals to relate sensory perception and motor behaviors with neural systems. However, most of the existed studies only aim at dealing with the continuous signal of neural systems, while lacking a unique feature of biological neurons, termed spike, which is the fundamental information unit for neural computation as well as a building block for brain-machine interface. Aiming at these limitations, we propose a transcoding framework to encode multi-modal sensory information into neural spikes and then reconstruct stimuli from spikes. Sensory information can be compressed into 10% in terms of neural spikes, yet re-extract 100% of information by reconstruction. Our framework can not only feasibly and accurately reconstruct dynamical visual and auditory scenes, but also rebuild the stimulus patterns from functional magnetic resonance imaging (fMRI) brain activities. More importantly, it has a superb ability of noise immunity for various types of artificial noises and background signals. The proposed framework provides efficient ways to perform multimodal feature representation and reconstruction in a high-throughput fashion, with potential usage for efficient neuromorphic computing in a noisy environment.

Convolutional neural networks trained on object recognition derive inspiration from the neural architecture of the visual system in primates, and have been used as models of the feedforward computation performed in the primate ventral stream. In contrast to the deep hierarchical organization of primates, the visual system of the mouse has a shallower arrangement. Since mice and primates are both capable of visually guided behavior, this raises questions about the role of architecture in neural computation. In this work, we introduce a novel framework for building a biologically constrained convolutional neural network model of the mouse visual cortex. The architecture and structural parameters of the network are derived from experimental measurements, specifically the 100-micrometer resolution interareal connectome, the estimates of numbers of neurons in each area and cortical layer, and the statistics of connections between cortical layers. This network is constructed to support detailed task-optimized models of mouse visual cortex, with neural populations that can be compared to specific corresponding populations in the mouse brain. Using a well-studied image classification task as our working example, we demonstrate the computational capability of this mouse-sized network. Given its relatively small size, MouseNet achieves roughly 2/3rds the performance level on ImageNet as VGG16. In combination with the large scale Allen Brain Observatory Visual Coding dataset, we use representational similarity analysis to quantify the extent to which MouseNet recapitulates the neural representation in mouse visual cortex. Importantly, we provide evidence that optimizing for task performance does not improve similarity to the corresponding biological system beyond a certain point. We demonstrate that the distributions of some physiological quantities are closer to the observed distributions in the mouse brain after task training. We encourage the use of the MouseNet architecture by making the code freely available.Author summaryTask-driven deep neural networks have shown great potential in predicting functional responses of biological neurons. Nevertheless, they are not precise biological analogues, raising questions about how they should be interpreted. Here, we build new deep neural network models of the mouse visual cortex (MouseNet) that are biologically constrained in detail, not only in terms of the basic structure of their connectivity, but also in terms of the count and hence density of neurons within each area, and the spatial extent of their projections. Equipped with the MouseNet model, we can address key questions about mesoscale brain architecture and its role in task learning and performance.We ask, and provide a first set of answers, to: What is the performance of a mouse brain-sized – and mouse brain-structured – model on benchmark image classification tasks? How does the training of a network on this task affect the functional properties of specified layers within the biologically constrained architecture – both overall, and in comparison with recorded function of mouse neurons? We anticipate much future work on allied questions, and the development of more sophisticated models in both mouse and other species, based on the freely available MouseNet model and code which we develop and provide here.

Also part of the Information Systems and Applications, incl. Internet/Web, and HCI book sub series (LNISA, volume 12925); International audience; The problem of identifying the degree of semantic similarity between two textual statements automatically has grown in importance in recent times. Its impact on various computer-related domains and recent breakthroughs in neural computation has increased the opportunities for better solutions to be developed. This research takes the research efforts a step further by designing and developing a novel neurofuzzy approach for semantic textual similarity that uses neural networks and fuzzy logics. The fundamental notion is to combine the remarkable capabilities of the current neural models for working with text with the possibilities that fuzzy logic provides for aggregating numerical information in a tailored manner. The results of our experiments suggest that this approach is capable of accurately determining semantic textual similarity.

Το αντικείμενο της διδακτορικής διατριβής αϕορά στην ανάπτυξη και στη θεμελίωση μεθόδων για αξιόπιστους νευρωνικούς υπολογισμούς. Ένα σημαντικό τμήμα της διατριβής είναι αφιερωμένο στη μαθηματική θεμελίωση των μεθόδων. Ειδικότερα, η διατριβή εστιάζει στη διαχείριση της αβεβαιότητας που είναι σύμϕυτη με την επεξεργασία των δεδομένων δηλαδή με τους αλγορίθμους επεξεργασίας. Τα ζητήματα που αναδεικνύουν και χαρακτηρίζουν την αβεβαιότητα σε επίπεδο επεξεργασίας των δεδομένων σχετίζονται με τις αρχικές και τις οριακές τιμές των αλγορίθμων, τη σθεναρότητα των αλγορίθμων σε σϕάλματα των υπολογισμών ή των δεδομένων, την πληθώρα των ευρετικών που υποκαθιστούν τη μαθηματική μοντελοποίηση άγνωστων παραμέτρων κλπ. Για τη μοντελοποίηση και την επίλυση ζητημάτων, όπως τα ανωτέρω, χρησιμοποιήθηκαν κύρια οι έννοιες και οι μέθοδοι της Ανάλυσης Διαστημάτων καθώς και μαθηματικές προσεγγίσεις όπως η ολική βελτιστοποίηση και η μηδενική στάθμη διανυσμάτων. Επί πλέον, αξιοποιήθηκαν κάποιες τεχνικές ομαδοποίησης και εξόρυξης δεδομένων για συγκεκριμένες περιπτώσεις ανάλυσης δεδομένων. Στα πλαίσια της διατριβής μελετώνται, κυρίως, πολυστρωματικά δίκτυα τύπου perceptron. Η διατύπωση αξιόπιστων αλγορίθμων σχετίζεται με τη διαχείριση της αβεβαιότητας στα ακόλουθα ζητήματα: α) αρχικοποίηση των συναπτικών βαρών, β) βελτιστοποίηση της σχεδίασης του δικτύου, γ) καθορισμός του χώρου εντός του οποίου εγγυημένα βρίσκεται κάποιος ολικός ελάχιστοποιητής της συνάρτησης κόστους της εξόδου του δικτύου, και δ) αξιόπιστη εκτίμηση του πεδίου εγκυρότητας ενός δικτύου.

A central goal of artificial intelligence is to design algorithms that are both generalisable and interpretable. We combine brain-inspired neural computation principles and scalable deep learning architectures to design compact neural controllers for task-specific compartments of a full-stack autonomous vehicle control system. We show that a single algorithm with 19 control neurons, connecting 32 encapsulated input features to outputs by 253 synapses, learns to map high-dimensional inputs into steering commands. This system shows superior generalisability, interpretability and robustness compared with orders-of-magnitude larger black-box learning systems. The obtained neural agents enable high-fidelity autonomy for task-specific parts of a complex autonomous system.

This paper aims at addressing a challenging problem on reliably estimating image motion against highly variable natural signals which artificial dynamic vision systems are faced with. Previously, the visual system response always represents fluctuation and high variance influenced by spatial contrast, the local difference between neighbouring luminance values. Effective contrast computation is therefore a prerequisite for robust motion vision. In this regard, sighted animals such as flies are remarkably adept at estimating image motion regardless of image statistics by rapidly adjusting contrast sensitivity. Current artificial visual systems, however, cannot account for this capability. Learning from neuroscience, here we propose contrast vision computation to improve a state-of-the-art, bio-inspired neural network model for background motion estimation. This includes mainly two neural computation schemes of (1) an instantaneous, feedback divisive contrast normalisation prior to motion correlation in the ON and OFF pathways to reduce local contrast sensitivity, (2) parallel contrast pathways influencing ON/OFF motion signals, negatively, at the pooling output layer to suppress high-contrast optic flows. We created a dataset of many shifting natural images with high input variability to investigate the proposed method. The experiments have demonstrated the effectiveness and robustness of the proposed contrast vision computation to reduce response fluctuation and variance against natural signals. The fidelity of motion perception thus has been significantly increased. The proposed methods could be generic to other motion vision models dealing with high-contrast visual scenes.

1AbstractA cornerstone of theoretical neuroscience is the circuit model: a system of equations that captures a hypothesized neural mechanism. Such models are valuable when they give rise to an experimentally observed phenomenon – whether behavioral or a pattern of neural activity – and thus can offer insights into neural computation. The operation of these circuits, like all models, critically depends on the choice of model parameters. A key step is then to identify the model parameters consistent with observed phenomena: to solve the inverse problem. In this work, we present a novel technique, emergent property inference (EPI), that brings the modern probabilistic modeling toolkit to theoretical neuroscience. When theorizing circuit models, theoreticians predominantly focus on reproducing computational properties rather than a particular dataset. Our method uses deep neural networks to learn parameter distributions with these computational properties. This methodology is introduced through a motivational example inferring conductance parameters in a circuit model of the stomatogastric ganglion. Then, with recurrent neural networks of increasing size, we show that EPI allows precise control over the behavior of inferred parameters, and that EPI scales better in parameter dimension than alternative techniques. In the remainder of this work, we present novel theoretical findings gained through the examination of complex parametric structure captured by EPI. In a model of primary visual cortex, we discovered how connectivity with multiple inhibitory subtypes shapes variability in the excitatory population. Finally, in a model of superior colliculus, we identified and characterized two distinct regimes of connectivity that facilitate switching between opposite tasks amidst interleaved trials, characterized each regime via insights afforded by EPI, and found conditions where these circuit models reproduce results from optogenetic silencing experiments. Beyond its scientific contribution, this work illustrates the variety of analyses possible once deep learning is harnessed towards solving theoretical inverse problems.

There are rich structures in off-task neural activity which are hypothesized to reflect fundamental computations across a broad spectrum of cognitive functions. Here, we develop an analysis toolkit – temporal delayed linear modelling (TDLM) – for analysing such activity. TDLM is a domain-general method for finding neural sequences that respect a pre-specified transition graph. It combines nonlinear classification and linear temporal modelling to test for statistical regularities in sequences of task-related reactivations. TDLM is developed on the non-invasive neuroimaging data and is designed to take care of confounds and maximize sequence detection ability. Notably, as a linear framework, TDLM can be easily extended, without loss of generality, to capture rodent replay in electrophysiology, including in continuous spaces, as well as addressing second-order inference questions, for example, its temporal and spatial varying pattern. We hope TDLM will advance a deeper understanding of neural computation and promote a richer convergence between animal and human neuroscience.

In this paper, the computational performance of a Spiking Feed-forward Neural Network (SFNN) is investigated based on a brain-inspired Intrinsic Plasticity (IP) mechanism, which is a membrane potential adaptive tuning scheme used to change the intrinsic excitability of individual neuron. This learning rule has the ability of regulating neural activity in a relative homeostatic level even if the external input of a neuron is extremely low or extremely high. The effectiveness of IP on SFNN model has been studied and evaluated through the MNIST handwritten digits classification. The training of network weights starts from a conventional artificial neural network by backpropagation and then the rate-based neurons are transformed into spiking neuron models with IP learning. Our results show that both over-activation and under-activation of neuronal response which commonly exist during the computation of neural networks can be effectively avoided. Without loss of accuracy, the calculation speed of SFNN with IP learning is extremely higher than that of the other models. Besides, when the input intensity and data noise are taken into account, both of the learning speed and accuracy of the model can be greatly improved by the application of IP learning. This biologically inspired SFNN model is simple and effective which may give insights for the optimization of neural computation.

Numerous efforts to generate "connectomes," or synaptic wiring diagrams, of large neural circuits or entire nervous systems are currently underway. These efforts promise an abundance of data to guide theoretical models of neural computation and test their predictions. However, there is not yet a standard set of tools for incorporating the connectivity constraints that these datasets provide into the models typically studied in theoretical neuroscience. This article surveys recent approaches to building models with constrained wiring diagrams and the insights they have provided. It also describes challenges and the need for new techniques to scale these approaches to ever more complex datasets.

Visual perception is an important way for organisms to perceive the external world. Simulation of visual cognitive process can enhance the cognitive ability of machine vision. Therefore, how to simulate visual perception system and make computer have a high world understanding ability is a hotspot of current neurocomputing. Based on the information processing mechanism of visual perception system, this paper establishes a neural computing model based on visual perception mechanism. The simulation results of standard face database and natural landscape images show that the proposed method can recognize face samples better when other noise samples are added to the face image samples. In landscape contour fitting simulation, the results show that although this method has little advantage for large contour image recognition, but for small contour recognition, this method is obviously superior to other methods.

The ability to compute the location and direction of sounds is a crucial perceptual skill to efficiently interact with dynamic environments. How the human brain implements spatial hearing is however poorly understood. In our study, we used fMRI to characterize the brain activity of male and female humans listening to left, right, up and down moving as well as static sounds. Whole brain univariate results contrasting moving and static sounds varying in their location revealed a robust functional preference for auditory motion in bilateral human Planum Temporale (hPT). Using independently localized hPT, we show that this region contains information about auditory motion directions and, to a lesser extent, sound source locations. Moreover, hPT showed an axis of motion organization reminiscent of the functional organization of the middle-temporal cortex (hMT+/V5) for vision. Importantly, whereas motion direction and location rely on partially shared pattern geometries in hPT, as demonstrated by successful cross-condition decoding, the responses elicited by static and moving sounds were however significantly distinct. Altogether our results demonstrate that the hPT codes for auditory motion and location but that the underlying neural computation linked to motion processing is more reliable and partially distinct from the one supporting sound source location.SIGNIFICANCE STATEMENTIn comparison to what we know about visual motion, little is known about how the brain implements spatial hearing. Our study reveals that motion directions and sound source locations can be reliably decoded in the human Planum Temporale (hPT) and that they rely on partially shared pattern geometries. Our study therefore sheds important new lights on how computing the location or direction of sounds are implemented in the human auditory cortex by showing that those two computations rely on partially shared neural codes. Furthermore, our results show that the neural representation of moving sounds in hPT follows a “preferred axis of motion” organization, reminiscent of the coding mechanisms typically observed in the occipital hMT+/V5 region for computing visual motion.

Deep neural networks (DNN) have set new standards at predicting responses of neural populations to visual input. Most such DNNs consist of a convolutional network (core) shared across all neurons which learns a representation of neural computation in visual cortex and a neuron-specific readout that linearly combines the relevant features in this representation. The goal of this paper is to test whether such a representation is indeed generally characteristic for visual cortex, i.e. gener-alizes between animals of a species, and what factors contribute to obtaining such a generalizing core. To push all non-linear computations into the core where the generalizing cortical features should be learned, we devise a novel readout that reduces the number of parameters per neuron in the readout by up to two orders of magnitude compared to the previous state-of-the-art. It does so by taking advantage of retinotopy and learns a Gaussian distribution over the neuron’s receptive field po-sition. With this new readout we train our network on neural responses from mouse primary visual cortex (V1) and obtain a gain in performance of 7% compared to the previous state-of-the-art network. We then investigate whether the convolutional core indeed captures general cortical features by using the core in transfer learning to a different animal. When transferring a core trained on thousands of neurons from various animals and scans we exceed the performance of training directly on that animal by 12%, and outperform a commonly used VGG16 core pre-trained on imagenet by 33%. In addition, transfer learning with our data-driven core is more data-efficient than direct training, achieving the same performance with only 40% of the data. Our model with its novel readout thus sets a new state-of-the-art for neural response prediction in mouse visual cortex from natural images, generalizes between animals, and captures better characteristic cortical features than current task-driven pre-training approaches such as VGG16.

The question of how the collective activity of neural populations gives rise to complex behaviour is fundamental to neuroscience. At the core of this question lie considerations about how neural circuits can perform computations that enable sensory perception, decision making, and motor control. It is thought that such computations are implemented through the dynamical evolution of distributed activity in recurrent circuits. Thus, identifying dynamical structure in neural population activity is a key challenge towards a better understanding of neural computation. At the same time, interpreting this structure in light of the computation of interest is essential for linking the time-varying activity patterns of the neural population to ongoing computational processes. Here, we review methods that aim to quantify structure in neural population recordings through a dynamical system defined in a low-dimensional latent variable space. We discuss advantages and limitations of different modelling approaches and address future challenges for the field.

Neuromorphic computing currently relies heavily on complicated hardware design to implement asynchronous, parallel and very large-scale brain simulations. This dependency slows down the migration of biological insights into technology. It typically takes several years from idea to finished hardware and once developed the hardware is not broadly available to the community. In this contribution, we present the CloudBrain research platform, an alternative based on modern cloud computing and event stream processing technology. Typical neuromorphic design goals, such as small form factor and low power consumption, are traded for 1) no constraints on the model elements, 2) access to all events and parameters during and after the simulation, 3) online reconfiguration of the network, and 4) real-time simulation. We explain principles for how neuron, synapse and network models can be implemented and we demonstrate that our implementation can be used to control a physical robot in real-time. CloudBrain is open source and can run on commodity hardware or in the cloud, thus providing the community a new platform with a different set of features supporting research into, for example, neuron models, structural plasticity and three-factor learning.

Producción Científica, Since two decades ago, research on resistive memories has continuosly grown, gathering relevance through the variety of different technologies that fit into the non-volatile memories’ area. In this study, we discuss the performance and electrical characteristics of RRAM cells constituted by MIM stacks with dielectric formed by Nb2O5-doped SiO2. We report experimental results that show a clear improvement in the resistive behavior of the devices and an excellent analogical control of the intermediate levels between high-resistance and low-resistance states., Ministerio de Ciencia, Innovación y Universidades con el apoyo de fondos Feder (grant TEC2017-84321-C4-2-R), Finnish Centre of Excellence in Atomic Layer Deposition (284623), Estonian Research Agency (PRG753)

Convolutional neural networks trained on object recognition derive some inspiration from the neuroscience of the visual system in primates, and have been used as models of the feedforward computation performed in the primate ventral stream. In contrast to the hierarchical organization of primates, the visual system of the mouse has flatter hierarchy. Since mice are capable of visually guided behavior, this raises questions about the role of architecture in neural computation. In this work, we introduce a framework for building a biologically constrained convolutional neural network model of lateral areas of the mouse visual cortex. The structural parameters of the network are derived from experimental measurements, specifically estimates of numbers of neurons in each area and cortical layer, the interareal connec-tome, and the statistics of connections between cortical layers. This network is constructed to support detailed task-optimized models of mouse visual cortex, with neural populations that can be compared to specific corresponding populations in the mouse brain. The code is freely available to support such research.

Our real-time actions in everyday life reflect a range of spatiotemporal dynamic brain activity patterns, the consequence of neuronal computation with spikes in the brain. Most existing models with spiking neurons aim at solving static pattern recognition tasks such as image classification. Compared with static features, spatiotemporal patterns are more complex due to their dynamics in both space and time domains. Spatiotemporal pattern recognition based on learning algorithms with spiking neurons therefore remains challenging. We propose an end-to-end recurrent spiking neural network model trained with an algorithm based on spike latency and temporal difference backpropagation. Our model is a cascaded network with three layers of spiking neurons where the input and output layers are the encoder and decoder, respectively. In the hidden layer, the recurrently connected neurons with transmission delays carry out high-dimensional computation to incorporate the spatiotemporal dynamics of the inputs. The test results based on the data sets of spiking activities of the retinal neurons show that the proposed framework can recognize dynamic spatiotemporal patterns much better than using spike counts. Moreover, for 3D trajectories of a human action data set, the proposed framework achieves a test accuracy of 83.6% on average. Rapid recognition is achieved through the learning methodology–based on spike latency and the decoding process using the first spike of the output neurons. Taken together, these results highlight a new model to extract information from activity patterns of neural computation in the brain and provide a novel approach for spike-based neuromorphic computing.

International audience; Recent literature considers that Spiking Neural Networks are now a serious alternative of Formal Neural Networksfor embedded artificial intelligence. The changes in the information coding and the elementary neural computation makethem more efficient than FNNs in terms of power consumption and chip surface occupation. However, these results are often based on simple neural network topologies with basic data-sets. In this paper, we study the behavior of Spiking Convolutional Neural Networks when applied to two different classification tasks. To do so, we analyze the spiking activity on both MNIST and GTSRB data-sets using different rate-based and temporal coding schemes. Notably, the Spike Select method is confronted to First Spike and Jittered Periodic methods in terms of prediction accuracy and spiking activity. Finally, we conclude about spike generation within spiking CNNs for embedded hardware classification.

In temporal lobe epilepsy, the ability of the dentate gyrus to limit excitatory cortical input to the hippocampus breaks down, leading to seizures. The dentate gyrus is also thought to help discriminate between similar memories by performing pattern separation, but whether epilepsy leads to a breakdown in this neural computation, and thus to mnemonic discrimination impairments, remains unknown. Here we show that temporal lobe epilepsy is characterized by behavioral deficits in mnemonic discrimination tasks, in both humans (females and males) and mice (C57Bl6 males, systemic low-dose kainate model). Using a recently developed assay in brain slices of the same epileptic mice, we reveal a decreased ability of the dentate gyrus to perform certain forms of pattern separation. This is due to a subset of granule cells with abnormal bursting that can develop independently of early EEG abnormalities. Overall, our results linking physiology, computation and cognition in the same mice advance our understanding of episodic memory mechanisms and their dysfunction in epilepsy.Significance StatementPeople with temporal lobe epilepsy (TLE) often have learning and memory impairments, sometimes occurring earlier than the first seizure, but those symptoms and their biological underpinnings are poorly understood. We focused on the dentate gyrus, a brain region that is critical to avoid confusion between similar memories and is anatomically disorganized in TLE. We show that both humans and mice with TLE experience confusion between similar situations. This impairment coincides with a failure of the dentate gyrus to disambiguate similar input signals because of pathological bursting in a subset of neurons. Our work bridges seizure-oriented and memory-oriented views of the dentate gyrus function, suggests a mechanism for cognitive symptoms in TLE and supports a long-standing hypothesis of episodic memory theories.

Machine learning (ML) architectures such as Deep Neural Networks (DNNs) have achieved unprecedented accuracy on modern applications such as image classification and speech recognition. With power dissipation becoming a major concern in ML architectures, computer architects have focused on designing both energy-efficient hardware platforms as well as optimizing ML algorithms. To dramatically reduce power consumption and increase parallelism in neural network accelerators, disruptive technology such as silicon photonics has been proposed which can improve the performance-per-Watt when compared to electrical implementation. In this paper, we propose PIXEL - Photonic Neural Network Accelerator that efficiently implements the fundamental operation in neural computation, namely the multiply and accumulate (MAC) functionality using photonic components such as microring resonators (MRRs) and Mach-Zehnder interferometer (MZI). We design two versions of PIXEL - a hybrid version that multiplies optically and accumulates electrically and a fully optical version that multiplies and accumulates optically. We perform a detailed power, area and timing analysis of the different versions of photonic and electronic accelerators for different convolution neural networks (AlexNet, VGG16, and others). Our results indicate a significant improvement in the energy-delay product for both PIXEL designs over traditional electrical designs (48.4% for OE and 73.9% for OO) while minimizing latency, at the cost of increased area over electrical designs.

From the Publisher: A hands-on guide to the design and implementation of neural networks -- A comprehensive source of reference for all neural network users, designers and implementers -- Provides an information pathway between scientists and engineers in different disciplines who apply neural networks to generically similar problems -- Offers access to timely information in a rapidly changing field -- World Wide Web access to the complete handbook available free to electronic edition subscribers In recent years, neural computation has developed from a specialized research discipline into a broadly based and dynamic activity with applications in an astonishing variety of fields. Many scientists, engineers and other practitioners are now using neural networks to tackle problems that are either intractable or unrealistically time consuming to solve through traditional computational strategies. The inaugural volume in the Computational Intelligence Library provides speeedy dissemination of new ideas to a broad spectrum of neural network users, designers and implementers. Devoted to network fundamentals, models, algorithms and applications, the work is intended to become the standard reference resource for the neural network community. As the field expands and develops, leading researchers will report on an analyze promising new approaches. In this way, the Handbook will become an evolving compendium on the state of the art of neural computation. Available in loose-leaf print form as well as in an electronic edition that combines both CD-ROM and on-line (World Wide Web) access to its contents, the Handbook of Neural Computation is available on a subscription basis, with regularly publishedsupplements keeping readers abreast of late-breaking developments and new advances in this rapidly developing field.

Neurons in the brain represent information in their collective activity. The fidelity of this neural population code depends on whether and how variability in the response of one neuron is shared with other neurons. Two decades of studies have investigated the influence of these noise correlations on the properties of neural coding. We provide an overview of the theoretical developments on the topic. Using simple, qualitative and general arguments, we discuss, categorize, and relate the various published results. We emphasize the relevance of the fine structure of noise correlation, and we present a new approach to the issue. Throughout we emphasize a geometrical picture of how noise correlations impact the neural code., 30 pages; 3 figures; review article

There are two main lines of research on visual question answering (VQA): compositional model with explicit multi-hop reasoning, and monolithic network with implicit reasoning in the latent feature space. The former excels in interpretability and compositionality but fails on real-world images, while the latter usually achieves better performance due to model flexibility and parameter efficiency. We aim to combine the two to build an interpretable framework for real-world compositional VQA. In our framework, images and questions are disentangled into scene graphs and programs, and a symbolic program executor runs on them with full transparency to select the attention regions, which are then iteratively passed to a visual-linguistic pre-trained encoder to predict answers. Experiments conducted on the GQA benchmark demonstrate that our framework outperforms the compositional prior arts and achieves competitive accuracy among monolithic ones. With respect to the validity, plausibility and distribution metrics, our framework surpasses others by a considerable margin.

Deep learning is a kind of machine learning which happens in a certain type of artificial neural networks called deep networks. Artificial deep networks, which exhibit many similarities with biological ones, have consistently shown human-like performance in many intelligent tasks. This poses the question whether this performance is caused by such similarities. After reviewing the structure and learning processes of artificial and biological neural networks, we outline two important reasons for the success of deep learning, namely the extraction of successively higher level features and the multiple layer structure, which are closely related to each other. Then some indications about the framing of this heated debate are given. After that, an assessment of the value of artificial deep networks as models of the human brain is given from the similarity perspective of model representation. Finally, a new version of computational functionalism is proposed which addresses the specificity of deep neural computation better than classic, program based computational functionalism.

Motion in the visual world provides critical information to guide the behavior of sighted animals. Furthermore, as visual motion estimation requires comparisons of signals across inputs and over time, it represents a paradigmatic and generalizable neural computation. Focusing on the Drosophila visual system, where an explosion of technological advances has recently accelerated experimental progress, we review our understanding of how, algorithmically and mechanistically, motion signals are first computed.

© 2017, The Natural Computing Applications Forum. This paper proposes a novel method for robust visual tracking of arbitrary objects, based on the combination of image-based prediction and position refinement by weighted correlation. The effectiveness of the proposed approach is demonstrated on a challenging set of dynamic video sequences, extracted from the final of triple jump at the London 2012 Summer Olympics. A comparison is made against five baseline tracking systems. The novel system shows remarkable superior performances with respect to the other methods, in all considered cases characterized by changing background, and a large variety of articulated motions. The novel architecture, from here onward named 2D Recurrent Neural Network (2D-RNN), is derived from the well-known recurrent neural network model and adopts nearest neighborhood connections between the input and context layers in order to store the temporal information content of the video. Starting from the selection of the object of interest in the first frame, neural computation is applied to predict the position of the target in each video frame. Normalized cross-correlation is then applied to refine the predicted target position. 2D-RNN ensures limited complexity, great adaptability and a very fast learning time. At the same time, it shows on the considered dataset fast execution times and very good accuracy, making this approach an excellent candidate for automated analysis of complex video streams.

Computationalism about the brain is the view that the brain literally performs computations. For the view to be interesting, we need an account of computation. The most well-developed account of computation is Turing Machine computation, the account provided by theoretical computer science which provides the basis for contemporary digital computers. Some have thought that, given the seemingly-close analogy between the all-or-nothing nature of neural spikes in brains and the binary nature of digital logic, neural computation could be a species of digital computation. A few recent authors have offered arguments against this idea; here, I review recent findings in neuroscience that further cement the implausibility of this view. However, I argue that we can retain the view that the brain is a computer if we expand what we mean by “computation” to include analog computation. I articulate an account of analog computation as the manipulation of analog representations based on previous work on the difference between analog and non-analog representations, extending a view originally articulated in Shagrir (Stud Hist Philos Sci 41(3):271–279, 2010). Given that analog computation constitutes a significant chapter in the history of computation, this revision of computationalism to include analog computation is not an ad hoc addition. Brains may well be computers, but of the analog kind, rather than the digital kind.

We show that Langevin Markov chain Monte Carlo inference in an energy-based model with latent variables has the property that the early steps of inference, starting from a stationary point, correspond to propagating error gradients into internal layers, similar to backpropagation. The backpropagated error is with respect to output units that have received an outside driving force pushing them away from the stationary point. Backpropagated error gradients correspond to temporal derivatives with respect to the activation of hidden units. These lead to a weight update proportional to the product of the presynaptic firing rate and the temporal rate of change of the postsynaptic firing rate. Simulations and a theoretical argument suggest that this rate-based update rule is consistent with those associated with spike-timing-dependent plasticity. The ideas presented in this article could be an element of a theory for explaining how brains perform credit assignment in deep hierarchies as efficiently as backpropagation does, with neural computation corresponding to both approximate inference in continuous-valued latent variables and error backpropagation, at the same time.

Deep neural networks (NNs) are the state-of-the-art models for understanding the content of images and videos. However, implementing deep NNs in embedded systems is a challenging task, e.g., a typical deep belief network could exhaust gigabytes of memory and result in bandwidth and computational bottlenecks. To address this challenge, this paper presents an algorithm and hardware codesign for efficient deep neural computation. A hardware-oriented deep learning algorithm, named the deep adaptive network, is proposed to explore the sparsity of neural connections. By adaptively removing the majority of neural connections and robustly representing the reserved connections using binary integers, the proposed algorithm could save up to 99.9% memory utility and computational resources without undermining classification accuracy. An efficient sparse-mapping-memory-based hardware architecture is proposed to fully take advantage of the algorithmic optimization. Different from traditional Von Neumann architecture, the deep-adaptive network on chip (DANoC) brings communication and computation in close proximity to avoid power-hungry parameter transfers between on-board memory and on-chip computational units. Experiments over different image classification benchmarks show that the DANoC system achieves competitively high accuracy and efficiency comparing with the state-of-the-art approaches.

Recent studies have demonstrated the effectiveness of supervised learning in spiking neural networks (SNNs). A trainable SNN provides a valuable tool not only for engineering applications but also for theoretical neuroscience studies. Here, we propose a modified SpikeProp learning algorithm, which ensures better learning stability for SNNs and provides more diverse network structures and coding schemes. Specifically, we designed a spike gradient threshold rule to solve the well-known gradient exploding problem in SNN training. In addition, regulation rules on firing rates and connection weights are proposed to control the network activity during training. Based on these rules, biologically realistic features such as lateral connections, complex synaptic dynamics, and sparse activities are included in the network to facilitate neural computation. We demonstrate the versatility of this framework by implementing three well-known temporal codes for different types of cognitive tasks, namely, handwritten digit recognition, spatial coordinate transformation, and motor sequence generation. Several important features observed in experimental studies, such as selective activity, excitatory-inhibitory balance, and weak pairwise correlation, emerged in the trained model. This agreement between experimental and computational results further confirmed the importance of these features in neural function. This work provides a new framework, in which various neural behaviors can be modeled and the underlying computational mechanisms can be studied.

In the electroencephalogram recorded data are often confounded with artifacts, especially in the case of eye blinks. Different methods for artifact detection and removal are discussed in the literature, including automatic detection and removal. Here, an automatic method of eye blink detection and correction is proposed where sparse coding is used for an electroencephalogram dataset. In this method, a hybrid dictionary based on a ridgelet transformation is used to capture prominent features by analyzing independent components extracted from a different number of electroencephalogram channels. In this study, the proposed method has been tested and validated with five different datasets for artifact detection and correction. Results show that the proposed technique is promising as it successfully extracted the exact locations of eye blinking artifacts. The accuracy of the method (automatic detection) is 89.6% which represents a better estimate than that obtained by an extreme machine learning classifier.

Reconstructing three-dimensional (3D) scenes from two-dimensional (2D) retinal images is an ill-posed problem. Despite this, 3D perception of the world based on 2D retinal images is seemingly accurate and precise. The integration of distinct visual cues is essential for robust 3D perception in humans, but it is unclear whether this is true for non-human primates (NHPs). Here, we assessed 3D perception in macaque monkeys using a planar surface orientation discrimination task., Reconstructing three-dimensional (3D) scenes from two-dimensional (2D) retinal images is an ill-posed problem. Despite this, 3D perception of the world based on 2D retinal images is seemingly accurate and precise. The integration of distinct visual cues is essential for robust 3D perception in humans, but it is unclear whether this is true for non-human primates (NHPs). Here, we assessed 3D perception in macaque monkeys using a planar surface orientation discrimination task. Perception was accurate across a wide range of spatial poses (orientations and distances), but precision was highly dependent on the plane’s pose. The monkeys achieved robust 3D perception by dynamically reweighting the integration of stereoscopic and perspective cues according to their pose-dependent reliabilities. Errors in performance could be explained by a prior resembling the 3D orientation statistics of natural scenes. We used neural network simulations based on 3D orientation-selective neurons recorded from the same monkeys to assess how neural computation might constrain perception. The perceptual data were consistent with a model in which the responses of two independent neuronal populations representing stereoscopic cues and perspective cues (with perspective signals from the two eyes combined using nonlinear canonical computations) were optimally integrated through linear summation. Perception of combined-cue stimuli was optimal given this architecture. However, an alternative architecture in which stereoscopic cues, left eye perspective cues, and right eye perspective cues were represented by three independent populations yielded two times greater precision than the monkeys. This result suggests that, due to canonical computations, cue integration for 3D perception is optimized but not maximized.