Spiking neural networks (SNNs) circumvent the need for large scale arithmetic using techniques inspired by biology.However, SNNs are designed with fundamentally different algorithms from ANNs, which have benefited from a rich history of theoretical advances and an increasingly mature software stack.In this paper we explore the potential of a new technique that lies between these two approaches, one that can leverage the software and system level optimizations of ANNs while utilizing biologically inspired circuits for energy efficient computation.The resulting hardware represents the traditional weight of an ANN as nothing more than a delay element and the degree of activation as nothing more than arrival time of a digital signal.Building on these fundamental operations, we can implement complete ANNs through several innovations: spatial and temporal reuse that facilitates classical dataflows, reducing memory system demands for ANN temporal operations; a new noise-tolerant temporal summation operation; novel hybrid digital/temporal memories; and the integration of temporal memory circuits for shepherding inter-layer activations.Using the MLPerf Tiny benchmark suite, we demonstrate how several architectural parameters can impact inference accuracy, that our proposed systolic array can provide 11× better energy consumption with a 4× improvement in latency compared to SNNs, and when equipped with temporal memories provides 3.5× improvements in energy compared to the most aggressive 8-bit digital systolic arrays.

Visual neuroprostheses (bionic eyes) aim to restore a rudimentary form of vision by translating camera input into patterns of electrical stimulation. To improve scene understanding under extreme resolution and bandwidth constraints, prior work has explored computer vision techniques such as semantic segmentation and depth estimation. However, presenting all task-relevant information simultaneously can overwhelm users in cluttered environments. We compare two complementary approaches to semantic preprocessing in immersive virtual reality: SemanticEdges, which highlights all relevant objects at once, and SemanticRaster, which staggers object categories over time to reduce visual clutter. Using a biologically grounded simulation of bionic vision, 18 sighted participants performed a wayfinding task in a dynamic urban environment across three conditions: edge-based baseline (Control), SemanticEdges, and SemanticRaster. Both semantic strategies improved performance and user experience relative to the baseline, with each offering distinct trade-offs: SemanticEdges increased the odds of success, while SemanticRaster boosted the likelihood of collision-free completions. These findings underscore the value of adaptive semantic preprocessing for bionic vision and, more broadly, may inform the design of low-bandwidth visual interfaces in XR that must balance information density, task relevance, and perceptual clarity.

Neuroprostheses show potential in restoring lost sensory function and enhancing human capabilities, but the sensations produced by current devices often seem unnatural or distorted. Exact placement of implants and differences in individual perception lead to significant variations in stimulus response, making personalized stimulus optimization a key challenge. Bayesian optimization could be used to optimize patient-specific stimulation parameters with limited noisy observations, but is not feasible for high-dimensional stimuli. Alternatively, deep learning models can optimize stimulus encoding strategies, but typically assume perfect knowledge of patient-specific variations. Here we propose a novel, practically feasible approach that overcomes both of these fundamental limitations. First, a deep encoder network is trained to produce optimal stimuli for any individual patient by inverting a forward model mapping electrical stimuli to visual percepts. Second, a preferential Bayesian optimization strategy utilizes this encoder to learn the optimal patient-specific parameters for a new patient, using a minimal number of pairwise comparisons between candidate stimuli. We demonstrate the viability of this approach on a novel, state-of-the-art visual prosthesis model. Our approach quickly learns a personalized stimulus encoder and leads to dramatic improvements in the quality of restored vision, outperforming existing encoding strategies. Further, this approach is robust to noisy patient feedback and misspecifications in the underlying forward model. Overall, our results suggest that combining the strengths of deep learning and Bayesian optimization could significantly improve the perceptual experience of patients fitted with visual prostheses and may prove a viable solution for a range of neuroprosthetic technologies

Convolutional neural networks (CNNs) have recently emerged as promising models of the ventral visual stream, despite their lack of biological specificity. While current state-of-the-art models of the primary visual cortex (V1) have surfaced from training with adversarial examples and extensively augmented data, these models are still unable to explain key neural properties observed in V1 that arise from biological circuitry. To address this gap, we systematically incorporated neuroscience-derived architectural components into CNNs to identify a set of mechanisms and architectures that more comprehensively explain V1 activity. Upon enhancing task-driven CNNs with architectural components that simulate center-surround antagonism, local receptive fields, tuned normalization, and cortical magnification, we uncover models with latent representations that yield state-of-the-art explanation of V1 neural activity and tuning properties. Moreover, analyses of the learned parameters of these components and stimuli that maximally activate neurons of the evaluated networks provide support for their role in explaining neural properties of V1. Our results highlight an important advancement in the field of NeuroAI, as we systematically establish a set of architectural components that contribute to unprecedented explanation of V1. The neuroscience insights that could be gleaned from increasingly accurate in-silico models of the brain have the potential to greatly advance the fields of both neuroscience and artificial intelligence.

Despite their immense success as a model of macaque visual cortex, deep convolutional neural networks (CNNs) have struggled to predict activity in visual cortex of the mouse, which is thought to be strongly dependent on the animal’s behavioral state. Furthermore, most computational models focus on predicting neural responses to static images presented under head fixation, which are dramatically different from the dynamic, continuous visual stimuli that arise during movement in the real world. Consequently, it is still unknown how natural visual input and different behavioral variables may integrate over time to generate responses in primary visual cortex (V1). To address this, we introduce a multimodal recurrent neural network that integrates gaze-contingent visual input with behavioral and temporal dynamics to explain V1 activity in freely moving mice. We show that the model achieves state-of-the-art predictions of V1 activity during free exploration and demonstrate the importance of each component in an extensive ablation study. Analyzing our model using maximally activating stimuli and saliency maps, we reveal new insights into cortical function, including the prevalence of mixed selectivity for behavioral variables in mouse V1. In summary, our model offers a comprehensive deep-learning framework for exploring the computational principles underlying V1 neurons in freely-moving animals engaged in natural behavior.

Sensory neuroprostheses are emerging as a promising technology to restore lost sensory function or augment human capabilities. However, sensations elicited by current devices often appear artificial and distorted. Although current models can predict the neural or perceptual response to an electrical stimulus, an optimal stimulation strategy solves the inverse problem: what is the required stimulus to produce a desired response? Here, we frame this as an end-to-end optimization problem, where a deep neural network stimulus encoder is trained to invert a known and fixed forward model that approximates the underlying biological system. As a proof of concept, we demonstrate the effectiveness of this Hybrid Neural Autoencoder (HNA) in visual neuroprostheses. We find that HNA produces high-fidelity patient-specific stimuli representing handwritten digits and segmented images of everyday objects, and significantly outperforms conventional encoding strategies across all simulated patients. Overall this is an important step towards the long-standing challenge of restoring high-quality vision to people living with incurable blindness and may prove a promising solution for a variety of neuroprosthetic technologies.

Visual neuroprostheses (bionic eyes) have the potential to treat degenerative eye diseases that often result in low vision or complete blindness. These devices rely on an external camera to capture the visual scene, which is then translated frame-by-frame into an electrical stimulation pattern that is sent to the implant in the eye. To highlight more meaningful information in the scene, recent studies have tested the effectiveness of deep-learning based computer vision techniques, such as depth estimation to highlight nearby obstacles (DepthOnly mode) and semantic edge detection to outline important objects in the scene (EdgesOnly mode). However, nobody has yet attempted to combine the two, either by presenting them together (EdgesAndDepth) or by giving the user the ability to flexibly switch between them (EdgesOrDepth). Here, we used a neurobiologically inspired model of simulated prosthetic vision (SPV) in an immersive virtual reality (VR) environment to test the relative importance of semantic edges and relative depth cues to support the ability to avoid obstacles and identify objects. We found that participants were significantly better at avoiding obstacles using depth-based cues as opposed to relying on edge information alone, and that roughly half the participants preferred the flexibility to switch between modes (EdgesOrDepth). This study highlights the relative importance of depth cues for SPV mobility and is an important first step towards a visual neuroprosthesis that uses computer vision to improve a user’s scene understanding.

Supported by recent computational studies, there is increasing evidence that a wide range of neuronal responses can be understood as an emergent property of nonnegative sparse coding (NSC), an efficient population coding scheme based on dimensionality reduction and sparsity constraints.We review evidence that NSC might be employed by sensory areas to efficiently encode external stimulus spaces, by some associative areas to conjunctively represent multiple behaviorally relevant variables, and possibly by the basal ganglia to coordinate movement.In addition, NSC might provide a useful theoretical framework under which to understand the often complex and nonintuitive response properties of neurons in other brain areas.Although NSC might not apply to all brain areas (for example, motor or executive function areas) the success of NSC-based models, especially in sensory areas, warrants further investigation for neural correlates in other regions. Author summaryBrains face the fundamental challenge of extracting relevant information from high-dimensional external stimuli in order to form the neural basis that can guide an organism's behavior and its interaction with the world.One potential approach to addressing this challenge is to reduce the number of variables required to represent a particular input space (i.e., dimensionality reduction).We review compelling evidence that a range of neuronal responses can be understood as an emergent property of nonnegative sparse coding (NSC)-a form of efficient population coding due to dimensionality reduction and sparsity constraints.

Large-scale spiking neural network (SNN) simulations are challenging to implement, due to the memory and computation required to iteratively process the large set of neural state dynamics and updates. To meet these challenges, we have developed CARLsim 4, a user-friendly SNN library written in C++ that can simulate large biologically detailed neural networks. Improving on the efficiency and scalability of earlier releases, the present release allows for the simulation using multiple GPUs and multiple CPU cores concurrently in a heterogeneous computing cluster. Benchmarking results demonstrate simulation of 8.6 million neurons and 0.48 billion synapses using 4 GPUs and up to 60x speedup for multi-GPU implementations over a single-threaded CPU implementation, making CARLsim 4 well-suited for large-scale SNN models in the presence of real-time constraints. Additionally, the present release adds new features, such as leaky-integrate-and-fire (LIF), 9-parameter Izhikevich, multi-compartment neuron models, and fourth order Runge Kutta integration.

Spiking neural network (SNN) models describe key aspects of neural function in a computationally efficient manner and have been used to construct large-scale brain models. Large-scale SNNs are challenging to implement, as they demand high-bandwidth communication, a large amount of memory, and are computationally intensive. Additionally, tuning parameters of these models becomes more difficult and time-consuming with the addition of biologically accurate descriptions. To meet these challenges, we have developed CARLsim 3, a user-friendly, GPU-accelerated SNN library written in C/C++ that is capable of simulating biologically detailed neural models. The present release of CARLsim provides a number of improvements over our prior SNN library to allow the user to easily analyze simulation data, explore synaptic plasticity rules, and automate parameter tuning. In the present paper, we provide examples and performance benchmarks highlighting the library's features.

Neuromorphic engineering takes inspiration from biology to design brain-like systems that are extremely low-power, fault-tolerant, and capable of adaptation to complex environments. The design of these artificial nervous systems involves both the development of neuromorphic hardware devices and the development neuromorphic simulation tools. In this paper, we describe a simulation environment that can be used to design, construct, and run spiking neural networks (SNNs) quickly and efficiently using graphics processing units (GPUs). We then explain how the design of the simulation environment utilizes the parallel processing power of GPUs to simulate large-scale SNNs and describe recent modeling experiments performed using the simulator. Finally, we present an automated parameter tuning framework that utilizes the simulation environment and evolutionary algorithms to tune SNNs. We believe the simulation environment and associated parameter tuning framework presented here can accelerate the development of neuromorphic software and hardware applications by making the design, construction, and tuning of SNNs an easier task.

Road and lane detection play an important role in autonomous driving and commercial driver-assistance systems. Vision-based road detection is an essential step towards autonomous driving, yet a challenging task due to illumination and complexity of the visual scenery. Urban scenes may present additional challenges such as intersections, multi-lane scenarios, or clutter due to heavy traffic. This paper presents an integrative approach to ego-lane detection that aims to be as simple as possible to enable real-time computation while being able to adapt to a variety of urban and rural traffic scenarios. The approach at hand combines and extends a road segmentation method in an illumination-invariant color image, lane markings detection using a ridge operator, and road geometry estimation using RANdom SAmple Consensus (RANSAC). Employing the segmented road region as a prior for lane markings extraction significantly improves the execution time and success rate of the RANSAC algorithm, and makes the detection of weakly pronounced ridge structures computationally tractable, thus enabling ego-lane detection even in the absence of lane markings. Segmentation performance is shown to increase when moving from a color-based to a histogram correlation-based model. The power and robustness of this algorithm has been demonstrated in a car simulation system as well as in the challenging KITTI data base of real-world urban traffic scenarios.