Neural inference at the frontier of energy, space, and time.

Computing, since its inception, has been processor-centric, with memory separated from compute. Inspired by the organic brain and optimized for inorganic silicon, NorthPole is a neural inference architecture that blurs this boundary by eliminating off-chip memory, intertwining compute with memory on-chip, and appearing externally as an active memory chip. NorthPole is a low-precision, massively parallel, densely interconnected, energy-efficient, and spatial computing architecture with a co-optimized, high-utilization programming model. On the ResNet50 benchmark image classification network, relative to a graphics processing unit (GPU) that uses a comparable 12-nanometer technology process, NorthPole achieves a 25 times higher energy metric of frames per second (FPS) per watt, a 5 times higher space metric of FPS per transistor, and a 22 times lower time metric of latency. Similar results are reported for the Yolo-v4 detection network. NorthPole outperforms all prevalent architectures, even those that use more-advanced technology processes.

Spontaneous emergence of fast attractor dynamics in a model of developing primary visual cortex.

Recent evidence suggests that neurons in primary sensory cortex arrange into competitive groups, representing stimuli by their joint activity rather than as independent feature analysers. A possible explanation for these results is that sensory cortex implements attractor dynamics, although this proposal remains controversial. Here we report that fast attractor dynamics emerge naturally in a computational model of a patch of primary visual cortex endowed with realistic plasticity (at both feedforward and lateral synapses) and mutual inhibition. When exposed to natural images (but not random pixels), the model spontaneously arranges into competitive groups of reciprocally connected, similarly tuned neurons, while developing realistic, orientation-selective receptive fields. Importantly, the same groups are observed in both stimulus-evoked and spontaneous (stimulus-absent) activity. The resulting network is inhibition-stabilized and exhibits fast, non-persistent attractor dynamics. Our results suggest that realistic plasticity, mutual inhibition and natural stimuli are jointly necessary and sufficient to generate attractor dynamics in primary sensory cortex.

Convolutional networks for fast, energy-efficient neuromorphic computing.

Deep networks are now able to achieve human-level performance on a broad spectrum of recognition tasks. Independently, neuromorphic computing has now demonstrated unprecedented energy-efficiency through a new chip architecture based on spiking neurons, low precision synapses, and a scalable communication network. Here, we demonstrate that neuromorphic computing, despite its novel architectural primitives, can implement deep convolution networks that (i) approach state-of-the-art classification accuracy across eight standard datasets encompassing vision and speech, (ii) perform inference while preserving the hardware's underlying energy-efficiency and high throughput, running on the aforementioned datasets at between 1,200 and 2,600 frames/s and using between 25 and 275 mW (effectively >6,000 frames/s per Watt), and (iii) can be specified and trained using backpropagation with the same ease-of-use as contemporary deep learning. This approach allows the algorithmic power of deep learning to be merged with the efficiency of neuromorphic processors, bringing the promise of embedded, intelligent, brain-inspired computing one step closer.

Imagery May Arise from Associations Formed through Sensory Experience: A Network of Spiking Neurons Controlling a Robot Learns Visual Sequences in Order to Perform a Mental Rotation Task.

Mental imagery occurs "when a representation of the type created during the initial phases of perception is present but the stimulus is not actually being perceived." How does the capability to perform mental imagery arise? Extending the idea that imagery arises from learned associations, we propose that mental rotation, a specific form of imagery, could arise through the mechanism of sequence learning-that is, by learning to regenerate the sequence of mental images perceived while passively observing a rotating object. To demonstrate the feasibility of this proposal, we constructed a simulated nervous system and embedded it within a behaving humanoid robot. By observing a rotating object, the system learns the sequence of neural activity patterns generated by the visual system in response to the object. After learning, it can internally regenerate a similar sequence of neural activations upon briefly viewing the static object. This system learns to perform a mental rotation task in which the subject must determine whether two objects are identical despite differences in orientation. As with human subjects, the time taken to respond is proportional to the angular difference between the two stimuli. Moreover, as reported in humans, the system fills in intermediate angles during the task, and this putative mental rotation activates the same pathways that are activated when the system views physical rotation. This work supports the proposal that mental rotation arises through sequence learning and the idea that mental imagery aids perception through learned associations, and suggests testable predictions for biological experiments.

Temporal sequence learning in winner-take-all networks of spiking neurons demonstrated in a brain-based device.

Animal behavior often involves a temporally ordered sequence of actions learned from experience. Here we describe simulations of interconnected networks of spiking neurons that learn to generate patterns of activity in correct temporal order. The simulation consists of large-scale networks of thousands of excitatory and inhibitory neurons that exhibit short-term synaptic plasticity and spike-timing dependent synaptic plasticity. The neural architecture within each area is arranged to evoke winner-take-all (WTA) patterns of neural activity that persist for tens of milliseconds. In order to generate and switch between consecutive firing patterns in correct temporal order, a reentrant exchange of signals between these areas was necessary. To demonstrate the capacity of this arrangement, we used the simulation to train a brain-based device responding to visual input by autonomously generating temporal sequences of motor actions.

Versatile networks of simulated spiking neurons displaying winner-take-all behavior.

We describe simulations of large-scale networks of excitatory and inhibitory spiking neurons that can generate dynamically stable winner-take-all (WTA) behavior. The network connectivity is a variant of center-surround architecture that we call center-annular-surround (CAS). In this architecture each neuron is excited by nearby neighbors and inhibited by more distant neighbors in an annular-surround region. The neural units of these networks simulate conductance-based spiking neurons that interact via mechanisms susceptible to both short-term synaptic plasticity and STDP. We show that such CAS networks display robust WTA behavior unlike the center-surround networks and other control architectures that we have studied. We find that a large-scale network of spiking neurons with separate populations of excitatory and inhibitory neurons can give rise to smooth maps of sensory input. In addition, we show that a humanoid brain-based-device (BBD) under the control of a spiking WTA neural network can learn to reach to target positions in its visual field, thus demonstrating the acquisition of sensorimotor coordination.

Embodied models of delayed neural responses: spatiotemporal categorization and predictive motor control in brain based devices.

In order to respond appropriately to environmental stimuli, organisms must integrate over time spatiotemporal signals that reflect object motion and self-movement. One possible mechanism to achieve this spatiotemporal transformation is to delay or lag neural responses. This paper reviews our recent modeling work testing the sufficiency of delayed responses in the nervous system in two different behavioral tasks: (1) Categorizing spatiotemporal tactile cues with thalamic "lag" cells and downstream coincidence detectors, and (2) Predictive motor control was achieved by the cerebellum through a delayed eligibility trace rule at cerebellar synapses. Since the timing of these neural signals must closely match real-world dynamics, we tested these ideas using the brain based device (BBD) approach in which a simulated nervous system is embodied in a robotic device. In both tasks, biologically inspired neural simulations with delayed neural responses were critical for successful behavior by the device.

A cerebellar model for predictive motor control tested in a brain-based device.

The cerebellum is known to be critical for accurate adaptive control and motor learning. We propose here a mechanism by which the cerebellum may replace reflex control with predictive control. This mechanism is embedded in a learning rule (the delayed eligibility trace rule) in which synapses onto a Purkinje cell or onto a cell in the deep cerebellar nuclei become eligible for plasticity only after a fixed delay from the onset of suprathreshold presynaptic activity. To investigate the proposal that the cerebellum is a general-purpose predictive controller guided by a delayed eligibility trace rule, a computer model based on the anatomy and dynamics of the cerebellum was constructed. It contained components simulating cerebellar cortex and deep cerebellar nuclei, and it received input from a middle temporal visual area and the inferior olive. The model was incorporated in a real-world brain-based device (BBD) built on a Segway robotic platform that learned to traverse curved paths. The BBD learned which visual motion cues predicted impending collisions and used this experience to avoid path boundaries. During learning, the BBD adapted its velocity and turning rate to successfully traverse various curved paths. By examining neuronal activity and synaptic changes during this behavior, we found that the cerebellar circuit selectively responded to motion cues in specific receptive fields of simulated middle temporal visual areas. The system described here prompts several hypotheses about the relationship between perception and motor control and may be useful in the development of general-purpose motor learning systems for machines.

Visual binding through reentrant connectivity and dynamic synchronization in a brain-based device.

Effective visual object recognition requires mechanisms to bind object features (e.g. color, shape and motion) while distinguishing distinct objects. Synchronously active neuronal circuits among reentrantly connected cortical areas may provide a basis for visual binding. To assess the potential of this mechanism, we have constructed a mobile brain-based device, Darwin VIII, which is guided by simulated analogues of cortical and sub-cortical areas required for visual processing, decision-making, reward and motor responses. These simulated areas are reentrantly connected and each area contains neuronal units representing both the mean activity level and the relative timing of the activity of groups of neurons. Darwin VIII learns to discriminate among multiple objects with shared visual features and associates 'target' objects with innately preferred auditory cues. We observed the co-activation of globally distributed neuronal circuits that corresponded to distinct objects in Darwin VIII's visual field. These circuits, which are constrained by a reentrant neuroanatomy and modulated by behavior and synaptic plasticity, are necessary for successful discrimination. By situating Darwin VIII in a rich real-world environment involving continual changes in the size and location of visual stimuli due to self-generated movement, and by recording its behavioral and neuronal responses in detail, we were able to show that reentrant connectivity and dynamic synchronization provide an effective mechanism for binding the features of visual objects.