Catalyzing next-generation Artificial Intelligence through NeuroAI.

Neuroscience has long been an essential driver of progress in artificial intelligence (AI). We propose that to accelerate progress in AI, we must invest in fundamental research in NeuroAI. A core component of this is the embodied Turing test, which challenges AI animal models to interact with the sensorimotor world at skill levels akin to their living counterparts. The embodied Turing test shifts the focus from those capabilities like game playing and language that are especially well-developed or uniquely human to those capabilities - inherited from over 500 million years of evolution - that are shared with all animals. Building models that can pass the embodied Turing test will provide a roadmap for the next generation of AI.

Dendrocentric learning for synthetic intelligence.

Artificial intelligence now advances by performing twice as many floating-point multiplications every two months, but the semiconductor industry tiles twice as many multipliers on a chip every two years. Moreover, the returns from tiling these multipliers ever more densely now diminish because signals must travel relatively farther and farther. Although travel can be shortened by stacking tiled multipliers in a three-dimensional chip, such a solution acutely reduces the available surface area for dissipating heat. Here I propose to transcend this three-dimensional thermal constraint by moving away from learning with synapses to learning with dendrites. Synaptic inputs are not weighted precisely but rather ordered meticulously along a short stretch of dendrite, termed dendrocentric learning. With the help of a computational model of a dendrite and a conceptual model of a ferroelectric device that emulates it, I illustrate how dendrocentric learning artificial intelligence-or synthetic intelligence for short-could run not with megawatts in the cloud but rather with watts on a smartphone.

Optimal noise level for coding with tightly balanced networks of spiking neurons in the presence of transmission delays.

Neural circuits consist of many noisy, slow components, with individual neurons subject to ion channel noise, axonal propagation delays, and unreliable and slow synaptic transmission. This raises a fundamental question: how can reliable computation emerge from such unreliable components? A classic strategy is to simply average over a population of N weakly-coupled neurons to achieve errors that scale as [Formula: see text]. But more interestingly, recent work has introduced networks of leaky integrate-and-fire (LIF) neurons that achieve coding errors that scale superclassically as 1/N by combining the principles of predictive coding and fast and tight inhibitory-excitatory balance. However, spike transmission delays preclude such fast inhibition, and computational studies have observed that such delays can cause pathological synchronization that in turn destroys superclassical coding performance. Intriguingly, it has also been observed in simulations that noise can actually improve coding performance, and that there exists some optimal level of noise that minimizes coding error. However, we lack a quantitative theory that describes this fascinating interplay between delays, noise and neural coding performance in spiking networks. In this work, we elucidate the mechanisms underpinning this beneficial role of noise by deriving analytical expressions for coding error as a function of spike propagation delay and noise levels in predictive coding tight-balance networks of LIF neurons. Furthermore, we compute the minimal coding error and the associated optimal noise level, finding that they grow as power-laws with the delay. Our analysis reveals quantitatively how optimal levels of noise can rescue neural coding performance in spiking neural networks with delays by preventing the build up of pathological synchrony without overwhelming the overall spiking dynamics. This analysis can serve as a foundation for the further study of precise computation in the presence of noise and delays in efficient spiking neural circuits.

Cortical state dynamics and selective attention define the spatial pattern of correlated variability in neocortex.

Correlated activity fluctuations in the neocortex influence sensory responses and behavior. Neural correlations reflect anatomical connectivity but also change dynamically with cognitive states such as attention. Yet, the network mechanisms defining the population structure of correlations remain unknown. We measured correlations within columns in the visual cortex. We show that the magnitude of correlations, their attentional modulation, and dependence on lateral distance are explained by columnar On-Off dynamics, which are synchronous activity fluctuations reflecting cortical state. We developed a network model in which the On-Off dynamics propagate across nearby columns generating spatial correlations with the extent controlled by attentional inputs. This mechanism, unlike previous proposals, predicts spatially non-uniform changes in correlations during attention. We confirm this prediction in our columnar recordings by showing that in superficial layers the largest changes in correlations occur at intermediate lateral distances. Our results reveal how spatially structured patterns of correlated variability emerge through interactions of cortical state dynamics, anatomical connectivity, and attention.

Selective modulation of cortical state during spatial attention.

Neocortical activity is permeated with endogenously generated fluctuations, but how these dynamics affect goal-directed behavior remains a mystery. We found that ensemble neural activity in primate visual cortex spontaneously fluctuated between phases of vigorous (On) and faint (Off) spiking synchronously across cortical layers. These On-Off dynamics, reflecting global changes in cortical state, were also modulated at a local scale during selective attention. Moreover, the momentary phase of local ensemble activity predicted behavioral performance. Our results show that cortical state is controlled locally within a cortical map according to cognitive demands and reveal the impact of these local changes in cortical state on goal-directed behavior.

Potassium conductance dynamics confer robust spike-time precision in a neuromorphic model of the auditory brain stem.

A fundamental question in neuroscience is how neurons perform precise operations despite inherent variability. This question also applies to neuromorphic engineering, where low-power microchips emulate the brain using large populations of diverse silicon neurons. Biological neurons in the auditory pathway display precise spike timing, critical for sound localization and interpretation of complex waveforms such as speech, even though they are a heterogeneous population. Silicon neurons are also heterogeneous, due to a key design constraint in neuromorphic engineering: smaller transistors offer lower power consumption and more neurons per unit area of silicon, but also more variability between transistors and thus between silicon neurons. Utilizing this variability in a neuromorphic model of the auditory brain stem with 1,080 silicon neurons, we found that a low-voltage-activated potassium conductance (g(KL)) enables precise spike timing via two mechanisms: statically reducing the resting membrane time constant and dynamically suppressing late synaptic inputs. The relative contribution of these two mechanisms is unknown because blocking g(KL) in vitro eliminates dynamic adaptation but also lengthens the membrane time constant. We replaced g(KL) with a static leak in silico to recover the short membrane time constant and found that silicon neurons could mimic the spike-time precision of their biological counterparts, but only over a narrow range of stimulus intensities and biophysical parameters. The dynamics of g(KL) were required for precise spike timing robust to stimulus variation across a heterogeneous population of silicon neurons, thus explaining how neural and neuromorphic systems may perform precise operations despite inherent variability.

Design and validation of a real-time spiking-neural-network decoder for brain-machine interfaces.

OBJECTIVE: Cortically-controlled motor prostheses aim to restore functions lost to neurological disease and injury. Several proof of concept demonstrations have shown encouraging results, but barriers to clinical translation still remain. In particular, intracortical prostheses must satisfy stringent power dissipation constraints so as not to damage cortex. APPROACH: One possible solution is to use ultra-low power neuromorphic chips to decode neural signals for these intracortical implants. The first step is to explore in simulation the feasibility of translating decoding algorithms for brain-machine interface (BMI) applications into spiking neural networks (SNNs). MAIN RESULTS: Here we demonstrate the validity of the approach by implementing an existing Kalman-filter-based decoder in a simulated SNN using the Neural Engineering Framework (NEF), a general method for mapping control algorithms onto SNNs. To measure this system's robustness and generalization, we tested it online in closed-loop BMI experiments with two rhesus monkeys. Across both monkeys, a Kalman filter implemented using a 2000-neuron SNN has comparable performance to that of a Kalman filter implemented using standard floating point techniques. SIGNIFICANCE: These results demonstrate the tractability of SNN implementations of statistical signal processing algorithms on different monkeys and for several tasks, suggesting that a SNN decoder, implemented on a neuromorphic chip, may be a feasible computational platform for low-power fully-implanted prostheses. The validation of this closed-loop decoder system and the demonstration of its robustness and generalization hold promise for SNN implementations on an ultra-low power neuromorphic chip using the NEF.

Inferior olive mirrors joint dynamics to implement an inverse controller.

To produce smooth and coordinated motion, our nervous systems need to generate precisely timed muscle activation patterns that, due to axonal conduction delay, must be generated in a predictive and feedforward manner. Kawato proposed that the cerebellum accomplishes this by acting as an inverse controller that modulates descending motor commands to predictively drive the spinal cord such that the musculoskeletal dynamics are canceled out. This and other cerebellar theories do not, however, account for the rich biophysical properties expressed by the olivocerebellar complex's various cell types, making these theories difficult to verify experimentally. Here we propose that a multizonal microcomplex's (MZMC) inferior olivary neurons use their subthreshold oscillations to mirror a musculoskeletal joint's underdamped dynamics, thereby achieving inverse control. We used control theory to map a joint's inverse model onto an MZMC's biophysics, and we used biophysical modeling to confirm that inferior olivary neurons can express the dynamics required to mirror biomechanical joints. We then combined both techniques to predict how experimentally injecting current into the inferior olive would affect overall motor output performance. We found that this experimental manipulation unmasked a joint's natural dynamics, as observed by motor output ringing at the joint's natural frequency, with amplitude proportional to the amount of current. These results support the proposal that the cerebellum-in particular an MZMC-is an inverse controller; the results also provide a biophysical implementation for this controller and allow one to make an experimentally testable prediction.

A superposable silicon synapse with programmable reversal potential.

We present a novel log-domain silicon synapse designed for subthreshold analog operation that emulates common synaptic interactions found in biology. Our circuit models the dynamic gating of ion-channel conductances by emulating the processes of neurotransmitter release-reuptake and receptor binding-unbinding in a superposable fashion: Only a single circuit is required to model the entire population of synapses (of a given type) that a biological neuron receives. Unlike previous designs, which are strictly excitatory or inhibitory, our silicon synapse implements-for the first time in the log-domain-a programmable reversal potential (i.e., driving force). To demonstrate our design's scalability, we fabricated in 180nm CMOS an array of 64K silicon neurons, each with four independent superposable synapse circuits occupying 11.0×21.5 µm(2) apiece. After verifying that these synapses have the predicted effect on the neurons' spike rate, we explored a recurrent network where the synapses' reversal potentials are set near the neurons' threshold, acting as shunts. These shunting synapses synchronized neuronal spiking more robustly than nonshunting synapses, confirming that reversal potentials can have important network-level implications.

Neuromorphic silicon neuron circuits.

Hardware implementations of spiking neurons can be extremely useful for a large variety of applications, ranging from high-speed modeling of large-scale neural systems to real-time behaving systems, to bidirectional brain-machine interfaces. The specific circuit solutions used to implement silicon neurons depend on the application requirements. In this paper we describe the most common building blocks and techniques used to implement these circuits, and present an overview of a wide range of neuromorphic silicon neurons, which implement different computational models, ranging from biophysically realistic and conductance-based Hodgkin-Huxley models to bi-dimensional generalized adaptive integrate and fire models. We compare the different design methodologies used for each silicon neuron design described, and demonstrate their features with experimental results, measured from a wide range of fabricated VLSI chips.

Deep cerebellar neurons mirror the spinal cord's gain to implement an inverse controller.

Smooth and coordinated motion requires precisely timed muscle activation patterns, which due to biophysical limitations, must be predictive and executed in a feed-forward manner. In a previous study, we tested Kawato's original proposition, that the cerebellum implements an inverse controller, by mapping a multizonal microcomplex's (MZMC) biophysics to a joint's inverse transfer function and showing that inferior olivary neuron may use their intrinsic oscillations to mirror a joint's oscillatory dynamics. Here, to continue to validate our mapping, we propose that climbing fiber input into the deep cerebellar nucleus (DCN) triggers rebounds, primed by Purkinje cell inhibition, implementing gain on IO's signal to mirror the spinal cord reflex's gain thereby achieving inverse control. We used biophysical modeling to show that Purkinje cell inhibition and climbing fiber excitation interact in a multiplicative fashion to set DCN's rebound strength; where the former primes the cell for rebound by deinactivating its T-type Ca2(+) channels and the latter triggers the channels by rapidly depolarizing the cell. We combined this result with our control theory mapping to predict how experimentally injecting current into DCN will affect overall motor output performance, and found that injecting current will proportionally scale the output and unmask the joint's natural response as observed by motor output ringing at the joint's natural frequency. Experimental verification of this prediction will lend support to a MZMC as a joint's inverse controller and the role we assigned underlying biophysical principles that enable it.

Silicon-Neuron Design: A Dynamical Systems Approach.

We present an approach to design spiking silicon neurons based on dynamical systems theory. Dynamical systems theory aids in choosing the appropriate level of abstraction, prescribing a neuron model with the desired dynamics while maintaining simplicity. Further, we provide a procedure to transform the prescribed equations into subthreshold current-mode circuits. We present a circuit design example, a positive-feedback integrate-and-fire neuron, fabricated in 0.25 µm CMOS. We analyze and characterize the circuit, and demonstrate that it can be configured to exhibit desired behaviors, including spike-frequency adaptation and two forms of bursting.

Space coding by gamma oscillations in the barn owl optic tectum.

Gamma-band (25-140 Hz) oscillations of the local field potential (LFP) are evoked by sensory stimuli in the mammalian forebrain and may be strongly modulated in amplitude when animals attend to these stimuli. The optic tectum (OT) is a midbrain structure known to contribute to multimodal sensory processing, gaze control, and attention. We found that presentation of spatially localized stimuli, either visual or auditory, evoked robust gamma oscillations with distinctive properties in the superficial (visual) layers and in the deep (multimodal) layers of the owl's OT. Across layers, gamma power was tuned sharply for stimulus location and represented space topographically. In the superficial layers, induced LFP power peaked strongly in the low-gamma band (25-90 Hz) and increased gradually with visual contrast across a wide range of contrasts. Spikes recorded in these layers included presumptive axonal (input) spikes that encoded stimulus properties nearly identically with gamma oscillations and were tightly phase locked with the oscillations, suggesting that they contribute to the LFP oscillations. In the deep layers, induced LFP power was distributed across the low and high (90-140 Hz) gamma-bands and tended to reach its maximum value at relatively low visual contrasts. In these layers, gamma power was more sharply tuned for stimulus location, on average, than were somatic spike rates, and somatic spikes synchronized with gamma oscillations. Such gamma synchronized discharges of deep-layer neurons could provide a high-resolution temporal code for signaling the location of salient sensory stimuli.

A Brain-Machine Interface Operating with a Real-Time Spiking Neural Network Control Algorithm.

Motor prostheses aim to restore function to disabled patients. Despite compelling proof of concept systems, barriers to clinical translation remain. One challenge is to develop a low-power, fully-implantable system that dissipates only minimal power so as not to damage tissue. To this end, we implemented a Kalman-filter based decoder via a spiking neural network (SNN) and tested it in brain-machine interface (BMI) experiments with a rhesus monkey. The Kalman filter was trained to predict the arm's velocity and mapped on to the SNN using the Neural Engineering Framework (NEF). A 2,000-neuron embedded Matlab SNN implementation runs in real-time and its closed-loop performance is quite comparable to that of the standard Kalman filter. The success of this closed-loop decoder holds promise for hardware SNN implementations of statistical signal processing algorithms on neuromorphic chips, which may offer power savings necessary to overcome a major obstacle to the successful clinical translation of neural motor prostheses.

Spiking Neural Network Decoder for Brain-Machine Interfaces.

We used a spiking neural network (SNN) to decode neural data recorded from a 96-electrode array in premotor/motor cortex while a rhesus monkey performed a point-to-point reaching arm movement task. We mapped a Kalman-filter neural prosthetic decode algorithm developed to predict the arm's velocity on to the SNN using the Neural Engineering Framework and simulated it using Nengo, a freely available software package. A 20,000-neuron network matched the standard decoder's prediction to within 0.03% (normalized by maximum arm velocity). A 1,600-neuron version of this network was within 0.27%, and run in real-time on a 3GHz PC. These results demonstrate that a SNN can implement a statistical signal processing algorithm widely used as the decoder in high-performance neural prostheses (Kalman filter), and achieve similar results with just a few thousand neurons. Hardware SNN implementations-neuromorphic chips-may offer power savings, essential for realizing fully-implantable cortically controlled prostheses.

Nonlinear influence of T-channels in an in silico relay neuron.

Thalamic relay cells express distinctive response modes based on the state of a low-threshold calcium channel (T-channel). When the channel is fully active (burst mode), the cell responds to inputs with a high-frequency burst of spikes; with the channel inactive ( tonic mode), the cell responds at a rate proportional to the input. Due to the T-channel's dynamics, we expect the cell's response to become more nonlinear as the channel becomes more active. To test this hypothesis, we study the response of an in silico relay cell to Poisson spike trains. We first validate our model cell by comparing its responses with in vitro responses. To characterize the model cell's nonlinearity, we calculate Poisson kernels, an approach akin to white noise analysis but using the randomness of Poisson input spikes instead of Gaussian white noise. We find that a relay cell with active T-channels requires at least a third-order system to achieve a characterization as good as a second-order system for a relay cell without T-channels.

Synchrony in silicon: the gamma rhythm.

In this paper, we present a network of silicon interneurons that synchronize in the gamma frequency range (20-80 Hz). The gamma rhythm strongly influences neuronal spike timing within many brain regions, potentially playing a crucial role in computation. Yet it has largely been ignored in neuromorphic systems, which use mixed analog and digital circuits to model neurobiology in silicon. Our neurons synchronize by using shunting inhibition (conductance based) with a synaptic rise time. Synaptic rise time promotes synchrony by delaying the effect of inhibition, providing an opportune period for interneurons to spike together. Shunting inhibition, through its voltage dependence, inhibits interneurons that spike out of phase more strongly (delaying the spike further), pushing them into phase (in the next cycle). We characterize the interneuron, which consists of soma (cell body) and synapse circuits, fabricated in a 0.25-microm complementary metal-oxide-semiconductor (CMOS). Further, we show that synchronized interneurons (population of 256) spike with a period that is proportional to the synaptic rise time. We use these interneurons to entrain model excitatory principal neurons and to implement a form of object binding.

Neurotech for neuroscience: unifying concepts, organizing principles, and emerging tools.

The ability to tackle analysis of the brain at multiple levels simultaneously is emerging from rapid methodological developments. The classical research strategies of "measure," "model," and "make" are being applied to the exploration of nervous system function. These include novel conceptual and theoretical approaches, creative use of mathematical modeling, and attempts to build brain-like devices and systems, as well as other developments including instrumentation and statistical modeling (not covered here). Increasingly, these efforts require teams of scientists from a variety of traditional scientific disciplines to work together. The potential of such efforts for understanding directed motor movement, emergence of cognitive function from neuronal activity, and development of neuromimetic computers are described by a team that includes individuals experienced in behavior and neuroscience, mathematics, and engineering. Funding agencies, including the National Science Foundation, explore the potential of these changing frontiers of research for developing research policies and long-term planning.

Functional circuitry for peripheral suppression in Mammalian Y-type retinal ganglion cells.

A retinal ganglion cell receptive field is made up of an excitatory center and an inhibitory surround. The surround has two components: one driven by horizontal cells at the first synaptic layer and one driven by amacrine cells at the second synaptic layer. Here we characterized how amacrine cells inhibit the center response of on- and off-center Y-type ganglion cells in the in vitro guinea pig retina. A high spatial frequency grating (4-5 cyc/mm), beyond the spatial resolution of horizontal cells, drifted in the ganglion cell receptive field periphery to stimulate amacrine cells. The peripheral grating suppressed the ganglion cell spiking response to a central spot. Suppression of spiking was strongest and observed most consistently in off cells. In intracellular recordings, the grating suppressed the subthreshold membrane potential in two ways: a reduced slope (gain) of the stimulus-response curve by approximately 20-30% and, in off cells, a tonic approximately 1-mV hyperpolarization. In voltage clamp, the grating increased an inhibitory conductance in all cells and simultaneously decreased an excitatory conductance in off cells. To determine whether center response inhibition was presynaptic or postsynaptic (shunting), we measured center response gain under voltage-clamp and current-clamp conditions. Under both conditions, the peripheral grating reduced center response gain similarly. This result suggests that reduced gain in the ganglion cell subthreshold center response reflects inhibition of presynaptic bipolar terminals. Thus amacrine cells suppressed ganglion cell center response gain primarily by inhibiting bipolar cell glutamate release.