Learning to represent signals spike by spike

Spiking networks that efficiently process dynamic sensory features explain receptor information mixing in somatosensory cortex

How do biological neural systems efficiently encode, transform and propagate information between the sensory periphery and the sensory cortex about sensory features evolving at different time scales? Are these computations efficient in normative information processing terms? While previous work has suggested that biologically plausible models of of such neural information processing may be implemented efficiently within a single processing layer, how such computations extend across several processing layers is less clear. Here, we model propagation of multiple time-varying sensory features across a sensory pathway, by extending the theory of efficient coding with spikes to efficient encoding, transformation and transmission of sensory signals. These computations are optimally realized by a multilayer spiking network with feedforward networks of spiking neurons (receptor layer) and recurrent excitatory-inhibitory networks of generalized leaky integrate-and-fire neurons (recurrent layers). Our model efficiently realizes a broad class of feature transformations, including positive and negative interaction across features, through specific and biologically plausible structures of feedforward connectivity. We find that mixing of sensory features in the activity of single neurons is beneficial because it lowers the metabolic cost at the network level. We apply the model to the somatosensory pathway by constraining it with parameters measured empirically and include in its last node, analogous to the primary somatosensory cortex (S1), two types of inhibitory neurons: parvalbumin-positive neurons realizing lateral inhibition, and somatostatin-positive neurons realizing winner-take-all inhibition. By implementing a negative interaction across stimulus features, this model captures several intriguing empirical observations from the somatosensory system of the mouse, including a decrease of sustained responses from subcortical networks to S1, a non-linear effect of the knock-out of receptor neuron types on the activity in S1, and amplification of weak signals from sensory neurons across the pathway.

Desiderata for Normative Models of Synaptic Plasticity

Normative models of synaptic plasticity use computational rationales to arrive at predictions of behavioral and network-level adaptive phenomena. In recent years, there has been an explosion of theoretical work in this realm, but experimental confirmation remains limited. In this review, we organize work on normative plasticity models in terms of a set of desiderata that, when satisfied, are designed to ensure that a given model demonstrates a clear link between plasticity and adaptive behavior, is consistent with known biological evidence about neural plasticity and yields specific testable predictions. As a prototype, we include a detailed analysis of the REINFORCE algorithm. We also discuss how new models have begun to improve on the identified criteria and suggest avenues for further development. Overall, we provide a conceptual guide to help develop neural learning theories that are precise, powerful, and experimentally testable.

The impact of spike timing precision and spike emission reliability on decoding accuracy

Precisely timed and reliably emitted spikes are hypothesized to serve multiple functions, including improving the accuracy and reproducibility of encoding stimuli, memories, or behaviours across trials. When these spikes occur as a repeating sequence, they can be used to encode and decode a potential time series. Here, we show both analytically and in simulations that the error incurred in approximating a time series with precisely timed and reliably emitted spikes decreases linearly with the number of neurons or spikes used in the decoding. This was verified numerically with synthetically generated patterns of spikes. Further, we found that if spikes were imprecise in their timing, or unreliable in their emission, the error incurred in decoding with these spikes would be sub-linear. However, if the spike precision or spike reliability increased with network size, the error incurred in decoding a time-series with sequences of spikes would maintain a linear decrease with network size. The spike precision had to increase linearly with network size, while the probability of spike failure had to decrease with the square-root of the network size. Finally, we identified a candidate circuit to test this scaling relationship: the repeating sequences of spikes with sub-millisecond precision in area HVC (proper name) of the zebra finch. This scaling relationship can be tested using both neural data and song-spectrogram-based recordings while taking advantage of the natural fluctuation in HVC network size due to neurogenesis.

Structure, dynamics, coding and optimal biophysical parameters of efficient excitatory-inhibitory spiking networks

The principle of efficient coding posits that sensory cortical networks are designed to encode maximal sensory information with minimal metabolic cost. Despite the major influence of efficient coding in neuro-science, it has remained unclear whether fundamental empirical properties of neural network activity can be explained solely based on this normative principle. Here, we rigorously derive the structural, coding, biophysical and dynamical properties of excitatory-inhibitory recurrent networks of spiking neurons that emerge directly from imposing that the network minimizes an instantaneous loss function and a time-averaged performance measure enacting efficient coding. The optimal network has biologically-plausible biophysical features, including realistic integrate-and-fire spiking dynamics, spike-triggered adaptation, and a non-stimulus-specific excitatory external input regulating metabolic cost. The efficient network has excitatory-inhibitory recurrent connectivity between neurons with similar stimulus tuning implementing feature-specific competition, similar to that recently found in visual cortex. Networks with unstructured connectivity cannot reach comparable levels of coding efficiency. The optimal biophysical parameters include 4 to 1 ratio of excitatory vs inhibitory neurons and 3 to 1 ratio of mean inhibitory-to-inhibitory vs. excitatory-to-inhibitory connectivity that closely match those of cortical sensory networks. The efficient network has biologically-plausible spiking dynamics, with a tight instantaneous E-I balance that makes them capable to achieve efficient coding of external stimuli varying over multiple time scales. Together, these results explain how efficient coding may be implemented in cortical networks and suggests that key properties of biological neural networks may be accounted for by efficient coding.

Approximating Nonlinear Functions With Latent Boundaries in Low-Rank Excitatory-Inhibitory Spiking Networks

Deep feedforward and recurrent neural networks have become successful functional models of the brain, but they neglect obvious biological details such as spikes and Dale's law. Here we argue that these details are crucial in order to understand how real neural circuits operate. Towards this aim, we put forth a new framework for spike-based computation in low-rank excitatory-inhibitory spiking networks. By considering populations with rank-1 connectivity, we cast each neuron's spiking threshold as a boundary in a low-dimensional input-output space. We then show how the combined thresholds of a population of inhibitory neurons form a stable boundary in this space, and those of a population of excitatory neurons form an unstable boundary. Combining the two boundaries results in a rank-2 excitatory-inhibitory (EI) network with inhibition-stabilized dynamics at the intersection of the two boundaries. The computation of the resulting networks can be understood as the difference of two convex functions and is thereby capable of approximating arbitrary non-linear input-output mappings. We demonstrate several properties of these networks, including noise suppression and amplification, irregular activity and synaptic balance, as well as how they relate to rate network dynamics in the limit that the boundary becomes soft. Finally, while our work focuses on small networks (5-50 neurons), we discuss potential avenues for scaling up to much larger networks. Overall, our work proposes a new perspective on spiking networks that may serve as a starting point for a mechanistic understanding of biological spike-based computation.

Establishing brain states in neuroimaging data

The definition of a brain state remains elusive, with varying interpretations across different sub-fields of neuroscience-from the level of wakefulness in anaesthesia, to activity of individual neurons, voltage in EEG, and blood flow in fMRI. This lack of consensus presents a significant challenge to the development of accurate models of neural dynamics. However, at the foundation of dynamical systems theory lies a definition of what constitutes the 'state' of a system-i.e., a specification of the system's future. Here, we propose to adopt this definition to establish brain states in neuroimaging timeseries by applying Dynamic Causal Modelling (DCM) to low-dimensional embedding of resting and task condition fMRI data. We find that ~90% of subjects in resting conditions are better described by first-order models, whereas ~55% of subjects in task conditions are better described by second-order models. Our work calls into question the status quo of using first-order equations almost exclusively within computational neuroscience and provides a new way of establishing brain states, as well as their associated phase space representations, in neuroimaging datasets.

Desiderata for normative models of synaptic plasticity

Normative models of synaptic plasticity use a combination of mathematics and computational simulations to arrive at predictions of behavioral and network-level adaptive phenomena. In recent years, there has been an explosion of theoretical work on these models, but experimental confirmation is relatively limited. In this review, we organize work on normative plasticity models in terms of a set of desiderata which, when satisfied, are designed to guarantee that a model has a clear link between plasticity and adaptive behavior, consistency with known biological evidence about neural plasticity, and specific testable predictions. We then discuss how new models have begun to improve on these criteria and suggest avenues for further development. As prototypes, we provide detailed analyses of two specific models - REINFORCE and the Wake-Sleep algorithm. We provide a conceptual guide to help develop neural learning theories that are precise, powerful, and experimentally testable.

Omission responses in local field potentials in rat auditory cortex

BACKGROUND

Non-invasive recordings of gross neural activity in humans often show responses to omitted stimuli in steady trains of identical stimuli. This has been taken as evidence for the neural coding of prediction or prediction error. However, evidence for such omission responses from invasive recordings of cellular-scale responses in animal models is scarce. Here, we sought to characterise omission responses using extracellular recordings in the auditory cortex of anaesthetised rats. We profiled omission responses across local field potentials (LFP), analogue multiunit activity (AMUA), and single/multi-unit spiking activity, using stimuli that were fixed-rate trains of acoustic noise bursts where 5% of bursts were randomly omitted.

RESULTS

Significant omission responses were observed in LFP and AMUA signals, but not in spiking activity. These omission responses had a lower amplitude and longer latency than burst-evoked sensory responses, and omission response amplitude increased as a function of the number of preceding bursts.

CONCLUSIONS

Together, our findings show that omission responses are most robustly observed in LFP and AMUA signals (relative to spiking activity). This has implications for models of cortical processing that require many neurons to encode prediction errors in their spike output.

Biophysical parameters control signal transfer in spiking network

Introduction

Information transmission and representation in both natural and artificial networks is dependent on connectivity between units. Biological neurons, in addition, modulate synaptic dynamics and post-synaptic membrane properties, but how these relate to information transmission in a population of neurons is still poorly understood. A recent study investigated local learning rules and showed how a spiking neural network can learn to represent continuous signals. Our study builds on their model to explore how basic membrane properties and synaptic delays affect information transfer.

Methods

The system consisted of three input and output units and a hidden layer of 300 excitatory and 75 inhibitory leaky integrate-and-fire (LIF) or adaptive integrate-and-fire (AdEx) units. After optimizing the connectivity to accurately replicate the input patterns in the output units, we transformed the model to more biologically accurate units and included synaptic delay and concurrent action potential generation in distinct neurons. We examined three different parameter regimes which comprised either identical physiological values for both excitatory and inhibitory units (Comrade), more biologically accurate values (Bacon), or the Comrade regime whose output units were optimized for low reconstruction error (HiFi). We evaluated information transmission and classification accuracy of the network with four distinct metrics: coherence, Granger causality, transfer entropy, and reconstruction error.

Results

Biophysical parameters showed a major impact on information transfer metrics. The classification was surprisingly robust, surviving very low firing and information rates, whereas information transmission overall and particularly low reconstruction error were more dependent on higher firing rates in LIF units. In AdEx units, the firing rates were lower and less information was transferred, but interestingly the highest information transmission rates were no longer overlapping with the highest firing rates.

Discussion

Our findings can be reflected on the predictive coding theory of the cerebral cortex and may suggest information transfer qualities as a phenomenological quality of biological cells.

Computational methods to study information processing in neural circuits

The brain is an information processing machine and thus naturally lends itself to be studied using computational tools based on the principles of information theory. For this reason, computational methods based on or inspired by information theory have been a cornerstone of practical and conceptual progress in neuroscience. In this Review, we address how concepts and computational tools related to information theory are spurring the development of principled theories of information processing in neural circuits and the development of influential mathematical methods for the analyses of neural population recordings. We review how these computational approaches reveal mechanisms of essential functions performed by neural circuits. These functions include efficiently encoding sensory information and facilitating the transmission of information to downstream brain areas to inform and guide behavior. Finally, we discuss how further progress and insights can be achieved, in particular by studying how competing requirements of neural encoding and readout may be optimally traded off to optimize neural information processing.

Where is the error? Hierarchical predictive coding through dendritic error computation

Top-down feedback in cortex is critical for guiding sensory processing, which has prominently been formalized in the theory of hierarchical predictive coding (hPC). However, experimental evidence for error units, which are central to the theory, is inconclusive and it remains unclear how hPC can be implemented with spiking neurons. To address this, we connect hPC to existing work on efficient coding in balanced networks with lateral inhibition and predictive computation at apical dendrites. Together, this work points to an efficient implementation of hPC with spiking neurons, where prediction errors are computed not in separate units, but locally in dendritic compartments. We then discuss the correspondence of this model to experimentally observed connectivity patterns, plasticity, and dynamics in cortex.

Predictive coding is a consequence of energy efficiency in recurrent neural networks

Predictive coding is a promising framework for understanding brain function. It postulates that the brain continuously inhibits predictable sensory input, ensuring preferential processing of surprising elements. A central aspect of this view is its hierarchical connectivity, involving recurrent message passing between excitatory bottom-up signals and inhibitory top-down feedback. Here we use computational modeling to demonstrate that such architectural hardwiring is not necessary. Rather, predictive coding is shown to emerge as a consequence of energy efficiency. When training recurrent neural networks to minimize their energy consumption while operating in predictive environments, the networks self-organize into prediction and error units with appropriate inhibitory and excitatory interconnections and learn to inhibit predictable sensory input. Moving beyond the view of purely top-down-driven predictions, we demonstrate, via virtual lesioning experiments, that networks perform predictions on two timescales: fast lateral predictions among sensory units and slower prediction cycles that integrate evidence over time.

Volitional Generation of Reproducible, Efficient Temporal Patterns

One of the extraordinary characteristics of the biological brain is the low energy expense it requires to implement a variety of biological functions and intelligence as compared to the modern artificial intelligence (AI). Spike-based energy-efficient temporal codes have long been suggested as a contributor for the brain to run on low energy expense. Despite this code having been largely reported in the sensory cortex, whether this code can be implemented in other brain areas to serve broader functions and how it evolves throughout learning have remained unaddressed. In this study, we designed a novel brain-machine interface (BMI) paradigm. Two macaques could volitionally generate reproducible energy-efficient temporal patterns in the primary motor cortex (M1) by learning the BMI paradigm. Moreover, most neurons that were not directly assigned to control the BMI did not boost their excitability, and they demonstrated an overall energy-efficient manner in performing the task. Over the course of learning, we found that the firing rates and temporal precision of selected neurons co-evolved to generate the energy-efficient temporal patterns, suggesting that a cohesive rather than dissociable processing underlies the refinement of energy-efficient temporal patterns.

Error-based or target-based? A unified framework for learning in recurrent spiking networks

The field of recurrent neural networks is over-populated by a variety of proposed learning rules and protocols. The scope of this work is to define a generalized framework, to move a step forward towards the unification of this fragmented scenario. In the field of supervised learning, two opposite approaches stand out, error-based and target-based. This duality gave rise to a scientific debate on which learning framework is the most likely to be implemented in biological networks of neurons. Moreover, the existence of spikes raises the question of whether the coding of information is rate-based or spike-based. To face these questions, we proposed a learning model with two main parameters, the rank of the feedback learning matrix R and the tolerance to spike timing τ_⋆. We demonstrate that a low (high) rank R accounts for an error-based (target-based) learning rule, while high (low) tolerance to spike timing promotes rate-based (spike-based) coding. We show that in a store and recall task, high-ranks allow for lower MSE values, while low-ranks enable a faster convergence. Our framework naturally lends itself to Behavioral Cloning and allows for efficiently solving relevant closed-loop tasks, investigating what parameters (R,τ⋆) are optimal to solve a specific task. We found that a high R is essential for tasks that require retaining memory for a long time (Button and Food). On the other hand, this is not relevant for a motor task (the 2D Bipedal Walker). In this case, we find that precise spike-based coding enables optimal performances. Finally, we show that our theoretical formulation allows for defining protocols to estimate the rank of the feedback error in biological networks. We release a PyTorch implementation of our model supporting GPU parallelization.

Learning in biological or artificial networks means changing the laws governing the network dynamics in order to better behave in a specific situation. However, there exists no consensus on what rules regulate learning in biological systems. To face these questions, we propose a novel theoretical formulation for learning with two main parameters, the number of learning constraints (R) and the tolerance to spike timing (τ_⋆). We demonstrate that a low (high) rank R accounts for an error-based (target-based) learning rule, while high (low) tolerance to spike timing τ_⋆ promotes rate-based (spike-based) coding.

Our approach naturally lends itself to Imitation Learning (and Behavioral Cloning in particular) and we apply it to solve relevant closed-loop tasks such as the button-and-food task, and the 2D Bipedal Walker. The button-and-food is a navigation task that requires retaining a long-term memory, and benefits from a high R. On the other hand, the 2D Bipedal Walker is a motor task and benefits from a low τ_⋆.

Finally, we show that our theoretical formulation suggests protocols to deduce the structure of learning feedback in biological networks.

The geometry of robustness in spiking neural networks

Neural systems are remarkably robust against various perturbations, a phenomenon that still requires a clear explanation. Here, we graphically illustrate how neural networks can become robust. We study spiking networks that generate low-dimensional representations, and we show that the neurons’ subthreshold voltages are confined to a convex region in a lower-dimensional voltage subspace, which we call a 'bounding box'. Any changes in network parameters (such as number of neurons, dimensionality of inputs, firing thresholds, synaptic weights, or transmission delays) can all be understood as deformations of this bounding box. Using these insights, we show that functionality is preserved as long as perturbations do not destroy the integrity of the bounding box. We suggest that the principles underlying robustness in these networks — low-dimensional representations, heterogeneity of tuning, and precise negative feedback — may be key to understanding the robustness of neural systems at the circuit level.

Local dendritic balance enables learning of efficient representations in networks of spiking neurons

How can neural networks learn to efficiently represent complex and high-dimensional inputs via local plasticity mechanisms? Classical models of representation learning assume that feedforward weights are learned via pairwise Hebbian-like plasticity. Here, we show that pairwise Hebbian-like plasticity works only under unrealistic requirements on neural dynamics and input statistics. To overcome these limitations, we derive from first principles a learning scheme based on voltage-dependent synaptic plasticity rules. Here, recurrent connections learn to locally balance feedforward input in individual dendritic compartments and thereby can modulate synaptic plasticity to learn efficient representations. We demonstrate in simulations that this learning scheme works robustly even for complex high-dimensional inputs and with inhibitory transmission delays, where Hebbian-like plasticity fails. Our results draw a direct connection between dendritic excitatory-inhibitory balance and voltage-dependent synaptic plasticity as observed in vivo and suggest that both are crucial for representation learning.

Supervised training of spiking neural networks for robust deployment on mixed-signal neuromorphic processors

Mixed-signal analog/digital circuits emulate spiking neurons and synapses with extremely high energy efficiency, an approach known as "neuromorphic engineering". However, analog circuits are sensitive to process-induced variation among transistors in a chip ("device mismatch"). For neuromorphic implementation of Spiking Neural Networks (SNNs), mismatch causes parameter variation between identically-configured neurons and synapses. Each chip exhibits a different distribution of neural parameters, causing deployed networks to respond differently between chips. Current solutions to mitigate mismatch based on per-chip calibration or on-chip learning entail increased design complexity, area and cost, making deployment of neuromorphic devices expensive and difficult. Here we present a supervised learning approach that produces SNNs with high robustness to mismatch and other common sources of noise. Our method trains SNNs to perform temporal classification tasks by mimicking a pre-trained dynamical system, using a local learning rule from non-linear control theory. We demonstrate our method on two tasks requiring temporal memory, and measure the robustness of our approach to several forms of noise and mismatch. We show that our approach is more robust than common alternatives for training SNNs. Our method provides robust deployment of pre-trained networks on mixed-signal neuromorphic hardware, without requiring per-device training or calibration.

Optimal learning with excitatory and inhibitory synapses

Characterizing the relation between weight structure and input/output statistics is fundamental for understanding the computational capabilities of neural circuits. In this work, I study the problem of storing associations between analog signals in the presence of correlations, using methods from statistical mechanics. I characterize the typical learning performance in terms of the power spectrum of random input and output processes. I show that optimal synaptic weight configurations reach a capacity of 0.5 for any fraction of excitatory to inhibitory weights and have a peculiar synaptic distribution with a finite fraction of silent synapses. I further provide a link between typical learning performance and principal components analysis in single cases. These results may shed light on the synaptic profile of brain circuits, such as cerebellar structures, that are thought to engage in processing time-dependent signals and performing on-line prediction.

Poisson balanced spiking networks

An important problem in computational neuroscience is to understand how networks of spiking neurons can carry out various computations underlying behavior. Balanced spiking networks (BSNs) provide a powerful framework for implementing arbitrary linear dynamical systems in networks of integrate-and-fire neurons. However, the classic BSN model requires near-instantaneous transmission of spikes between neurons, which is biologically implausible. Introducing realistic synaptic delays leads to an pathological regime known as "ping-ponging", in which different populations spike maximally in alternating time bins, causing network output to overshoot the target solution. Here we document this phenomenon and provide a novel solution: we show that a network can have realistic synaptic delays while maintaining accuracy and stability if neurons are endowed with conditionally Poisson firing. Formally, we propose two alternate formulations of Poisson balanced spiking networks: (1) a "local" framework, which replaces the hard integrate-and-fire spiking rule within each neuron by a "soft" threshold function, such that firing probability grows as a smooth nonlinear function of membrane potential; and (2) a "population" framework, which reformulates the BSN objective function in terms of expected spike counts over the entire population. We show that both approaches offer improved robustness, allowing for accurate implementation of network dynamics with realistic synaptic delays between neurons. Both Poisson frameworks preserve the coding accuracy and robustness to neuron loss of the original model and, moreover, produce positive correlations between similarly tuned neurons, a feature of real neural populations that is not found in the deterministic BSN. This work unifies balanced spiking networks with Poisson generalized linear models and suggests several promising avenues for future research.