Encoding surprise by retinal ganglion cells.

The efficient coding hypothesis posits that early sensory neurons transmit maximal information about sensory stimuli, given internal constraints. A central prediction of this theory is that neurons should preferentially encode stimuli that are most surprising. Previous studies suggest this may be the case in early visual areas, where many neurons respond strongly to rare or surprising stimuli. For example, previous research showed that when presented with a rhythmic sequence of full-field flashes, many retinal ganglion cells (RGCs) respond strongly at the instance the flash sequence stops, and when another flash would be expected. This phenomenon is called the 'omitted stimulus response'. However, it is not known whether the responses of these cells varies in a graded way depending on the level of stimulus surprise. To investigate this, we presented retinal neurons with extended sequences of stochastic flashes. With this stimulus, the surprise associated with a particular flash/silence, could be quantified analytically, and varied in a graded manner depending on the previous sequences of flashes and silences. Interestingly, we found that RGC responses could be well explained by a simple normative model, which described how they optimally combined their prior expectations and recent stimulus history, so as to encode surprise. Further, much of the diversity in RGC responses could be explained by the model, due to the different prior expectations that different neurons had about the stimulus statistics. These results suggest that even as early as the retina many cells encode surprise, relative to their own, internally generated expectations.

Scalable Gaussian process inference of neural responses to natural images.

Predicting the responses of sensory neurons is a long-standing neuroscience goal. However, while there has been much progress in modeling neural responses to simple and/or artificial stimuli, predicting responses to natural stimuli remains an ongoing challenge. On the one hand, deep neural networks perform very well on certain datasets but can fail when data are limited. On the other hand, Gaussian processes (GPs) perform well on limited data but are poor at predicting responses to high-dimensional stimuli, such as natural images. Here, we show how structured priors, e.g., for local and smooth receptive fields, can be used to scale up GPs to model neural responses to high-dimensional stimuli. With this addition, GPs largely outperform a deep neural network trained to predict retinal responses to natural images, with the largest differences observed when both models are trained on a small dataset. Further, since they allow us to quantify the uncertainty in their predictions, GPs are well suited to closed-loop experiments, where stimuli are chosen actively so as to collect "informative" neural data. We show how GPs can be used to actively select which stimuli to present, so as to i) efficiently learn a model of retinal responses to natural images, using few data, and ii) rapidly distinguish between competing models (e.g., a linear vs. a nonlinear model). In the future, our approach could be applied to other sensory areas, beyond the retina.

Human-in-the-Loop Optimization for Deep Stimulus Encoding in Visual Prostheses.

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 optimize 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. We show that our approach quickly learns a personalized stimulus encoder, leads to dramatic improvements in the quality of restored vision, and 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.

Human-in-the-loop optimization of visual prosthetic stimulation.

Objective. Retinal prostheses are a promising strategy to restore sight to patients with retinal degenerative diseases. These devices compensate for the loss of photoreceptors by electrically stimulating neurons in the retina. Currently, the visual function that can be recovered with such devices is very limited. This is due, in part, to current spread, unintended axonal activation, and the limited resolution of existing devices. Here we show, using a recent model of prosthetic vision, that optimizing how visual stimuli are encoded by the device can help overcome some of these limitations, leading to dramatic improvements in visual perception.Approach. We propose a strategy to do this in practice, using patients' feedback in a visual task. The main challenge of our approach comes from the fact that, typically, one only has access to a limited number of noisy responses from patients. We propose two ways to deal with this: first, we use a model of prosthetic vision to constrain and simplify the optimization. We show that, if one knew the parameters of this model for a given patient, it would be possible to greatly improve their perceptual performance. Second we propose a preferential Bayesian optimization to efficiently learn these model parameters for each patient, using minimal trials.Main results. To test our approach, we presented healthy subjects with visual stimuli generated by a recent model of prosthetic vision, to replicate the perceptual experience of patients fitted with an implant. Our optimization procedure led to significant and robust improvements in perceived image quality, that transferred to increased performance in other tasks.Significance. Importantly, our strategy is agnostic to the type of prosthesis and thus could readily be implemented in existing implants.

Inferring the function performed by a recurrent neural network.

A central goal in systems neuroscience is to understand the functions performed by neural circuits. Previous top-down models addressed this question by comparing the behaviour of an ideal model circuit, optimised to perform a given function, with neural recordings. However, this requires guessing in advance what function is being performed, which may not be possible for many neural systems. To address this, we propose an inverse reinforcement learning (RL) framework for inferring the function performed by a neural network from data. We assume that the responses of each neuron in a network are optimised so as to drive the network towards 'rewarded' states, that are desirable for performing a given function. We then show how one can use inverse RL to infer the reward function optimised by the network from observing its responses. This inferred reward function can be used to predict how the neural network should adapt its dynamics to perform the same function when the external environment or network structure changes. This could lead to theoretical predictions about how neural network dynamics adapt to deal with cell death and/or varying sensory stimulus statistics.

Toward a unified theory of efficient, predictive, and sparse coding.

A central goal in theoretical neuroscience is to predict the response properties of sensory neurons from first principles. To this end, "efficient coding" posits that sensory neurons encode maximal information about their inputs given internal constraints. There exist, however, many variants of efficient coding (e.g., redundancy reduction, different formulations of predictive coding, robust coding, sparse coding, etc.), differing in their regimes of applicability, in the relevance of signals to be encoded, and in the choice of constraints. It is unclear how these types of efficient coding relate or what is expected when different coding objectives are combined. Here we present a unified framework that encompasses previously proposed efficient coding models and extends to unique regimes. We show that optimizing neural responses to encode predictive information can lead them to either correlate or decorrelate their inputs, depending on the stimulus statistics; in contrast, at low noise, efficiently encoding the past always predicts decorrelation. Later, we investigate coding of naturalistic movies and show that qualitatively different types of visual motion tuning and levels of response sparsity are predicted, depending on whether the objective is to recover the past or predict the future. Our approach promises a way to explain the observed diversity of sensory neural responses, as due to multiple functional goals and constraints fulfilled by different cell types and/or circuits.

Sensory noise predicts divisive reshaping of receptive fields.

In order to respond reliably to specific features of their environment, sensory neurons need to integrate multiple incoming noisy signals. Crucially, they also need to compete for the interpretation of those signals with other neurons representing similar features. The form that this competition should take depends critically on the noise corrupting these signals. In this study we show that for the type of noise commonly observed in sensory systems, whose variance scales with the mean signal, sensory neurons should selectively divide their input signals by their predictions, suppressing ambiguous cues while amplifying others. Any change in the stimulus context alters which inputs are suppressed, leading to a deep dynamic reshaping of neural receptive fields going far beyond simple surround suppression. Paradoxically, these highly variable receptive fields go alongside and are in fact required for an invariant representation of external sensory features. In addition to offering a normative account of context-dependent changes in sensory responses, perceptual inference in the presence of signal-dependent noise accounts for ubiquitous features of sensory neurons such as divisive normalization, gain control and contrast dependent temporal dynamics.

Neural oscillations as a signature of efficient coding in the presence of synaptic delays.

Cortical networks exhibit 'global oscillations', in which neural spike times are entrained to an underlying oscillatory rhythm, but where individual neurons fire irregularly, on only a fraction of cycles. While the network dynamics underlying global oscillations have been well characterised, their function is debated. Here, we show that such global oscillations are a direct consequence of optimal efficient coding in spiking networks with synaptic delays and noise. To avoid firing unnecessary spikes, neurons need to share information about the network state. Ideally, membrane potentials should be strongly correlated and reflect a 'prediction error' while the spikes themselves are uncorrelated and occur rarely. We show that the most efficient representation is when: (i) spike times are entrained to a global Gamma rhythm (implying a consistent representation of the error); but (ii) few neurons fire on each cycle (implying high efficiency), while (iii) excitation and inhibition are tightly balanced. This suggests that cortical networks exhibiting such dynamics are tuned to achieve a maximally efficient population code.

Efficiency turns the table on neural encoding, decoding and noise.

Sensory neurons are usually described with an encoding model, for example, a function that predicts their response from the sensory stimulus using a receptive field (RF) or a tuning curve. However, central to theories of sensory processing is the notion of 'efficient coding'. We argue here that efficient coding implies a completely different neural coding strategy. Instead of a fixed encoding model, neural populations would be described by a fixed decoding model (i.e. a model reconstructing the stimulus from the neural responses). Because the population solves a global optimization problem, individual neurons are variable, but not noisy, and have no truly invariant tuning curve or receptive field. We review recent experimental evidence and implications for neural noise correlations, robustness and adaptation.

Attention as reward-driven optimization of sensory processing.

Attention causes diverse changes to visual neuron responses, including alterations in receptive field structure, and firing rates. A common theoretical approach to investigate why sensory neurons behave as they do is based on the efficient coding hypothesis: that sensory processing is optimized toward the statistics of the received input. We extend this approach to account for the influence of task demands, hypothesizing that the brain learns a probabilistic model of both the sensory input and reward received for performing different actions. Attention-dependent changes to neural responses reflect optimization of this internal model to deal with changes in the sensory environment (stimulus statistics) and behavioral demands (reward statistics). We use this framework to construct a simple model of visual processing that is able to replicate a number of attention-dependent changes to the responses of neurons in the midlevel visual cortices. The model is consistent with and provides a normative explanation for recent divisive normalization models of attention (Reynolds & Heeger, 2009).

Complexity and specificity of experimentally-induced expectations in motion perception.

Our perceptions are fundamentally altered by our expectations, i.e., priors about the world. In previous statistical learning experiments (Chalk, Seitz, & Seriès, 2010), we investigated how such priors are formed by presenting subjects with white low contrast moving dots on a blank screen and using a bimodal distribution of motion directions such that two directions were more frequently presented than the others. We found that human observers quickly and automatically developed expectations for the most frequently presented directions of motion. Here, we examine the specificity of these expectations. Can one learn simultaneously to expect different motion directions for dots of different colors? We interleaved moving dot displays of two different colors, either red or green, with different motion direction distributions. When one distribution was bimodal while the other was uniform, we found that subjects learned a single bimodal prior for the two stimuli. On the contrary, when both distributions were similarly structured, we found evidence for the formation of two distinct priors, which significantly influenced the subjects' behavior when no stimulus was presented. Our results can be modeled using a Bayesian framework and discussed in terms of a suboptimality of the statistical learning process under some conditions.

Rapidly learned stimulus expectations alter perception of motion.

Expectations broadly influence our experience of the world. However, the process by which they are acquired and then shape our sensory experiences is not well understood. Here, we examined whether expectations of simple stimulus features can be developed implicitly through a fast statistical learning procedure. We found that participants quickly and automatically developed expectations for the most frequently presented directions of motion and that this altered their perception of new motion directions, inducing attractive biases in the perceived direction as well as visual hallucinations in the absence of a stimulus. Further, the biases in motion direction estimation that we observed were well explained by a model that accounted for participants' behavior using a Bayesian strategy, combining a learned prior of the stimulus statistics (the expectation) with their sensory evidence (the actual stimulus) in a probabilistically optimal manner. Our results demonstrate that stimulus expectations are rapidly learned and can powerfully influence perception of simple visual features.

Attention reduces stimulus-driven gamma frequency oscillations and spike field coherence in V1.

Rhythmic activity of neuronal ensembles has been proposed to play an important role in cognitive functions such as attention, perception, and memory. Here we investigate whether rhythmic activity in V1 of the macaque monkey (macaca mulatta) is affected by top-down visual attention. We measured the local field potential (LFP) and V1 spiking activity while monkeys performed an attention-demanding detection task. We show that gamma oscillations were strongly modulated by the stimulus and by attention. Stimuli that engaged inhibitory mechanisms induced the largest gamma LFP oscillations and the largest spike field coherence. Directing attention toward a visual stimulus at the receptive field of the recorded neurons decreased LFP gamma power and gamma spike field coherence. This decrease could reflect an attention-mediated reduction of surround inhibition. Changes in synchrony in V1 would thus be a byproduct of reduced inhibitory drive, rather than a mechanism that directly aids perceptual processing.