The equivalence of information-theoretic and likelihood-based methods for neural dimensionality reduction

Electrical Input Filters of Ganglion Cells in Wild Type and Degenerating rd10 Mouse Retina as a Template for Selective Electrical Stimulation

Bionic vision systems are currently limited by indiscriminate activation of all retinal ganglion cells (RGCs)- despite the dozens of known RGC types which each encode a different visual message. Here, we use spike-triggered averaging to explore how electrical responsiveness varies across RGC types toward the goal of using this variation to create type-selective electrical stimuli. A battery of visual stimuli and a randomly distributed sequence of electrical pulses were delivered to healthy and degenerating (4-week-old rd10) mouse retinas. Ganglion cell spike trains were recorded during stimulation using a 60-channel microelectrode array. Hierarchical clustering divided the recorded RGC populations according to their visual and electrical response patterns. Novel electrical stimuli were presented to assess type-specific selectivity. In healthy retinas, responses fell into 35 visual patterns and 14 electrical patterns. In degenerating retinas, responses fell into 12 visual and 23 electrical patterns. Few correspondences between electrical and visual response patterns were found except for the known correspondence of ON visual type with upward deflecting electrical type and OFF cells with downward electrical profiles. Further refinement of the approach presented here may yet yield the elusive nuances necessary for type-selective stimulation. This study greatly deepens our understanding of electrical input filters in the context of detailed visual response characterization and includes the most complete examination yet of degenerating electrical input filters.

A convolutional neural network provides a generalizable model of natural sound coding by neural populations in auditory cortex

Convolutional neural networks (CNNs) can provide powerful and flexible models of neural sensory processing. However, the utility of CNNs in studying the auditory system has been limited by their requirement for large datasets and the complex response properties of single auditory neurons. To address these limitations, we developed a population encoding model: a CNN that simultaneously predicts activity of several hundred neurons recorded during presentation of a large set of natural sounds. This approach defines a shared spectro-temporal space and pools statistical power across neurons. Population models of varying architecture performed consistently and substantially better than traditional linear-nonlinear models on data from primary and non-primary auditory cortex. Moreover, population models were highly generalizable. The output layer of a model pre-trained on one population of neurons could be fit to data from novel single units, achieving performance equivalent to that of neurons in the original fit data. This ability to generalize suggests that population encoding models capture a complete representational space across neurons in an auditory cortical field.

Quantitative models of auditory cortical processing

To generate insight from experimental data, it is critical to understand the inter-relationships between individual data points and place them in context within a structured framework. Quantitative modeling can provide the scaffolding for such an endeavor. Our main objective in this review is to provide a primer on the range of quantitative tools available to experimental auditory neuroscientists. Quantitative modeling is advantageous because it can provide a compact summary of observed data, make underlying assumptions explicit, and generate predictions for future experiments. Quantitative models may be developed to characterize or fit observed data, to test theories of how a task may be solved by neural circuits, to determine how observed biophysical details might contribute to measured activity patterns, or to predict how an experimental manipulation would affect neural activity. In complexity, quantitative models can range from those that are highly biophysically realistic and that include detailed simulations at the level of individual synapses, to those that use abstract and simplified neuron models to simulate entire networks. Here, we survey the landscape of recently developed models of auditory cortical processing, highlighting a small selection of models to demonstrate how they help generate insight into the mechanisms of auditory processing. We discuss examples ranging from models that use details of synaptic properties to explain the temporal pattern of cortical responses to those that use modern deep neural networks to gain insight into human fMRI data. We conclude by discussing a biologically realistic and interpretable model that our laboratory has developed to explore aspects of vocalization categorization in the auditory pathway.

How Stimulus Statistics Affect the Receptive Fields of Cells in Primary Visual Cortex

We studied the changes that neuronal receptive field (RF) models undergo when the statistics of the stimulus are changed from those of white Gaussian noise (WGN) to those of natural scenes (NSs), by fitting the models to multielectrode data recorded from primary visual cortex (V1) of female cats. This allowed the estimation of both a cascade of linear filters on the stimulus, as well as the static nonlinearities that map the output of the filters to the neuronal spike rates. We found that cells respond differently to these two classes of stimuli, with mostly higher spike rates and shorter response latencies to NSs than to WGN. The most striking finding was that NSs resulted in RFs that had additional uncovered filters compared with WGN. This finding was not an artifact of the higher spike rates observed for NSs relative to WGN, but rather was related to a change in coding. Our results reveal a greater extent of nonlinear processing in V1 neurons when stimulated using NSs compared with WGN. Our findings indicate the existence of nonlinear mechanisms that endow V1 neurons with context-dependent transmission of visual information.SIGNIFICANCE STATEMENT This study addresses a fundamental question about the concept of the receptive field (RF): does the encoding of information depend on the context or statistical regularities of the stimulus type? We applied state-of-the-art RF modeling techniques to data collected from multielectrode recordings from cat visual cortex in response to two statistically distinct stimulus types: white Gaussian noise and natural scenes. We find significant differences between the RFs that emerge from our data-driven modeling. Natural scenes result in far more complex RFs that combine multiple features in the visual input. Our findings reveal that different regimes or modes of operation are at work in visual cortical processing depending on the information present in the visual input. The complexity of V1 neural coding appears to be dependent on the complexity of the stimulus. We believe this new finding will have interesting implications for our understanding of the efficient transmission of information in sensory systems, which is an integral assumption of many computational theories (e.g., efficient and predictive coding of sensory processing in the brain).

Classification of pseudocalcium visual responses from mouse retinal ganglion cells

Recently, a detailed catalog of 32 retinal ganglion cell (RGC) visual response patterns in mouse has emerged. However, the 10,000 samples required for this catalog-based on fluorescent signals from a calcium indicator dye-are much harder to acquire from the extracellular spike train recordings underlying our bionic vision research. Therefore, we sought to convert spike trains into pseudocalcium signals so that our data could be directly matched to the 32 predefined, calcium signal-based groups. A microelectrode array (MEA) was used to record spike trains from mouse RGCs of 29 retinas. Visual stimuli were adapted from the Baden et al. study; including moving bars, full-field contrast and temporal frequency chirps, and black-white and UV-green color flashes. Spike train histograms were converted into pseudocalcium traces with an OGB-1 convolution kernel. Response features were extracted using sparse principal components analysis to match each RGC to one of the 32 RGC groups. These responses mapped onto of the 32 previously described groups; however, some of the groups remained unmatched. Thus, adaptation of the Baden et al. methodology for MEA recordings of spike trains instead of calcium recordings was partially successful. Different classification methods, however, will be needed to define clear RGC groups from MEA data for our bionic vision research. Nevertheless, others may pursue a pseudocalcium approach to reconcile spike trains with calcium signals. This work will help to guide them on the limitations and potential pitfalls of such an approach.

Continuous, multidimensional coding of 3D complex tactile stimuli by primary sensory neurons of the vibrissal system

Across all sensory modalities, first-stage sensory neurons are an information bottleneck: they must convey all information available for an animal to perceive and act in its environment. Our understanding of coding properties of primary sensory neurons in the auditory and visual systems has been aided by the use of increasingly complex, naturalistic stimulus sets. By comparison, encoding properties of primary somatosensory afferents are poorly understood. Here, we use the rodent whisker system to examine how tactile information is represented in primary sensory neurons of the trigeminal ganglion (Vg). Vg neurons have long been thought to segregate into functional classes associated with separate streams of information processing. However, this view is based on Vg responses to restricted stimulus sets which potentially underreport the coding capabilities of these neurons. In contrast, the current study records Vg responses to complex three-dimensional (3D) stimulation while quantifying the complete 3D whisker shape and mechanics, thereby beginning to reveal their full representational capabilities. The results show that individual Vg neurons simultaneously represent multiple mechanical features of a stimulus, do not preferentially encode principal components of the stimuli, and represent continuous and tiled variations of all available mechanical information. These results directly contrast with proposed codes in which subpopulations of Vg neurons encode select stimulus features. Instead, individual Vg neurons likely overcome the information bottleneck by encoding large regions of a complex sensory space. This proposed tiled and multidimensional representation at the Vg directly constrains the computations performed by more central neurons of the vibrissotrigeminal pathway.

Modeling statistical dependencies in multi-region spike train data

Neural computations underlying cognition and behavior rely on the coordination of neural activity across multiple brain areas. Understanding how brain areas interact to process information or generate behavior is thus a central question in neuroscience. Here we provide an overview of statistical approaches for characterizing statistical dependencies in multi-region spike train recordings. We focus on two classes of models in particular: regression-based models and shared latent variable models. Regression-based models describe interactions in terms of a directed transformation of information from one region to another. Shared latent variable models, on the other hand, seek to describe interactions in terms of sources that capture common fluctuations in spiking activity across regions. We discuss the advantages and limitations of each of these approaches and future directions for the field. We intend this review to be an introduction to the statistical methods in multi-region models for computational neuroscientists and experimentalists alike.

Complementary Effects of Adaptation and Gain Control on Sound Encoding in Primary Auditory Cortex

An important step toward understanding how the brain represents complex natural sounds is to develop accurate models of auditory coding by single neurons. A commonly used model is the linear-nonlinear spectro-temporal receptive field (STRF; LN model). The LN model accounts for many features of auditory tuning, but it cannot account for long-lasting effects of sensory context on sound-evoked activity. Two mechanisms that may support these contextual effects are short-term plasticity (STP) and contrast-dependent gain control (GC), which have inspired expanded versions of the LN model. Both models improve performance over the LN model, but they have never been compared directly. Thus, it is unclear whether they account for distinct processes or describe one phenomenon in different ways. To address this question, we recorded activity of neurons in primary auditory cortex (A1) of awake ferrets during presentation of natural sounds. We then fit models incorporating one nonlinear mechanism (GC or STP) or both (GC+STP) using this single dataset, and measured the correlation between the models' predictions and the recorded neural activity. Both the STP and GC models performed significantly better than the LN model, but the GC+STP model outperformed both individual models. We also quantified the equivalence of STP and GC model predictions and found only modest similarity. Consistent results were observed for a dataset collected in clean and noisy acoustic contexts. These results establish general methods for evaluating the equivalence of arbitrarily complex encoding models and suggest that the STP and GC models describe complementary processes in the auditory system.

Inferring synaptic inputs from spikes with a conductance-based neural encoding model

Descriptive statistical models of neural responses generally aim to characterize the mapping from stimuli to spike responses while ignoring biophysical details of the encoding process. Here, we introduce an alternative approach, the conductance-based encoding model (CBEM), which describes a mapping from stimuli to excitatory and inhibitory synaptic conductances governing the dynamics of sub-threshold membrane potential. Remarkably, we show that the CBEM can be fit to extracellular spike train data and then used to predict excitatory and inhibitory synaptic currents. We validate these predictions with intracellular recordings from macaque retinal ganglion cells. Moreover, we offer a novel quasi-biophysical interpretation of the Poisson generalized linear model (GLM) as a special case of the CBEM in which excitation and inhibition are perfectly balanced. This work forges a new link between statistical and biophysical models of neural encoding and sheds new light on the biophysical variables that underlie spiking in the early visual pathway.

Data-Driven Approaches to Understanding Visual Neuron Activity

With modern neurophysiological methods able to record neural activity throughout the visual pathway in the context of arbitrarily complex visual stimulation, our understanding of visual system function is becoming limited by the available models of visual neurons that can be directly related to such data. Different forms of statistical models are now being used to probe the cellular and circuit mechanisms shaping neural activity, understand how neural selectivity to complex visual features is computed, and derive the ways in which neurons contribute to systems-level visual processing. However, models that are able to more accurately reproduce observed neural activity often defy simple interpretations. As a result, rather than being used solely to connect with existing theories of visual processing, statistical modeling will increasingly drive the evolution of more sophisticated theories.

Characterizing and dissociating multiple time-varying modulatory computations influencing neuronal activity

In many brain areas, sensory responses are heavily modulated by factors including attentional state, context, reward history, motor preparation, learned associations, and other cognitive variables. Modelling the effect of these modulatory factors on sensory responses has proven challenging, mostly due to the time-varying and nonlinear nature of the underlying computations. Here we present a computational model capable of capturing and dissociating multiple time-varying modulatory effects on neuronal responses on the order of milliseconds. The model’s performance is tested on extrastriate perisaccadic visual responses in nonhuman primates. Visual neurons respond to stimuli presented around the time of saccades differently than during fixation. These perisaccadic changes include sensitivity to the stimuli presented at locations outside the neuron’s receptive field, which suggests a contribution of multiple sources to perisaccadic response generation. Current computational approaches cannot quantitatively characterize the contribution of each modulatory source in response generation, mainly due to the very short timescale on which the saccade takes place. In this study, we use a high spatiotemporal resolution experimental paradigm along with a novel extension of the generalized linear model framework (GLM), termed the sparse-variable GLM, to allow for time-varying model parameters representing the temporal evolution of the system with a resolution on the order of milliseconds. We used this model framework to precisely map the temporal evolution of the spatiotemporal receptive field of visual neurons in the middle temporal area during the execution of a saccade. Moreover, an extended model based on a factorization of the sparse-variable GLM allowed us to disassociate and quantify the contribution of individual sources to the perisaccadic response. Our results show that our novel framework can precisely capture the changes in sensitivity of neurons around the time of saccades, and provide a general framework to quantitatively track the role of multiple modulatory sources over time.

The sensory responses of neurons in many brain areas, particularly those in higher prefrontal or parietal areas, are strongly influenced by factors including task rules, attentional state, context, reward history, motor preparation, learned associations, and other cognitive variables. These modulations often occur in combination, or on fast timescales which present a challenge for both experimental and modelling approaches aiming to describe the underlying mechanisms or computations. Here we present a computational model capable of capturing and dissociating multiple time-varying modulatory effects on spiking responses on the order of milliseconds. The model’s performance is evaluated by testing its ability to reproduce and dissociate multiple changes in visual sensitivity occurring in extrastriate visual cortex around the time of rapid eye movements. No previous model is capable of capturing these changes with as fine a resolution as that presented here. Our model both provides specific insight into the nature and time course of changes in visual sensitivity around the time of eye movements, and offers a general framework applicable to a wide variety of contexts in which sensory processing is modulated dynamically by multiple time-varying cognitive or behavioral factors, to understand the neuronal computations underpinning these modulations and make predictions about the underlying mechanisms.

Multi-Input, Multi-Output Neuronal Mode Network Approach to Modeling the Encoding Dynamics and Functional Connectivity of Neural Systems

This letter proposes a novel method, multi-input, multi-output neuronal mode network (MIMO-NMN), for modeling encoding dynamics and functional connectivity in neural ensembles such as the hippocampus. Compared with conventional approaches such as the Volterra-Wiener model, linear-nonlinear-cascade (LNC) model, and generalized linear model (GLM), the NMN has several advantages in terms of estimation accuracy, model interpretation, and functional connectivity analysis. We point out the limitations of current neural spike modeling methods, especially the estimation biases caused by the imbalanced class problem when the number of zeros is significantly larger than ones in the spike data. We use synthetic data to test the performance of NMN with a comparison of the traditional methods, and the results indicate the NMN approach could reduce the imbalanced class problem and achieve better predictions. Subsequently, we apply the MIMO-NMN method to analyze data from the human hippocampus. The results indicate that the MIMO-NMN method is a promising approach to modeling neural dynamics and analyzing functional connectivity of multi-neuronal data.

Inferring single-trial neural population dynamics using sequential auto-encoders

Neuroscience is experiencing a revolution in which simultaneous recording of many thousands of neurons is revealing population dynamics that are not apparent from single-neuron responses. This structure is typically extracted from trial-averaged data, but deeper understanding requires studying single-trial phenomena, which is challenging due to incomplete sampling of the neural population, trial-to-trial variability, and fluctuations in action potential timing. We introduce Latent Factor Analysis via Dynamical Systems (LFADS), a deep learning method to infer latent dynamics from single-trial neural spiking data. LFADS uses a nonlinear dynamical system to infer the dynamics underlying observed spiking activity and to extract ‘de-noised’ single-trial firing rates. When applied to a variety of monkey and human motor cortical datasets, LFADS predicts observed behavioral variables with unprecedented accuracy, extracts precise estimates of neural dynamics on single trials, infers perturbations to those dynamics that correlate with behavioral choices, and combines data from non-overlapping recording sessions spanning months to improve inference of underlying dynamics.

Describing complex cells in primary visual cortex: a comparison of context and multifilter LN models

Receptive field (RF) models are an important tool for deciphering neural responses to sensory stimuli. The two currently popular RF models are multifilter linear-nonlinear (LN) models and context models. Models are, however, never correct, and they rely on assumptions to keep them simple enough to be interpretable. As a consequence, different models describe different stimulus-response mappings, which may or may not be good approximations of real neural behavior. In the current study, we take up two tasks: 1) we introduce new ways to estimate context models with realistic nonlinearities, that is, with logistic and exponential functions, and 2) we evaluate context models and multifilter LN models in terms of how well they describe recorded data from complex cells in cat primary visual cortex. Our results, based on single-spike information and correlation coefficients, indicate that context models outperform corresponding multifilter LN models of equal complexity (measured in terms of number of parameters), with the best increase in performance being achieved by the novel context models. Consequently, our results suggest that the multifilter LN-model framework is suboptimal for describing the behavior of complex cells: the context-model framework is clearly superior while still providing interpretable quantizations of neural behavior. NEW & NOTEWORTHY We used data from complex cells in primary visual cortex to estimate a wide variety of receptive field models from two frameworks that have previously not been compared with each other. The models included traditionally used multifilter linear-nonlinear models and novel variants of context models. Using mutual information and correlation coefficients as performance measures, we showed that context models are superior for describing complex cells and that the novel context models performed the best.

Modern Machine Learning as a Benchmark for Fitting Neural Responses

Neuroscience has long focused on finding encoding models that effectively ask "what predicts neural spiking?" and generalized linear models (GLMs) are a typical approach. It is often unknown how much of explainable neural activity is captured, or missed, when fitting a model. Here we compared the predictive performance of simple models to three leading machine learning methods: feedforward neural networks, gradient boosted trees (using XGBoost), and stacked ensembles that combine the predictions of several methods. We predicted spike counts in macaque motor (M1) and somatosensory (S1) cortices from standard representations of reaching kinematics, and in rat hippocampal cells from open field location and orientation. Of these methods, XGBoost and the ensemble consistently produced more accurate spike rate predictions and were less sensitive to the preprocessing of features. These methods can thus be applied quickly to detect if feature sets relate to neural activity in a manner not captured by simpler methods. Encoding models built with a machine learning approach accurately predict spike rates and can offer meaningful benchmarks for simpler models.

Capturing the Dynamical Repertoire of Single Neurons with Generalized Linear Models

A key problem in computational neuroscience is to find simple, tractable models that are nevertheless flexible enough to capture the response properties of real neurons. Here we examine the capabilities of recurrent point process models known as Poisson generalized linear models (GLMs). These models are defined by a set of linear filters and a point nonlinearity and are conditionally Poisson spiking. They have desirable statistical properties for fitting and have been widely used to analyze spike trains from electrophysiological recordings. However, the dynamical repertoire of GLMs has not been systematically compared to that of real neurons. Here we show that GLMs can reproduce a comprehensive suite of canonical neural response behaviors, including tonic and phasic spiking, bursting, spike rate adaptation, type I and type II excitation, and two forms of bistability. GLMs can also capture stimulus-dependent changes in spike timing precision and reliability that mimic those observed in real neurons, and can exhibit varying degrees of stochasticity, from virtually deterministic responses to greater-than-Poisson variability. These results show that Poisson GLMs can exhibit a wide range of dynamic spiking behaviors found in real neurons, making them well suited for qualitative dynamical as well as quantitative statistical studies of single-neuron and population response properties.

Neuron's eye view: Inferring features of complex stimuli from neural responses

Experiments that study neural encoding of stimuli at the level of individual neurons typically choose a small set of features present in the world—contrast and luminance for vision, pitch and intensity for sound—and assemble a stimulus set that systematically varies along these dimensions. Subsequent analysis of neural responses to these stimuli typically focuses on regression models, with experimenter-controlled features as predictors and spike counts or firing rates as responses. Unfortunately, this approach requires knowledge in advance about the relevant features coded by a given population of neurons. For domains as complex as social interaction or natural movement, however, the relevant feature space is poorly understood, and an arbitrary a priori choice of features may give rise to confirmation bias. Here, we present a Bayesian model for exploratory data analysis that is capable of automatically identifying the features present in unstructured stimuli based solely on neuronal responses. Our approach is unique within the class of latent state space models of neural activity in that it assumes that firing rates of neurons are sensitive to multiple discrete time-varying features tied to the stimulus, each of which has Markov (or semi-Markov) dynamics. That is, we are modeling neural activity as driven by multiple simultaneous stimulus features rather than intrinsic neural dynamics. We derive a fast variational Bayesian inference algorithm and show that it correctly recovers hidden features in synthetic data, as well as ground-truth stimulus features in a prototypical neural dataset. To demonstrate the utility of the algorithm, we also apply it to cluster neural responses and demonstrate successful recovery of features corresponding to monkeys and faces in the image set.

Many neuroscience experiments begin with a set of reduced stimuli designed to vary only along a small set of variables. Yet many phenomena of interest—natural movies, objects—are not easily parameterized by a small number of dimensions. Here, we develop a novel Bayesian model for clustering stimuli based solely on neural responses, allowing us to discover which latent features of complex stimuli actually drive neural activity. We demonstrate that this model allows us to recover key features of neural responses in a pair of well-studied paradigms.

Models of Neuronal Stimulus-Response Functions: Elaboration, Estimation, and Evaluation

Rich, dynamic, and dense sensory stimuli are encoded within the nervous system by the time-varying activity of many individual neurons. A fundamental approach to understanding the nature of the encoded representation is to characterize the function that relates the moment-by-moment firing of a neuron to the recent history of a complex sensory input. This review provides a unifying and critical survey of the techniques that have been brought to bear on this effort thus far—ranging from the classical linear receptive field model to modern approaches incorporating normalization and other nonlinearities. We address separately the structure of the models; the criteria and algorithms used to identify the model parameters; and the role of regularizing terms or “priors.” In each case we consider benefits or drawbacks of various proposals, providing examples for when these methods work and when they may fail. Emphasis is placed on key concepts rather than mathematical details, so as to make the discussion accessible to readers from outside the field. Finally, we review ways in which the agreement between an assumed model and the neuron's response may be quantified. Re-implemented and unified code for many of the methods are made freely available.

Network Receptive Field Modeling Reveals Extensive Integration and Multi-feature Selectivity in Auditory Cortical Neurons

Cortical sensory neurons are commonly characterized using the receptive field, the linear dependence of their response on the stimulus. In primary auditory cortex neurons can be characterized by their spectrotemporal receptive fields, the spectral and temporal features of a sound that linearly drive a neuron. However, receptive fields do not capture the fact that the response of a cortical neuron results from the complex nonlinear network in which it is embedded. By fitting a nonlinear feedforward network model (a network receptive field) to cortical responses to natural sounds, we reveal that primary auditory cortical neurons are sensitive over a substantially larger spectrotemporal domain than is seen in their standard spectrotemporal receptive fields. Furthermore, the network receptive field, a parsimonious network consisting of 1-7 sub-receptive fields that interact nonlinearly, consistently better predicts neural responses to auditory stimuli than the standard receptive fields. The network receptive field reveals separate excitatory and inhibitory sub-fields with different nonlinear properties, and interaction of the sub-fields gives rise to important operations such as gain control and conjunctive feature detection. The conjunctive effects, where neurons respond only if several specific features are present together, enable increased selectivity for particular complex spectrotemporal structures, and may constitute an important stage in sound recognition. In conclusion, we demonstrate that fitting auditory cortical neural responses with feedforward network models expands on simple linear receptive field models in a manner that yields substantially improved predictive power and reveals key nonlinear aspects of cortical processing, while remaining easy to interpret in a physiological context.

A Novel Nonparametric Approach for Neural Encoding and Decoding Models of Multimodal Receptive Fields

Pyramidal neurons recorded from the rat hippocampus and entorhinal cortex, such as place and grid cells, have diverse receptive fields, which are either unimodal or multimodal. Spiking activity from these cells encodes information about the spatial position of a freely foraging rat. At fine timescales, a neuron's spike activity also depends significantly on its own spike history. However, due to limitations of current parametric modeling approaches, it remains a challenge to estimate complex, multimodal neuronal receptive fields while incorporating spike history dependence. Furthermore, efforts to decode the rat's trajectory in one- or two-dimensional space from hippocampal ensemble spiking activity have mainly focused on spike history-independent neuronal encoding models. In this letter, we address these two important issues by extending a recently introduced nonparametric neural encoding framework that allows modeling both complex spatial receptive fields and spike history dependencies. Using this extended nonparametric approach, we develop novel algorithms for decoding a rat's trajectory based on recordings of hippocampal place cells and entorhinal grid cells. Results show that both encoding and decoding models derived from our new method performed significantly better than state-of-the-art encoding and decoding models on 6 minutes of test data. In addition, our model's performance remains invariant to the apparent modality of the neuron's receptive field.

The Essential Complexity of Auditory Receptive Fields

Encoding properties of sensory neurons are commonly modeled using linear finite impulse response (FIR) filters. For the auditory system, the FIR filter is instantiated in the spectro-temporal receptive field (STRF), often in the framework of the generalized linear model. Despite widespread use of the FIR STRF, numerous formulations for linear filters are possible that require many fewer parameters, potentially permitting more efficient and accurate model estimates. To explore these alternative STRF architectures, we recorded single-unit neural activity from auditory cortex of awake ferrets during presentation of natural sound stimuli. We compared performance of > 1000 linear STRF architectures, evaluating their ability to predict neural responses to a novel natural stimulus. Many were able to outperform the FIR filter. Two basic constraints on the architecture lead to the improved performance: (1) factorization of the STRF matrix into a small number of spectral and temporal filters and (2) low-dimensional parameterization of the factorized filters. The best parameterized model was able to outperform the full FIR filter in both primary and secondary auditory cortex, despite requiring fewer than 30 parameters, about 10% of the number required by the FIR filter. After accounting for noise from finite data sampling, these STRFs were able to explain an average of 40% of A1 response variance. The simpler models permitted more straightforward interpretation of sensory tuning properties. They also showed greater benefit from incorporating nonlinear terms, such as short term plasticity, that provide theoretical advances over the linear model. Architectures that minimize parameter count while maintaining maximum predictive power provide insight into the essential degrees of freedom governing auditory cortical function. They also maximize statistical power available for characterizing additional nonlinear properties that limit current auditory models.

Understanding how the brain solves sensory problems can provide useful insight for the development of automated systems such as speech recognizers and image classifiers. Recent developments in nonlinear regression and machine learning have produced powerful algorithms for characterizing the input-output relationship of complex systems. However, the complexity of sensory neural systems, combined with practical limitations on experimental data, make it difficult to apply arbitrarily complex analyses to neural data. In this study we pushed analysis in the opposite direction, toward simpler models. We asked how simple a model can be while still capturing the essential sensory properties of neurons in auditory cortex. We found that substantially simpler formulations of the widely-used spectro-temporal receptive field are able to perform as well as the best current models. These simpler formulations define new basis sets that can be incorporated into state-of-the-art machine learning algorithms for a more exhaustive exploration of sensory processing.