Active Learning for Discrete Latent Variable Models.

Active learning seeks to reduce the amount of data required to fit the parameters of a model, thus forming an important class of techniques in modern machine learning. However, past work on active learning has largely overlooked latent variable models, which play a vital role in neuroscience, psychology, and a variety of other engineering and scientific disciplines. Here we address this gap by proposing a novel framework for maximum-mutual-information input selection for discrete latent variable regression models. We first apply our method to a class of models known as mixtures of linear regressions (MLR). While it is well known that active learning confers no advantage for linear-gaussian regression models, we use Fisher information to show analytically that active learning can nevertheless achieve large gains for mixtures of such models, and we validate this improvement using both simulations and real-world data. We then consider a powerful class of temporally structured latent variable models given by a hidden Markov model (HMM) with generalized linear model (GLM) observations, which has recently been used to identify discrete states from animal decision-making data. We show that our method substantially reduces the amount of data needed to fit GLM-HMMs and outperforms a variety of approximate methods based on variational and amortized inference. Infomax learning for latent variable models thus offers a powerful approach for characterizing temporally structured latent states, with a wide variety of applications in neuroscience and beyond.

From connectome to effectome: learning the causal interaction map of the fly brain.

A long-standing goal of neuroscience is to obtain a causal model of the nervous system. This would allow neuroscientists to explain animal behavior in terms of the dynamic interactions between neurons. The recently reported whole-brain fly connectome [1-7] specifies the synaptic paths by which neurons can affect each other but not whether, or how, they do affect each other in vivo. To overcome this limitation, we introduce a novel combined experimental and statistical strategy for efficiently learning a causal model of the fly brain, which we refer to as the "effectome". Specifically, we propose an estimator for a dynamical systems model of the fly brain that uses stochastic optogenetic perturbation data to accurately estimate causal effects and the connectome as a prior to drastically improve estimation efficiency. We then analyze the connectome to propose circuits that have the greatest total effect on the dynamics of the fly nervous system. We discover that, fortunately, the dominant circuits significantly involve only relatively small populations of neurons-thus imaging, stimulation, and neuronal identification are feasible. Intriguingly, we find that this approach also re-discovers known circuits and generates testable hypotheses about their dynamics. Overall, our analyses of the connectome provide evidence that global dynamics of the fly brain are generated by a large collection of small and often anatomically localized circuits operating, largely, independently of each other. This in turn implies that a causal model of a brain, a principal goal of systems neuroscience, can be feasibly obtained in the fly.

Olfactory learning alters navigation strategies and behavioral variability in C. elegans.

Animals adjust their behavioral response to sensory input adaptively depending on past experiences. The flexible brain computation is crucial for survival and is of great interest in neuroscience. The nematode C. elegans modulates its navigation behavior depending on the association of odor butanone with food (appetitive training) or starvation (aversive training), and will then climb up the butanone gradient or ignore it, respectively. However, the exact change in navigation strategy in response to learning is still unknown. Here we study the learned odor navigation in worms by combining precise experimental measurement and a novel descriptive model of navigation. Our model consists of two known navigation strategies in worms: biased random walk and weathervaning. We infer weights on these strategies by applying the model to worm navigation trajectories and the exact odor concentration it experiences. Compared to naive worms, appetitive trained worms up-regulate the biased random walk strategy, and aversive trained worms down-regulate the weathervaning strategy. The statistical model provides prediction with $>90 \%$ accuracy of the past training condition given navigation data, which outperforms the classical chemotaxis metric. We find that the behavioral variability is altered by learning, such that worms are less variable after training compared to naive ones. The model further predicts the learning-dependent response and variability under optogenetic perturbation of the olfactory neuron AWC$^\mathrm{ON}$. Lastly, we investigate neural circuits downstream from AWC$^\mathrm{ON}$ that are differentially recruited for learned odor-guided navigation. Together, we provide a new paradigm to quantify flexible navigation algorithms and pinpoint the underlying neural substrates.

Efficient Decoding of Large-Scale Neural Population Responses With Gaussian-Process Multiclass Regression.

Neural decoding methods provide a powerful tool for quantifying the information content of neural population codes and the limits imposed by correlations in neural activity. However, standard decoding methods are prone to overfitting and scale poorly to high-dimensional settings. Here, we introduce a novel decoding method to overcome these limitations. Our approach, the gaussian process multiclass decoder (GPMD), is well suited to decoding a continuous low-dimensional variable from high-dimensional population activity and provides a platform for assessing the importance of correlations in neural population codes. The GPMD is a multinomial logistic regression model with a gaussian process prior over the decoding weights. The prior includes hyperparameters that govern the smoothness of each neuron's decoding weights, allowing automatic pruning of uninformative neurons during inference. We provide a variational inference method for fitting the GPMD to data, which scales to hundreds or thousands of neurons and performs well even in data sets with more neurons than trials. We apply the GPMD to recordings from primary visual cortex in three species: monkey, ferret, and mouse. Our decoder achieves state-of-the-art accuracy on all three data sets and substantially outperforms independent Bayesian decoding, showing that knowledge of the correlation structure is essential for optimal decoding in all three species.

Compact deep neural network models of visual cortex.

A powerful approach to understanding the computations carried out in visual cortex is to develop models that predict neural responses to arbitrary images. Deep neural network (DNN) models have worked remarkably well at predicting neural responses [1, 2, 3], yet their underlying computations remain buried in millions of parameters. Have we simply replaced one complicated system in vivo with another in silico? Here, we train a data-driven deep ensemble model that predicts macaque V4 responses ~50% more accurately than currently-used task-driven DNN models. We then compress this deep ensemble to identify compact models that have 5,000x fewer parameters yet equivalent accuracy as the deep ensemble. We verified that the stimulus preferences of the compact models matched those of the real V4 neurons by measuring V4 responses to both 'maximizing' and adversarial images generated using compact models. We then analyzed the inner workings of the compact models and discovered a common circuit motif: Compact models share a similar set of filters in early stages of processing but then specialize by heavily consolidating this shared representation with a precise readout. This suggests that a V4 neuron's stimulus preference is determined entirely by its consolidation step. To demonstrate this, we investigated the compression step of a dot-detecting compact model and found a set of simple computations that may be carried out by dot-selective V4 neurons. Overall, our work demonstrates that the DNN models currently used in computational neuroscience are needlessly large; our approach provides a new way forward for obtaining explainable, high-accuracy models of visual cortical neurons.

Flow-field inference from neural data using deep recurrent networks.

Computations involved in processes such as decision-making, working memory, and motor control are thought to emerge from the dynamics governing the collective activity of neurons in large populations. But the estimation of these dynamics remains a significant challenge. Here we introduce Flow-field Inference from Neural Data using deep Recurrent networks (FINDR), an unsupervised deep learning method that can infer low-dimensional nonlinear stochastic dynamics underlying neural population activity. Using population spike train data from frontal brain regions of rats performing an auditory decision-making task, we demonstrate that FINDR outperforms existing methods in capturing the heterogeneous responses of individual neurons. We further show that FINDR can discover interpretable low-dimensional dynamics when it is trained to disentangle task-relevant and irrelevant components of the neural population activity. Importantly, the low-dimensional nature of the learned dynamics allows for explicit visualization of flow fields and attractor structures. We suggest FINDR as a powerful method for revealing the low-dimensional task-relevant dynamics of neural populations and their associated computations.

Temporal integration is a robust feature of perceptual decisions.

Making informed decisions in noisy environments requires integrating sensory information over time. However, recent work has suggested that it may be difficult to determine whether an animal's decision-making strategy relies on evidence integration or not. In particular, strategies based on extrema-detection or random snapshots of the evidence stream may be difficult or even impossible to distinguish from classic evidence integration. Moreover, such non-integration strategies might be surprisingly common in experiments that aimed to study decisions based on integration. To determine whether temporal integration is central to perceptual decision-making, we developed a new model-based approach for comparing temporal integration against alternative 'non-integration' strategies for tasks in which the sensory signal is composed of discrete stimulus samples. We applied these methods to behavioral data from monkeys, rats, and humans performing a variety of sensory decision-making tasks. In all species and tasks, we found converging evidence in favor of temporal integration. First, in all observers across studies, the integration model better accounted for standard behavioral statistics such as psychometric curves and psychophysical kernels. Second, we found that sensory samples with large evidence do not contribute disproportionately to subject choices, as predicted by an extrema-detection strategy. Finally, we provide a direct confirmation of temporal integration by showing that the sum of both early and late evidence contributed to observer decisions. Overall, our results provide experimental evidence suggesting that temporal integration is an ubiquitous feature in mammalian perceptual decision-making. Our study also highlights the benefits of using experimental paradigms where the temporal stream of sensory evidence is controlled explicitly by the experimenter, and known precisely by the analyst, to characterize the temporal properties of the decision process.

Scalable Variational Inference for Low-Rank Spatiotemporal Receptive Fields.

An important problem in systems neuroscience is to characterize how a neuron integrates sensory inputs across space and time. The linear receptive field provides a mathematical characterization of this weighting function and is commonly used to quantify neural response properties and classify cell types. However, estimating receptive fields is difficult in settings with limited data and correlated or high-dimensional stimuli. To overcome these difficulties, we propose a hierarchical model designed to flexibly parameterize low-rank receptive fields. The model includes gaussian process priors over spatial and temporal components of the receptive field, encouraging smoothness in space and time. We also propose a new temporal prior, temporal relevance determination, which imposes a variable degree of smoothness as a function of time lag. We derive a scalable algorithm for variational Bayesian inference for both spatial and temporal receptive field components and hyperparameters. The resulting estimator scales to high-dimensional settings in which full-rank maximum likelihood or a posteriori estimates are intractable. We evaluate our approach on neural data from rat retina and primate cortex and show that it substantially outperforms a variety of existing estimators. Our modeling approach will have useful extensions to a variety of other high-dimensional inference problems with smooth or low-rank structure.

A General Framework for Inferring Bayesian Ideal Observer Models from Psychophysical Data.

A central question in neuroscience is how sensory inputs are transformed into percepts. At this point, it is clear that this process is strongly influenced by prior knowledge of the sensory environment. Bayesian ideal observer models provide a useful link between data and theory that can help researchers evaluate how prior knowledge is represented and integrated with incoming sensory information. However, the statistical prior employed by a Bayesian observer cannot be measured directly, and must instead be inferred from behavioral measurements. Here, we review the general problem of inferring priors from psychophysical data, and the simple solution that follows from assuming a prior that is a Gaussian probability distribution. As our understanding of sensory processing advances, however, there is an increasing need for methods to flexibly recover the shape of Bayesian priors that are not well approximated by elementary functions. To address this issue, we describe a novel approach that applies to arbitrary prior shapes, which we parameterize using mixtures of Gaussian distributions. After incorporating a simple approximation, this method produces an analytical solution for psychophysical quantities that can be numerically optimized to recover the shapes of Bayesian priors. This approach offers advantages in flexibility, while still providing an analytical framework for many scenarios. We provide a MATLAB toolbox implementing key computations described herein.

Correcting motion induced fluorescence artifacts in two-channel neural imaging.

Imaging neural activity in a behaving animal presents unique challenges in part because motion from an animal's movement creates artifacts in fluorescence intensity time-series that are difficult to distinguish from neural signals of interest. One approach to mitigating these artifacts is to image two channels simultaneously: one that captures an activity-dependent fluorophore, such as GCaMP, and another that captures an activity-independent fluorophore such as RFP. Because the activity-independent channel contains the same motion artifacts as the activity-dependent channel, but no neural signals, the two together can be used to identify and remove the artifacts. However, existing approaches for this correction, such as taking the ratio of the two channels, do not account for channel-independent noise in the measured fluorescence. Here, we present Two-channel Motion Artifact Correction (TMAC), a method which seeks to remove artifacts by specifying a generative model of the two channel fluorescence that incorporates motion artifact, neural activity, and noise. We use Bayesian inference to infer latent neural activity under this model, thus reducing the motion artifact present in the measured fluorescence traces. We further present a novel method for evaluating ground-truth performance of motion correction algorithms by comparing the decodability of behavior from two types of neural recordings; a recording that had both an activity-dependent fluorophore and an activity-independent fluorophore (GCaMP and RFP) and a recording where both fluorophores were activity-independent (GFP and RFP). A successful motion correction method should decode behavior from the first type of recording, but not the second. We use this metric to systematically compare five models for removing motion artifacts from fluorescent time traces. We decode locomotion from a GCaMP expressing animal 20x more accurately on average than from control when using TMAC inferred activity and outperforms all other methods of motion correction tested, the best of which were ~8x more accurate than control.

Probing the Relationship Between Latent Linear Dynamical Systems and Low-Rank Recurrent Neural Network Models.

A large body of work has suggested that neural populations exhibit low-dimensional dynamics during behavior. However, there are a variety of different approaches for modeling low-dimensional neural population activity. One approach involves latent linear dynamical system (LDS) models, in which population activity is described by a projection of low-dimensional latent variables with linear dynamics. A second approach involves low-rank recurrent neural networks (RNNs), in which population activity arises directly from a low-dimensional projection of past activity. Although these two modeling approaches have strong similarities, they arise in different contexts and tend to have different domains of application. Here we examine the precise relationship between latent LDS models and linear low-rank RNNs. When can one model class be converted to the other, and vice versa? We show that latent LDS models can only be converted to RNNs in specific limit cases, due to the non-Markovian property of latent LDS models. Conversely, we show that linear RNNs can be mapped onto LDS models, with latent dimensionality at most twice the rank of the RNN. A surprising consequence of our results is that a partially observed RNN is better represented by an LDS model than by an RNN consisting of only observed units.

Reduced neural activity but improved coding in rodent higher-order visual cortex during locomotion.

Running profoundly alters stimulus-response properties in mouse primary visual cortex (V1), but its effect in higher-order visual cortex is under-explored. Here we systematically investigate how visual responses vary with locomotive state across six visual areas and three cortical layers using a massive dataset from the Allen Brain Institute. Although previous work has shown running speed to be positively correlated with neural activity in V1, here we show that the sign of correlations between speed and neural activity varies across extra-striate cortex, and is even negative in anterior extra-striate cortex. Nevertheless, across all visual cortices, neural responses can be decoded more accurately during running than during stationary periods. We show that this effect is not attributable to changes in population activity structure, and propose that it instead arises from an increase in reliability of single-neuron responses during locomotion.

Detecting and correcting false transients in calcium imaging.

Population recordings of calcium activity are a major source of insight into neural function. Large datasets require automated processing, but this can introduce errors that are difficult to detect. Here we show that popular time course-estimation algorithms often contain substantial misattribution errors affecting 10-20% of transients. Misattribution, in which fluorescence is ascribed to the wrong cell, arises when overlapping cells and processes are imperfectly defined or not identified. To diagnose misattribution, we develop metrics and visualization tools for evaluating large datasets. To correct time courses, we introduce a robust estimator that explicitly accounts for contaminating signals. In one hippocampal dataset, removing contamination reduced the number of place cells by 15%, and 19% of place fields shifted by over 10 cm. Our methods are compatible with other cell-finding techniques, empowering users to diagnose and correct a potentially widespread problem that could alter scientific conclusions.

Opponent control of behavior by dorsomedial striatal pathways depends on task demands and internal state.

A classic view of the striatum holds that activity in direct and indirect pathways oppositely modulates motor output. Whether this involves direct control of movement, or reflects a cognitive process underlying movement, remains unresolved. Here we find that strong, opponent control of behavior by the two pathways of the dorsomedial striatum depends on the cognitive requirements of a task. Furthermore, a latent state model (a hidden Markov model with generalized linear model observations) reveals that-even within a single task-the contribution of the two pathways to behavior is state dependent. Specifically, the two pathways have large contributions in one of two states associated with a strategy of evidence accumulation, compared to a state associated with a strategy of repeating previous choices. Thus, both the demands imposed by a task, as well as the internal state of mice when performing a task, determine whether dorsomedial striatum pathways provide strong and opponent control of behavior.

Mice alternate between discrete strategies during perceptual decision-making.

Classical models of perceptual decision-making assume that subjects use a single, consistent strategy to form decisions, or that decision-making strategies evolve slowly over time. Here we present new analyses suggesting that this common view is incorrect. We analyzed data from mouse and human decision-making experiments and found that choice behavior relies on an interplay among multiple interleaved strategies. These strategies, characterized by states in a hidden Markov model, persist for tens to hundreds of trials before switching, and often switch multiple times within a session. The identified decision-making strategies were highly consistent across mice and comprised a single 'engaged' state, in which decisions relied heavily on the sensory stimulus, and several biased states in which errors frequently occurred. These results provide a powerful alternate explanation for 'lapses' often observed in rodent behavioral experiments, and suggest that standard measures of performance mask the presence of major changes in strategy across trials.

Brain kernel: A new spatial covariance function for fMRI data.

A key problem in functional magnetic resonance imaging (fMRI) is to estimate spatial activity patterns from noisy high-dimensional signals. Spatial smoothing provides one approach to regularizing such estimates. However, standard smoothing methods ignore the fact that correlations in neural activity may fall off at different rates in different brain areas, or exhibit discontinuities across anatomical or functional boundaries. Moreover, such methods do not exploit the fact that widely separated brain regions may exhibit strong correlations due to bilateral symmetry or the network organization of brain regions. To capture this non-stationary spatial correlation structure, we introduce the brain kernel, a continuous covariance function for whole-brain activity patterns. We define the brain kernel in terms of a continuous nonlinear mapping from 3D brain coordinates to a latent embedding space, parametrized with a Gaussian process (GP). The brain kernel specifies the prior covariance between voxels as a function of the distance between their locations in embedding space. The GP mapping warps the brain nonlinearly so that highly correlated voxels are close together in latent space, and uncorrelated voxels are far apart. We estimate the brain kernel using resting-state fMRI data, and we develop an exact, scalable inference method based on block coordinate descent to overcome the challenges of high dimensionality (10-100K voxels). Finally, we illustrate the brain kernel's usefulness with applications to brain decoding and factor analysis with multiple task-based fMRI datasets.

Neural anatomy and optical microscopy (NAOMi) simulation for evaluating calcium imaging methods.

BACKGROUND: The past decade has seen a multitude of new in vivo functional imaging methodologies. However, the lack of ground-truth comparisons or evaluation metrics makes the large-scale, systematic validation vital to the continued development and use of optical microscopy impossible. NEW-METHOD: We provide a new framework for evaluating two-photon microscopy methods via in silico Neural Anatomy and Optical Microscopy (NAOMi) simulation. Our computationally efficient model generates large anatomical volumes of mouse cortex, simulates neural activity, and incorporates optical propagation and scanning to create realistic calcium imaging datasets. RESULTS: We verify NAOMi simulations against in vivo two-photon recordings from mouse cortex. We leverage this in silico ground truth to directly compare different segmentation algorithms and optical designs. We find modern segmentation algorithms extract strong neural time-courses comparable to estimation using oracle spatial information, but with an increase in the false positive rate. Comparison between optical setups demonstrate improved resilience to motion artifacts in sparsely labeled samples using Bessel beams, increased signal-to-noise ratio and cell-count using low numerical aperture Gaussian beams and nuclear GCaMP, and more uniform spatial sampling with temporal focusing versus multi-plane imaging. COMPARISON WITH EXISTING METHODS: NAOMi is a first-of-its kind framework for assessing optical imaging modalities. Existing methods are either anatomical simulations or do not address functional imaging. Thus there is no competing method for simulating realistic functional optical microscopy data. CONCLUSIONS: By leveraging the rich accumulated knowledge of neural anatomy and optical physics, we provide a powerful new tool to assess and develop important methods in neural imaging.

Extracting the dynamics of behavior in sensory decision-making experiments.

Decision-making strategies evolve during training and can continue to vary even in well-trained animals. However, studies of sensory decision-making tend to characterize behavior in terms of a fixed psychometric function that is fit only after training is complete. Here, we present PsyTrack, a flexible method for inferring the trajectory of sensory decision-making strategies from choice data. We apply PsyTrack to training data from mice, rats, and human subjects learning to perform auditory and visual decision-making tasks. We show that it successfully captures trial-to-trial fluctuations in the weighting of sensory stimuli, bias, and task-irrelevant covariates such as choice and stimulus history. This analysis reveals dramatic differences in learning across mice and rapid adaptation to changes in task statistics. PsyTrack scales easily to large datasets and offers a powerful tool for quantifying time-varying behavior in a wide variety of animals and tasks.