Deep learning models deployed in real-world applications (e.g., medicine) face challenges because source models do not generalize well to domain-shifted target data. Many successful domain adaptation (DA) approaches require full access to source data. Yet, such requirements are unrealistic in scenarios where source data cannot be shared either because of privacy concerns or because it is too large and incurs prohibitive storage or computational costs. Moreover, resource constraints may limit the availability of labeled targets. We illustrate this challenge in a neuroscience setting where source data are unavailable, labeled target data are meager, and predictions involve continuous-valued outputs. We build upon Contradistinguisher (CUDA), an efficient framework that learns a shared model across the labeled source and unlabeled target samples, without intermediate representation alignment. Yet, CUDA was designed for unsupervised DA, with full access to source data, and for classification tasks. We develop CRAFT -- a Contradistinguisher-based Regularization Approach for Flexible Training -- for source-free (SF), semi-supervised transfer of pretrained models in regression tasks. We showcase the efficacy of CRAFT in two neuroscience settings: gaze prediction with electroencephalography (EEG) data and ``brain age'' prediction with structural MRI data. For both datasets, CRAFT yielded up to 9% improvement in root-mean-squared error (RMSE) over fine-tuned models when labeled training examples were scarce. Moreover, CRAFT leveraged unlabeled target data and outperformed four competing state-of-the-art source-free domain adaptation models by more than 3%. Lastly, we demonstrate the efficacy of CRAFT on two other real-world regression benchmarks. We propose CRAFT as an efficient approach for source-free, semi-supervised deep transfer for regression that is ubiquitous in biology and medicine.

Deep neural networks are increasingly applied in automated histopathology. Yet, whole-slide images (WSIs) are often acquired at gigapixel sizes, rendering them computationally infeasible to analyze entirely at high resolution. Diagnostic labels are largely available only at the slide-level, because expert annotation of images at a finer (patch) level is both laborious and expensive. Moreover, regions with diagnostic information typically occupy only a small fraction of the WSI, making it inefficient to examine the entire slide at full resolution. Here, we propose SASHA -- Sequential Attention-based Sampling for Histopathological Analysis -- a deep reinforcement learning approach for efficient analysis of histopathological images. First, SASHA learns informative features with a lightweight hierarchical, attention-based multiple instance learning (MIL) model. Second, SASHA samples intelligently and zooms selectively into a small fraction (10-20\%) of high-resolution patches to achieve reliable diagnoses. We show that SASHA matches state-of-the-art methods that analyze the WSI fully at high resolution, albeit at a fraction of their computational and memory costs. In addition, it significantly outperforms competing, sparse sampling methods. We propose SASHA as an intelligent sampling model for medical imaging challenges that involve automated diagnosis with exceptionally large images containing sparsely informative features. Model implementation is available at: https://github.com/coglabiisc/SASHA.

Reliable outlier detection is critical for real-world deployment of deep learning models. Although extensively studied, likelihoods produced by deep generative models have been largely dismissed as being impractical for outlier detection. First, deep generative model likelihoods are readily biased by low-level input statistics. Second, many recent solutions for correcting these biases are computationally expensive, or do not generalize well to complex, natural datasets. Here, we explore outlier detection with a state-of-the-art deep autoregressive model: PixelCNN++. We show that biases in PixelCNN++ likelihoods arise primarily from predictions based on local dependencies. We propose two families of bijective transformations -- ``stirring'' and ``shaking'' -- which ameliorate low-level biases and isolate the contribution of long-range dependencies to PixelCNN++ likelihoods. These transformations are inexpensive and readily computed at evaluation time. We test our approaches extensively with five grayscale and six natural image datasets and show that they achieve or exceed state-of-the-art outlier detection, particularly on datasets with complex, natural images. We also show that our solutions work well with other types of generative models (generative flows and variational autoencoders) and that their efficacy is governed by each model's reliance on local dependencies. In sum, lightweight remedies suffice to achieve robust outlier detection on image data with deep generative models.

Deep networks often make confident, yet, incorrect, predictions when tested with outlier data that is far removed from their training distributions. Likelihoods computed by deep generative models (DGMs) are a candidate metric for outlier detection with unlabeled data. Yet, previous studies have shown that DGM likelihoods are unreliable and can be easily biased by simple transformations to input data. Here, we examine outlier detection with variational autoencoders (VAEs), among the simplest of DGMs. We propose novel analytical and algorithmic approaches to ameliorate key biases with VAE likelihoods. Our bias corrections are sample-specific, computationally inexpensive, and readily computed for various decoder visible distributions. Next, we show that a well-known image pre-processing technique – contrast stretching – extends the effectiveness of bias correction to further improve outlier detection. Our approach achieves state-of-the-art accuracies with nine grayscale and natural image datasets, and demonstrates significant advantages – both with speed and performance – over four recent, competing approaches. In summary, lightweight remedies suffice to achieve robust outlier detection with VAEs.11Code is available at https://github.com/google-research/google-research/tree/master/vae_ood.

Despite possessing the capacity for selective attention, we often fail to notice the obvious. We investigated participants' (n = 39) failures to detect salient changes in a change blindness experiment. Surprisingly, change detection success varied by over two-fold across participants. These variations could not be readily explained by differences in scan paths or fixated visual features. Yet, two simple gaze metrics-mean duration of fixations and the variance of saccade amplitudes-systematically predicted change detection success. We explored the mechanistic underpinnings of these results with a neurally-constrained model based on the Bayesian framework of sequential probability ratio testing, with a posterior odds-ratio rule for shifting gaze. The model's gaze strategies and success rates closely mimicked human data. Moreover, the model outperformed a state-of-the-art deep neural network (DeepGaze II) with predicting human gaze patterns in this change blindness task. Our mechanistic model reveals putative rational observer search strategies for change detection during change blindness, with critical real-world implications.

Functional magnetic resonance imaging (fMRI) enables measuring human brain activity, in vivo. Yet, the fMRI hemodynamic response unfolds over very slow timescales (<0.1-1 Hz), orders of magnitude slower than millisecond timescales of neural spiking. It is unclear, therefore, if slow dynamics as measured with fMRI are relevant for cognitive function. We investigated this question with a novel application of Gaussian Process Factor Analysis (GPFA) and machine learning to fMRI data. We analyzed slowly sampled (1.4 Hz) fMRI data from 1000 healthy human participants (Human Connectome Project database), and applied GPFA to reduce dimensionality and extract smooth latent dynamics. GPFA dimensions with slow (<1 Hz) characteristic timescales identified, with high accuracy (>95%), the specific task that each subject was performing inside the fMRI scanner. Moreover, functional connectivity between slow GPFA latents accurately predicted inter-individual differences in behavioral scores across a range of cognitive tasks. Finally, infra-slow (<0.1 Hz) latent dynamics predicted CDR (Clinical Dementia Rating) scores of individual patients, and identified patients with mild cognitive impairment (MCI) who would progress to develop Alzheimer’s dementia (AD). Slow and infra-slow brain dynamics may be relevant for understanding the neural basis of cognitive function, in health and disease.

Brain processes occur at various timescales, ranging from milliseconds (neurons) to minutes and hours (behavior). Characterizing functional coupling among brain regions at these diverse timescales is key to understanding how the brain produces behavior. Here, we apply instantaneous and lag-based measures of conditional linear dependence, based on Granger-Geweke causality (GC), to infer network connections at distinct timescales from functional magnetic resonance imaging (fMRI) data. Due to the slow sampling rate of fMRI, it is widely held that GC produces spurious and unreliable estimates of functional connectivity when applied to fMRI data. We challenge this claim with simulations and a novel machine learning approach. First, we show, with simulated fMRI data, that instantaneous and lag-based GC identify distinct timescales and complementary patterns of functional connectivity. Next, we analyze fMRI scans from 500 subjects and show that a linear classifier trained on either instantaneous or lag-based GC connectivity reliably distinguishes task versus rest brain states, with ~80-85% cross-validation accuracy. Importantly, instantaneous and lag-based GC exploit markedly different spatial and temporal patterns of connectivity to achieve robust classification. Our approach enables identifying functionally connected networks that operate at distinct timescales in the brain.

We describe a neurobiologically plausible model to implement dynamic routing using the concept of neuronal communication through neuronal coherence. The model has a three-tier architecture: a raw input tier, a routing control tier, and an invariant output tier. The correct mapping between input and output tiers is re- alized by an appropriate alignment of the phases of their respective background oscillations by the routing control units. We present an example architecture, im- plemented on a neuromorphic chip, that is able to achieve circular-shift invariance. A simple extension to our model can accomplish circular-shift dynamic routing with only O(N) connections, compared to O(N 2) connections required by tradi- tional models. 1 Dynamic Routing Circuit Models for Circular-Shift Invariance Dynamic routing circuit models are among the most prominent neural models for invariant recogni- tion [1] (also see [2] for review). These models implement shift invariance by dynamically changing spatial connectivity to transform an object to a standard position or orientation. The connectivity between the raw input and invariant output layers is controlled by routing units, which turn certain subsets of connections on or off (Figure 1A). An important feature of this model is the explicit rep- resentation of what and where information in the main network and the routing units, respectively; the routing units use the where information to create invariant representations. Traditional solutions for shift invariance are neurobiologically implausible for at least two reasons. First, there are too many synaptic connections: for N input neurons, N output neurons and N possible input-output mappings, the network requires O(N 2) connections in the routing layer— between each of the N routing units and each set of N connections that that routing unit gates (Figure 1A). Second, these connections must be extremely precise: each routing unit must activate an input- output mapping (N individual connections) corresponding to the desired shift (as highlighted in Figure 1A). Other approaches that have been proposed, including invariant feature networks [3,4], also suffer from significant drawbacks, such as the inability to explicitly represent where information [2]. It remains an open question how biology could achieve shift invariance without profligate and precise connections. In this article, we propose a simple solution for shift invariance for quantities that are circular or periodic in nature—circular-shift invariance (CSI)—orientation invariance in vision and key invari- ance in music. The visual system may create orientation-invariant representations to aid recognition under conditions of object rotation or head-tilt [5,6]; a similar mechanism could be employed by the auditory system to create key-invariant representations under conditions where the same melody 1 Figure 1: Dynamic routing. A In traditional dynamic routing, connections from the (raw) input layer to the (invariant) output layer are gated by routing units. For instance, the mapping from A to 5, B to 6, . . . , F to 4 is achieved by turning on the highlighted routing unit. B In time-division multiplexing (TDM), the encoder samples input channels periodically (using a rotating switch) while the decoder sends each sample to the appropriate output channel (based on its time bin). TDM can be extended to achieve a circular-shift transformation by altering the angle between encoder and decoder switches (θ), thereby creating a rotated mapping between input and output channels (adapted from [7]). is played in different keys. Similar to orientation, which is a periodic quantity, musical notes one octave apart sound alike, a phenomenon known as octave equivalence [8]. Thus, the problems of key invariance and orientation invariance admit similar solutions. Deriving inspiration from time-division multiplexing (TDM), we propose a neural network for CSI that uses phase to encode and decode information. We modulate the temporal window of commu- nication between (raw) input and (invariant) output neurons to achieve the appropriate input–output mapping. Extending TDM, any particular circular-shift transformation can be accomplished by changing the relative angle, θ, between the rotating switches of the encoder (that encodes the raw input in time) and decoder (that decodes the invariant output in time) (Figure 1B). This obviates the need to hardwire routing control units that specifically modulate the strength of each possible input- output connection, thereby significantly reducing the complexity inherent in the traditional dynamic routing solution. Similarly, a remapping between the input and output neurons can be achieved by introducing a relative phase-shift in their background oscillations. 2 Dynamic Routing through Neuronal Coherence To modulate the temporal window of communication, the model uses a ring of neurons (the oscilla- tion ring) to select the pool of neurons (in the projection ring) that encode or decode information at a particular time (Figure 2A). Each projection pool encodes a specific value of the feature (for exam- ple, one of twelve musical notes). Upon activation by external input, each pool is active only when background inhibition generated by the oscillation ring (outer ring of neurons) is at a minimum. In addition to exciting 12 inhibitory interneurons in the projection ring, each oscillation ring neuron excites its nearest 18 neighbors in the clockwise direction around the oscillation ring. As a result, a wave of inhibition travels around the projection ring that allows only one pool to be excitable at any point in time. These neurons become excitable at roughly the same time (numbered sectors, inner ring) by virtue of recurrent excitatory intra-pool connections. Decoding is accomplished by a second tier of rings (Figure 2B). The projection ring of the first (in- put) tier connects all-to-all to the projection ring of the second (output) tier. The two oscillation rings create a window of excitability for the pools of neurons in their respective projection rings. Hence, the most effective communication occurs between input and output pools that become excitable at the same time (i.e. are oscillating in phase with one another [9]). The CSI problem is solved by introducing a phase-shift between the input and output tiers. If they are exactly in phase, then an input pool is simply mapped to the output pool directly above it. If their 2 Figure 2: Double-Ring Network for Encoding and Decoding. A The projection (inner) ring is divided into (numbered) pools. The oscillation (outer) ring modulates sub-threshold activity (wave- forms) of the projection ring by exciting (black distribution) inhibitory neurons that inhibit neigh- boring projection neurons. A wave of activity travels around the oscillation ring due to asymmetric excitatory connections, creating a corresponding wave of inhibitory activity in the projection ring, such that only one pool of projection neurons is excitable (spikes) at a given time. B Two instances of the double-ring structure from A. The input projection ring connects all-to-all to the output pro- jection ring (dashed lines). Because each input pool will spike only during a distinct time bin, and each output pool is excitable only in a certain time bin, communication occurs between input and output pools that are oscillating in phase with each other. Appropriate phase offset between input and output oscillation rings realizes the desired circular shift (input pool H to output pool 1, solid arrow). C Interactions among pools highlighted in B. phases are different, the input is dynamically routed to an appropriate circularly shifted position in the output tier. Such changes in phase are analogous to adjusting the angle of the rotating switch at either the encoder or the decoder in TDM (see Figure 1B). There is some evidence that neural systems could employ phase relationships of subthreshold oscillations to selectively target neural populations [9-11]. 3 Implementation in Silicon We implemented this solution to CSI on a neuromorphic silicon chip [12]. The neuromorphic chip has neurons whose properties resemble that of biological neurons; these neurons even have intrin- sic differences, thereby mimicking heterogeneity in real neurobiological systems. The chip uses a conductance-based spiking model for both inhibitory and excitatory neurons. Inhibitory neurons project to nearby excitatory and inhibitory neurons via a diffusor network that determines the spread of inhibition. A lookup table of excitatory synaptic connectivity is stored in a separate random- access memory (RAM) chip. Spikes occurring on-chip are converted to a neuron address, mapped to synapses (if any) via the lookup table, and routed to the targeted on-chip synapse. A universal serial bus (USB) interface chip communicates spikes to and from a computer, for external input and 3 Figure 3: Traveling-wave activity in the oscillation ring. A Population activity (5ms bins) of a pool of eighteen (adjacent) oscillation neurons. B Increasing the strength of feedforward excitation led to increasing frequencies of periodic firing in the θ and α range (1-10 Hz). Strength of excitation is the amplitude change in post-synaptic conductance due to a single pre-synaptic spike (measured relative to minimum amplitude used). data analysis, respectively. Simulations on the chip occur in real-time, making it an attractive option for implementing the model. We configured the following parameters: • Magnitude of a potassium M-current: increasing this current’s magnitude increased the post-spike repolarization time of the membrane potential, thereby constraining spiking to a single time bin per cycle. • The strength of excitatory and inhibitory synapses: a correct balance had to be established between excitation and inhibition to make only a small subset of neurons in the projection rings fire at a time—too much excitation led to widespread firing and too much inhibition led to neurons that were entirely silent or fired sporadically. • The space constant of inhibitory spread: increasing the spread was effective in preventing runaway excitation, which could occur due to the recurrent excitatory connections. We were able to create a stable traveling wave of background activity within the oscillation ring. We transiently stimulated a small subset of the neurons, which initiated a wave of activity that propagated in a stable manner around the ring after the transient external stimulation had ceased (Figure 3A). The network frequency determined from a Fourier transform of the network activity smoothed with a non-causal Gaussian kernel (FDHM = 80ms) was 7.4Hz. The frequency varied with the strength of the neurons’ excitatory connections (Figure 3B), measured as the amplitude of the step increase in membrane conductivity due to the arrival of a pre-synaptic spike. Over much of the range of the synaptic strengths tested, we observed stable oscillations in the θ and α bands (1-10Hz); the frequency appeared to increase logarithmically with synaptic strength. 4 Phase-based Encoding and Decoding In order to assess the best-case performance of the model, the background activity in the input and output projection rings was derived from the input oscillation ring. Their spikes were delivered to the appropriately circularly-shifted output oscillation neurons. The asymmetric feedforward con- nections were disabled in the output oscillation ring. For instance, in order to achieve a circular shift by k pools (i.e. mapping input projection pool 1 to output projection pool k + 1, input pool 2 to output pool k + 2, and so on), activity from the input oscillation neurons closest to input pool 1 was fed into the output oscillation neurons closest to output pool k. By providing the appropriate phase difference between input and output oscillation, we were able to assess the performance of the model under ideal conditions. In the Discussion section, we discuss a biologically plausible mechanism to control the relative phases. 4 Figure 4: Phase-based encoding. Rasters indicating activity of projection pools in 1ms bins, and mean phase of firing (×’s) for each pool (relative to arbitrary zero time). The abscissa shows firing time normalized by the period of oscillation (which may be converted to firing phase by multiplica- tion by 2π). Under constant input to the input projection ring, the input pools fire approximately in sequence. Two cycles of pool activity normalized by maximum firing rate for each pool are shown in left inset (for clarity, pools 1-6 are shown in the top panel and pools 7-12 are shown separately in the bottom panel); phase of background inhibition of pool 4 is shown (below) for reference. Phase-aligned average1 of activity (right inset) showed that the firing times were relatively tight and uniform across pools: a standard deviation of 0.0945 periods, or equivalently, a spread of 1.135 pools at any instant of time. We verified that the input projection pools fired in a phase-shifted fashion relative to one another, a property critical for accurate encoding (see Figure 2). We stimulated all pools in the input pro- jection ring simultaneously while the input oscillation ring provided a periodic wave of background inhibition. The mean phase of firing for each pool (relative to arbitrary zero time) increased nearly linearly with pool number, thereby providing evidence for accurate, phase-based encoding (Figure 4). The firing times of all pools are shown for two cycles of background oscillatory activity (Figure 4 left inset). A phase-aligned average1 showed that the timing was relatively tight (standard deviation 1.135 pools) and uniform across pools of neurons (Figure 4 right inset). We then characterized the system’s ability to correctly decode this encoding under a given circular shift. The shift was set to seven pools, mapping input pool 1 to output pool 8, and so on. Each input pool was stimulated in turn. We expected to see only the appropriately shifted output pool become highly active. In fact, not only was this pool active, but other pools around it were also active, though to a lesser extent (Figure 5A). Thus, the phase-encoded input was decoded successfully, and circularly shifted, except that the output units were broadly tuned. To quantify the overall precision of encoding and decoding, we constructed an input-locked aver- age of the tuning curves (Figure 5B): the curves were circularly shifted to the left by an amount corresponding to the stimulated input pool number, and the raw pool firing rates were averaged. If the phase-based encoding and decoding were perfect, the peak should occur at a shift of 7 pools. 1The phase-aligned average was constructed by shifting the pool-activity curves by the (# of the pool) × 12 of the period) to align activity across pools, which was then averaged. ( 1 5 Figure 5: Decoding phase-encoded input. A In order to assess decoding performance under a given circular shift (here 7 pools) each input pool was stimulated in turn and activity in each output pool was recorded and averaged over 500ms. The pool’s response, normalized by its maximum firing rate, is plotted for each stimulated input pool (arrows pointing to curves, color code as in Figure 4). Each input pool stimulation trial consistently resulted in peak activity in the appropriate output pool; however, adjacent pools were also active, but to a lesser extent, resulting in a broad tuning curve. B The best-fit Gaussian (dot-dashed grey curve, σ = 2.30 pools) to the input-locked average of the raw pool firing rates (see text for details) revealed a maximum between a shift of 7 and 8 pools (inverted grey triangle; expected peak at a shift of 7 pools). Indeed, the highest (average) firing rate corresponded to a shift of 7 pools. However, the activity corresponding to a shift of 8 pools was nearly equal to that of 7 pools, and the best fitting Gaus- sian curve to the activity histogram (grey dot-dashed line) peaked at a point between pools 7 and 8 (inverted grey triangle). The standard deviation (σ) was 2.30 pools, versus the expected ideal σ of 1.60, which corresponds to the encoding distribution (σ = 1.135 pools) convolved with itself.

We present a computational model of basal ganglia as a key player in exploratory behavior. The model describes exploration of a virtual rat in a simulated water pool experiment. The virtual rat is trained using a reward-based or reinforcement learning paradigm which requires units with stochastic behavior for exploration of the system's state space. We model the Subthalamic Nucleus-Globus Pallidus externa (STN-GPe) segment of the basal ganglia as a pair of neuronal layers with oscillatory dynamics, exhibiting a variety of dynamic regimes such as chaos, traveling waves and clustering. Invoking the property of chaotic systems to explore state-space, we suggest that the complex exploratory dynamics of STN-GPe system in conjunction with dopamine-based reward signaling from the Substantia Nigra pars compacta (SNc) present the two key ingredients of a reinforcement learning system.