Track Session Program - All times EST

DAY 1
Monday, December 14

12:30 PM - 2:00 PM: Track Session 2

Talk Title: Building Mental Models of our Networked World

Abstract: Human learners acquire not only disconnected bits of information, but complex interconnected networks of relational knowledge. The capacity for such learning naturally depends upon three factors: (i) the architecture of the knowledge network itself, (ii) the nature of our perceptive instrument, and (iii) the instantiation of that instrument in biological tissue. In this talk, I will walk through each factor in turn. l will begin by describing recent work assessing network constraints on the learnability of relational knowledge. I will then describe a computational model informed by the free energy principle, which offers an explanation of how such network constraints manifest in human perception. In the third section of the talk, I will describe how neural representations reflect network constraints. Throughout, I'll move from previously published work to unpublished data, and from the world outside to the world inside, before speculating on as-yet uncharted territory

Talk Title: Latent Neural Source Recovery via Transcoding of Simultaneous EEG-fMRI

Abstract: Simultaneous EEG-fMRI is a multi-modal neuroimaging technique that provides complementary spatial and temporal resolution for inferring a latent source space of neural activity. We address this inference problem within the framework of transcoding – mapping from a specific encoding (modality) to a decoding (the latent source space) and then encoding the latent source space to the other modality. Specifically, we develop a symmetric method consisting of a cyclic convolution transcoder that transcodes EEG to fMRI and vice versa. Without any prior knowledge of either the hemodynamic response function or lead field matrix,the method exploits the temporal and spatial relationships between the modalities and latent source spaces to learn these mappings. We show, for real EEG-fMRIdata, how well the modalities can be transcoded from one to another as well as the source spaces that are recovered, all on unseen data. In addition to enabling a new way to symmetrically infer a latent source space, the method can also be seen as low-cost computational neuroimaging – i.e. generating an ’expensive’ fMRIBOLD image from ’low cost’ EEG data.

Talk Title: The “AI Neuroscientist”: A New Imaging Framework Combining Real-Time fMRI and Machine Learning with Applications in Health and Disease

Abstract: Cognitive neuroscientists are often interested in broad research questions, yet use overly narrow experimental designs by considering only a small subset of possible experimental conditions. This limits the generalizability and reproducibility of many research findings. In this talk, I present an alternative approach that resolves these problems by combining real-time fMRI with a branch of machine learning, Bayesian optimization. Neuroadaptive Bayesian optimization is an active sampling approach that allows to intelligently search through large experiment spaces with the aim to optimize an unknown objective function. It thus provides a powerful strategy to efficiently explore many more experimental conditions than is currently possible with standard neuroimaging methodology. Here, I will present results from two different studies where we applied the method to: (1) better understand the functional role of frontoparietal networks and (2) map cognitive dysfunction in aphasic stroke patients. I will conclude my talk in discussing how Bayesian optimization can be combined with study preregistration to cover exploration, mitigating researcher bias more broadly and improving reproducibility.

4:00 PM - 5:30 PM: Track Session 4

Talk Title: Bringing AI and Neuroscience Together

Abstract: Despite its remarkable recent advances, AI is still far from achieving human-level intelligence, and making further progress in that direction may require developing fundamentally new approaches to move us from today’s mostly “narrow”/task-specific AI towards a “broad”/multi-functional, continually learning and robust AI. One promising avenue of research which is believed to have a potential for revolutionizing the field is to explore more biologically - inspired mechanisms behind the human intelligence. On the other hand, introducing AI ideas, models and techniques to neuroscience, psychology and mental health can greatly benefit those fields as well. In this talk, I plan to provide an overview of several ongoing efforts on the intersection between those fields, including application of machine-learning approaches to computational psychiatry and neuroimaging, such as mental state prediction from fMRI and EEG data, dialogue generation for depression therapy, as well as using some brain-inspired models (e.g., adult neurogenesis) to advance machine-learning algorithms.

Talk Title: Deep Graph Pose: A Semi-Supervised Deep Graphical Model for Improved Animal Pose Tracking

Noninvasive behavioral tracking of animals is crucial for many scientific investigations. Recent transfer learning approaches for behavioral tracking have considerably advanced the state of the art. Typically these methods only consider labeled frames in videos and treat each video frame and each object to be tracked independently. In this work, we improve on these methods (particularly in the regime of few training labels) by leveraging the rich spatiotemporal structures pervasive in behavioral video — specifically, the spatial statistics imposed by physical constraints (e.g., paw to elbow distance), and the temporal statistics imposed by smoothness from frame to frame. We propose a probabilistic graphical model built on top of deep neural networks, Deep Graph Pose (DGP), to leverage these useful spatial and temporal constraints, and develop an efficient structured variational approach to perform inference in this model. The resulting semi-supervised model exploits both labeled and unlabeled frames to achieve significantly more accurate and robust tracking while requiring users to label fewer training frames. More importantly, these tracking improvements enhance performance on downstream applications, including robust unsupervised segmentation of behavioral “syllables,” and estimation of interpretable “disentangled” low-dimensional representations of the full behavioral video.

Talk Title: Factor-Analytic Inverse Regression for High-Dimension, Small-Sample Dimensionality Reduction

Abstract: High-dimensional data is an inevitable challenge of modern problems of interest in neuroscience and machine learning. When mapping high-dimensional observations to a target variable, often many of the observed dimensions are uninformative or redundant. This motivates sufficient dimension reduction methods, which identify a reduced-dimension set of variables that preserve the input-output relationship. However, existing sufficient dimension reduction methods typically require more observations than the data-dimensionality ($N > p$), and their performance is sensitive to the scale of $N$. High-throughput neural recordings often fall into the opposite regime where $N < p$, which make sufficient dimension reduction methods intractable, even though mounting evidence suggests that neural activity varies within a small set of latent dimensions.

DAY 2
Tuesday, December 15

11:00 AM - 12:30 PM: Track Session 5

Talk Title: Primary Motor Cortex Does Drive Premotor Cortex During Movements

Abstract: The factor-analytic backbone of the model also readily incorporates priors for known structure of the data, such as smoothness in the projection axes, which we show can dramatically improve performance in the small sample regime. We demonstrate the effectiveness of CFAD with an application to functional magnetic resonance imaging (fMRI) measurements of brain activity during visual object recognition.

Talk Title: Inferring Learning Rules From Animal Decision-Making

Abstract: How do animals learn? This remains an elusive question in neuroscience. Whereas reinforcement learning often focuses on the design of algorithms that enable artificial agents to efficiently learn new tasks, we have developed a modeling framework to directly infer the empirical learning rules that animals use to acquire new behaviors. Our method efficiently infers the trial-to-trial changes in an animal’s policy, and decomposes those changes into a learning component and a noise component. Specifically, this allows us to: (i) compare different learning rules and objective functions that an animal may be using to update its policy; (ii) estimate distinct learning rates for different parameters of an animal’s policy; (iii) identify variations in learning across cohorts of animals; and (iv) uncover trial-to-trial changes that are not captured by normative learning rules. After validating our framework on simulated choice data, we applied our model to data from rats and mice learning perceptual decision-making tasks. We considered variants of the policy gradient learning rule known as REINFORCE [Williams, 1992], and found that certain rules were far more capable of explaining the trial-to-trial policy changes used by real animals. Whereas the average contribution of the conventional REINFORCE learning rule to the policy update for mice learning the International Brain Laboratory's task was just 30%, we found that adding reward-offset ('baseline') parameters allowed the learning rule to explain 92% of the animals' policy updates under our model. Intriguingly, the best-fitting learning rates and baseline values indicate that an animal's policy update, at each trial, does not occur in the direction that maximizes expected reward. Understanding how an animal transitions from chance-level to high-accuracy performance when learning a new task not only provides neuroscientists with insight into their animals, but also provides concrete examples of biological learning algorithms to the machine learning community.

Talk Title: Controversial Stimuli: Pitting Neural Networks Against Each Other as Models of Human Recognition

Abstract: Deep neural networks (DNNs) provide the leading model of biological object recognition, but their power and flexibility come at a price: different DNN models often make very similar predictions. To enable iterative testing and improvement of DNNs as scientific hypotheses about biological vision, we need to efficiently adjudicate between different candidate models. We suggest synthetizing controversial stimuli to achieve this aim. Controversial stimuli are inputs (e.g., images) whose classifications by two (or more) models are incompatible. Since human perceptual judgments of a controversial stimulus cannot be compatible with both models, such judgments are guaranteed to provide evidence against at least one of the models. Therefore, controversial stimuli allow to efficiently contrast the validity of different models. To demonstrate our approach, we have assembled a diverse set of deep neural networks, trained on either MNIST and CIFAR-10. While multiple networks achieve human-like accuracy on these two standard benchmarks, the various architectures cannot all be good models of how humans recognize MNIST digits or CIFAR-10 categories. For each pair of either MNIST or CIFAR-10 models, we sampled random noise images and optimized them to increase the incompatibility of their classifications by the two models. This resulted in a stimulus set for each task that induces disagreement among the models. We tested these stimuli on 90 human observers, who rated the presence of each of the categories from 0% to 100% in each image. Contrasting each model's outputs with the judgments of each observer revealed that the deep generative models outperform standard and adversarially trained discriminative deep neural networks at predicting the human responses. However, no model entirely explained the explainable variability of human responses for neither CIFAR-10 nor MNIST. I will present these results as well as more recent methods for synthesizing high-resolution controversial stimuli for ImageNet models and discuss controversial stimuli as an efficient experimental paradigm for comparing DNN models of vision and as a practical and conceptual generalization of adversarial examples.

2:30 PM - 4:00 PM: Track Session 7

Talk Title: Dynamical Modeling, Decoding, and Control of Multiscale Brain Networks: From Motor to Mood

Abstract: In this talk, I first discuss our recent work on modeling, decoding, and controlling multisite human brain dynamics underlying mood states. I present a multiscale dynamical modeling framework that allows us to decode mood variations and identify brain sites that are most predictive of mood. I then develop a system identification approach that can predict multiregional brain network dynamics (output) in response to electrical stimulation (input) toward enabling closed-loop control of brain network activity. Further, I demonstrate a novel modeling framework that can dissociate and uncover behaviorally relevant neural dynamics, such as those during naturalistic movements. Finally, I show how our framework can combine information from multiple spatiotemporal scales of activity and model their different time-scales and statistics. These dynamical models, decoders, and controllers can advance our understanding of neural mechanisms and facilitate future closed-loop therapies for neurological and neuropsychiatric disorders.

Talk Title: Bayesian Recurrent State Space Model for rs-fMRI

Abstract: Resting state fMRI (rs-fMRI) measures the intrinsic spontaneous activity across brain regions, in the absence of any sensory or cognitive stimulus, and has proven to be a useful tool in understanding the functional architecture of the brain. Alterations in these functional network patterns have been observed in neuropathologies such as Mild Cognitive Impairment (MCI), Alzheimer's, Depression, Schizophrenia, Autism and ADHD. We propose a recurrent state space model to uncover the spatio-temporal correlation patterns observed in rs-fMRI data, which allows us to model the state switching behavior across different diseases classes, and to delineate the network patterns shared across all classes. We evaluate our method on rs-fMRI data from patients with MCI. In addition to states shared across healthy and individuals MCI, we discover latent states that are predominantly observed in individuals with MCI. Our approach outperforms current state of the art deep learning method on ADNI2 dataset.

Talk Title: Data-Driven Automatic Neuromotor Disorder Detection

Abstract: Artificial intelligence is leading advances in automating various aspects of science and medicine. With recent advances in automatically estimating whole-body human pose from videos, neuromotor rehabilitation is the next frontier for artificial intelligence in medicine. In this talk, I will outline our research on using neural network models for automated neuromotor disorder detection in infants. Infants have complex aperiodic movements that make it difficult, expensive, and time-intensive to quantify what ‘typically developing’ movements look like using traditional methods. To solve this, we curated a large dataset of typical infant movements using computer vision on videos, against which to assess neuromotor development. Using unsupervised learning methods, we estimated the probability that a given infant’s movements are typical as a measure of neuromotor disorder risk. In addition to body movements, we also created a system to automatically classify infant emotion from facial cues, which are an important behavioral signal of development. This system lays the foundations for accessible and quantified automated infant neuro-developmental assessment.