Deep Learning seminars

In popular media, there is often a connection drawn between the advent of awareness in artificial agents and those same agents simultaneously achieving human or superhuman level intelligence. In this talk, I will examine the validity and potential application of this seemingly intuitive link between consciousness and intelligence. I will do so by examining the cognitive abilities associated with three contemporary theories of conscious function: Global Workspace Theory (GWT), Information Generation Theory (IGT), and Attention Schema Theory (AST), and demonstrating that all three theories specifically relate conscious function to some aspect of domain-general intelligence in humans. With this insight, we will turn to the field of Artificial Intelligence (AI) and find that, while still far from demonstrating general intelligence, many state-of-the-art deep learning methods have begun to incorporate key aspects of each of the three functional theories. Given this apparent trend, I will use the motivating example of mental time travel in humans to propose ways in which insights from each of the three theories may be combined into a unified model. I believe that doing so can enable the development of artificial agents which are not only more generally intelligent but are also consistent with multiple current theories of conscious function.

With recent advances in learning algorithms, recurrent networks of spiking neurons are achieving performance competitive with standard recurrent neural networks. Still, these learning algorithms are limited to small networks of simple spiking neurons and modest-length temporal sequences, as they impose high memory requirements, have difficulty training complex neuron models, and are incompatible with online learning.Taking inspiration from the concept of Liquid Time-Constant (LTCs), we introduce a novel class of spiking neurons, the Liquid Time-Constant Spiking Neuron (LTC-SN), resulting in functionality similar to the gating operation in LSTMs. We integrate these neurons in SNNs that are trained with FPTT and demonstrate that thus trained LTC-SNNs outperform various SNNs trained with BPTT on long sequences while enabling online learning and drastically reducing memory complexity. We show this for several classical benchmarks that can easily be varied in sequence length, like the Add Task and the DVS-gesture benchmark. We also show how FPTT-trained LTC-SNNs can be applied to large convolutional SNNs, where we demonstrate novel state-of-the-art for online learning in SNNs on a number of standard benchmarks (S-MNIST, R-MNIST, DVS-GESTURE) and also show that large feedforward SNNs can be trained successfully in an online manner to near (Fashion-MNIST, DVS-CIFAR10) or exceeding (PS-MNIST, R-MNIST) state-of-the-art performance as obtained with offline BPTT. Finally, the training and memory efficiency of FPTT enables us to directly train SNNs in an end-to-end manner at network sizes and complexity that was previously infeasible: we demonstrate this by training in an end-to-end fashion the first deep and performant spiking neural network for object localization and recognition. Taken together, we out contribution enable for the first time training large-scale complex spiking neural network architectures online and on long temporal sequences.

Biologically plausible spiking neural networks (SNNs) are an emerging architecture for deep learning tasks due to their energy efficiency when implemented on neuromorphic hardware. However, many of the biological features are at best irrelevant and at worst counterproductive when evaluated in the context of task performance and suitability for neuromorphic hardware. In this talk, I will present an alternative paradigm to design deep learning architectures with good task performance in real-world benchmarks while maintaining all the advantages of SNNs. We do this by focusing on two main features – event-based computation and activity sparsity. Starting from the performant gated recurrent unit (GRU) deep learning architecture, we modify it to make it event-based and activity-sparse. The resulting event-based GRU (EGRU) is extremely efficient for both training and inference. At the same time, it achieves performance close to conventional deep learning architectures in challenging tasks such as language modelling, gesture recognition and sequential MNIST.

Memory is a key component of biological neural systems that enables the retention of information over a huge range of temporal scales, ranging from hundreds of milliseconds up to years. While Hebbian plasticity is believed to play a pivotal role in biological memory, it has so far been analyzed mostly in the context of pattern completion and unsupervised learning. Here, we propose that Hebbian plasticity is fundamental for computations in biological neural systems. We introduce a novel spiking neural network (SNN) architecture that is enriched by Hebbian synaptic plasticity. We experimentally show that our memory-equipped SNN model outperforms state-of-the-art deep learning mechanisms in a sequential pattern-memorization task, as well as demonstrate superior out-of-distribution generalization capabilities compared to these models. We further show that our model can be successfully applied to one-shot learning and classification of handwritten characters, improving over the state-of-the-art SNN model. We also demonstrate the capability of our model to learn associations for audio to image synthesis from spoken and handwritten digits. Our SNN model further presents a novel solution to a variety of cognitive question answering tasks from a standard benchmark, achieving comparable performance to both memory-augmented ANN and SNN-based state-of-the-art solutions to this problem. Finally we demonstrate that our model is able to learn from rewards on an episodic reinforcement learning task and attain near-optimal strategy on a memory-based card game. Hence, our results show that Hebbian enrichment renders spiking neural networks surprisingly versatile in terms of their computational as well as learning capabilities. Since local Hebbian plasticity can easily be implemented in neuromorphic hardware, this also suggests that powerful cognitive neuromorphic systems can be build based on this principle.

Research in Neuroscience, as in many scientific disciplines, is undergoing a renaissance based on deep learning. Unique to Neuroscience, deep learning models can be used not only as a tool but interpreted as models of the brain. The central claims of recent deep learning-based models of brain circuits are that they shed light on fundamental functions being optimized or make novel predictions about neural phenomena. We show, through the case-study of grid cells in the entorhinal-hippocampal circuit, that one may get neither. We rigorously examine the claims of deep learning models of grid cells using large-scale hyperparameter sweeps and theory-driven experimentation, and demonstrate that the results of such models are more strongly driven by particular, non-fundamental, and post-hoc implementation choices than fundamental truths about neural circuits or the loss function(s) they might optimize. We discuss why these models cannot be expected to produce accurate models of the brain without the addition of substantial amounts of inductive bias, an informal No Free Lunch result for Neuroscience.

Studies of the mouse visual system have revealed a variety of visual brain areas in a roughly hierarchical arrangement, together with a multitude of behavioral capacities, ranging from stimulus-reward associations, to goal-directed navigation, and object-centric discriminations. However, an overall understanding of the mouse’s visual cortex organization, and how this organization supports visual behaviors, remains unknown. Here, we take a computational approach to help address these questions, providing a high-fidelity quantitative model of mouse visual cortex. By analyzing factors contributing to model fidelity, we identified key principles underlying the organization of mouse visual cortex. Structurally, we find that comparatively low-resolution and shallow structure were both important for model correctness. Functionally, we find that models trained with task-agnostic, unsupervised objective functions, based on the concept of contrastive embeddings were substantially better than models trained with supervised objectives. Finally, the unsupervised objective builds a general-purpose visual representation that enables the system to achieve better transfer on out-of-distribution visual, scene understanding and reward-based navigation tasks. Our results suggest that mouse visual cortex is a low-resolution, shallow network that makes best use of the mouse’s limited resources to create a light-weight, general-purpose visual system – in contrast to the deep, high-resolution, and more task-specific visual system of primates.

How could machines learn as efficiently as humans and animals? How could machines learn to reason and plan? How could machines learn representations of percepts and action plans at multiple levels of abstraction, enabling them to reason, predict, and plan at multiple time horizons? I will propose a possible path towards autonomous intelligent agents, based on a new modular cognitive architecture and a somewhat new self supervised training paradigm. The centerpiece of the proposed architecture is a configurable predictive world model that allows the agent to plan. Behavior and learning are driven by a set of differentiable intrinsic cost functions. The world model uses a new type of energy-based model architecture called H-JEPA (Hierarchical Joint Embedding Predictive Architecture). H-JEPA learns hierarchical abstract representations of the world that are simultaneously maximally informative and maximally predictable.

Research in the brain and cognitive sciences attempts to uncover the neural mechanisms underlying intelligent behavior in domains such as vision. Due to the complexities of brain processing, studies necessarily had to start with a narrow scope of experimental investigation and computational modeling. I argue that it is time for our field to take the next step: build system models that capture a range of visual intelligence behaviors along with the underlying neural mechanisms. To make progress on system models, we propose integrative benchmarking – integrating experimental results from many laboratories into suites of benchmarks that guide and constrain those models at multiple stages and scales. We show-case this approach by developing Brain-Score benchmark suites for neural (spike rates) and behavioral experiments in the primate visual ventral stream. By systematically evaluating a wide variety of model candidates, we not only identify models beginning to match a range of brain data (~50% explained variance), but also discover that models’ brain scores are predicted by their object categorization performance (up to 70% ImageNet accuracy). Using the integrative benchmarks, we develop improved state-of-the-art system models that more closely match shallow recurrent neuroanatomy and early visual processing to predict primate temporal processing and become more robust, and require fewer supervised synaptic updates. Taken together, these integrative benchmarks and system models are first steps to modeling the complexities of brain processing in an entire domain of intelligence.

Biologically plausible spiking neural networks (SNNs) are an emerging architecture for deep learning tasks due to their energy efficiency when implemented on neuromorphic hardware. However, many of the biological features are at best irrelevant and at worst counterproductive when evaluated in the context of task performance and suitability for neuromorphic hardware. In this talk, I will present an alternative paradigm to design deep learning architectures with good task performance in real-world benchmarks while maintaining all the advantages of SNNs. We do this by focusing on two main features -- event-based computation and activity sparsity. Starting from the performant gated recurrent unit (GRU) deep learning architecture, we modify it to make it event-based and activity-sparse. The resulting event-based GRU (EGRU) is extremely efficient for both training and inference. At the same time, it achieves performance close to conventional deep learning architectures in challenging tasks such as language modelling, gesture recognition and sequential MNIST

Computational imaging seeks to achieve novel capabilities and overcome conventional limitations by combining optics and algorithms. In this seminar, I will discuss two computational imaging technologies developed in Boston University Computational Imaging Systems lab, including Intensity Diffraction Tomography and Computational Miniature Mesoscope. In our intensity diffraction tomography system, we demonstrate 3D quantitative phase imaging on a simple LED array microscope. We develop both single-scattering and multiple-scattering models to image complex biological samples. In our Computational Miniature Mesoscope, we demonstrate single-shot 3D high-resolution fluorescence imaging across a wide field-of-view in a miniaturized platform. We develop methods to characterize 3D spatially varying aberrations and physical simulator-based deep learning strategies to achieve fast and accurate reconstructions. Broadly, I will discuss how synergies between novel optical instrumentation, physical modeling, and model- and learning-based computational algorithms can push the limits in biomedical microscopy and neural imaging.

Discriminating distinct objects and concepts from sensory stimuli is essential for survival. Our brains accomplish this feat by forming meaningful internal representations in deep sensory networks with plastic synaptic connections. Experience-dependent plasticity presumably exploits temporal contingencies between sensory inputs to build these internal representations. However, the precise mechanisms underlying plasticity remain elusive. We derive a local synaptic plasticity model inspired by self-supervised machine learning techniques that shares a deep conceptual connection to Bienenstock-Cooper-Munro (BCM) theory and is consistent with experimentally observed plasticity rules. We show that our plasticity model yields disentangled object representations in deep neural networks without the need for supervision and implausible negative examples. In response to altered visual experience, our model qualitatively captures neuronal selectivity changes observed in the monkey inferotemporal cortex in-vivo. Our work suggests a plausible learning rule to drive learning in sensory networks while making concrete testable predictions.

A central goal of vision science is to understand the principles underlying the perception and neural coding of the complex visual environment of our everyday experience. In the visual cortex, foundational work with artificial stimuli, and more recent work combining natural images and deep convolutional neural networks, have revealed much about the tuning of cortical neurons to specific image features. However, a major limitation of this existing work is its focus on single-neuron response strength to isolated images. First, during natural vision, the inputs to cortical neurons are not isolated but rather embedded in a rich spatial and temporal context. Second, the full structure of population activity—including the substantial trial-to-trial variability that is shared among neurons—determines encoded information and, ultimately, perception. In the first part of this talk, I will argue for a normative approach to study encoding of natural images in primary visual cortex (V1), which combines a detailed understanding of the sensory inputs with a theory of how those inputs should be represented. Specifically, we hypothesize that V1 response structure serves to approximate a probabilistic representation optimized to the statistics of natural visual inputs, and that contextual modulation is an integral aspect of achieving this goal. I will present a concrete computational framework that instantiates this hypothesis, and data recorded using multielectrode arrays in macaque V1 to test its predictions. In the second part, I will discuss how we are leveraging this framework to develop deep probabilistic algorithms for natural image and video segmentation.

Building conceptual abstractions from sensory information and then reasoning about them is central to human intelligence. Abstract reasoning both relies on, and is facilitated by, our ability to make analogies about concepts from known domains to novel domains. Structure Mapping Theory of human analogical reasoning posits that analogical mappings rely on (higher-order) relations and not on the sensory content of the domain. This enables humans to reason systematically about novel domains, a problem with which machine learning (ML) models tend to struggle. We introduce a two-stage neural net framework, which we label Neural Structure Mapping (NSM), to learn visual analogies from Raven's Progressive Matrices, an abstract visual reasoning test of fluid intelligence. Our framework uses (1) a multi-task visual relationship encoder to extract constituent concepts from raw visual input in the source domain, and (2) a neural module net analogy inference engine to reason compositionally about the inferred relation in the target domain. Our NSM approach (a) isolates the relational structure from the source domain with high accuracy, and (b) successfully utilizes this structure for analogical reasoning in the target domain.

Knowledge discovery with computers requires a huge amount of search. Analogical reasoning is effective for efficient knowledge discovery. Therefore, we proposed analogical reasoning systems based on first-order predicate logic using Neuro-Symbolic AI. Neuro-Symbolic AI is a combination of Symbolic AI and artificial neural networks and has features that are easy for human interpretation and robust against data ambiguity and errors. We have implemented analogical reasoning systems by Neuro-symbolic AI models with word embedding which can represent similarity between words. Using the proposed systems, we efficiently extracted unknown rules from knowledge bases described in Prolog. The proposed method is the first case of analogical reasoning based on the first-order predicate logic using deep learning.

The brain represents the external world through the bottleneck of sensory organs. The network of hierarchically organized neurons is thought to recover the causes of sensory inputs to reconstruct the reality in the brain in idiosyncratic ways depending on individuals and their internal states. How can we understand the world model represented in an individual’s brain, or the neuroverse? My lab has been working on brain decoding of visual perception and subjective experiences such as imagery and dreaming using machine learning and deep neural network representations. In this talk, I will outline the progress of brain decoding methods and present how subjective experiences are externalized as images and how they could be shared across individuals via neural code conversion. The prospects of these approaches in basic science and neurotechnology will be discussed.

Modeling neural responses to naturalistic stimuli has been instrumental in advancing our understanding of the visual system. Dominant computational modeling efforts in this direction have been deeply rooted in preconceived hypotheses. In contrast, hypothesis-neutral computational methodologies with minimal apriorism which bring neuroscience data directly to bear on the model development process are likely to be much more flexible and effective in modeling and understanding tuning properties throughout the visual system. In this study, we develop a hypothesis-neutral approach and characterize response selectivity in the human visual cortex exhaustively and systematically via response-optimized deep neural network models. First, we leverage the unprecedented scale and quality of the recently released Natural Scenes Dataset to constrain parametrized neural models of higher-order visual systems and achieve novel predictive precision, in some cases, significantly outperforming the predictive success of state-of-the-art task-optimized models. Next, we ask what kinds of functional properties emerge spontaneously in these response-optimized models? We examine trained networks through structural ( feature visualizations) as well as functional analysis (feature verbalizations) by running `virtual' fMRI experiments on large-scale probe datasets. Strikingly, despite no category-level supervision, since the models are solely optimized for brain response prediction from scratch, the units in the networks after optimization act as detectors for semantic concepts like `faces' or `words', thereby providing one of the strongest evidences for categorical selectivity in these visual areas. The observed selectivity in model neurons raises another question: are the category-selective units simply functioning as detectors for their preferred category or are they a by-product of a non-category-specific visual processing mechanism? To investigate this, we create selective deprivations in the visual diet of these response-optimized networks and study semantic selectivity in the resulting `deprived' networks, thereby also shedding light on the role of specific visual experiences in shaping neuronal tuning. Together with this new class of data-driven models and novel model interpretability techniques, our study illustrates that DNN models of visual cortex need not be conceived as obscure models with limited explanatory power, rather as powerful, unifying tools for probing the nature of representations and computations in the brain.

To better understand human scene understanding, we extracted features from images using CLIP, a neural network model of visual concept trained with supervision from natural language. We then constructed voxelwise encoding models to explain whole brain responses arising from viewing natural images from the Natural Scenes Dataset (NSD) - a large-scale fMRI dataset collected at 7T. Our results reveal that CLIP, as compared to convolution based image classification models such as ResNet or AlexNet, as well as language models such as BERT, gives rise to representations that enable better prediction performance - up to a 0.86 correlation with test data and an r-square of 0.75 - in higher-level visual cortex in humans. Moreover, CLIP representations explain distinctly unique variance in these higher-level visual areas as compared to models trained with only images or text. Control experiments show that the improvement in prediction observed with CLIP is not due to architectural differences (transformer vs. convolution) or to the encoding of image captions per se (vs. single object labels). Together our results indicate that CLIP and, more generally, multimodal models trained jointly on images and text, may serve as better candidate models of representation in human higher-level visual cortex. The bridge between language and vision provided by jointly trained models such as CLIP also opens up new and more semantically-rich ways of interpreting the visual brain.

In the deep reinforcement learning (RL) community, motor control problems are usually approached from a reward-based learning perspective. However, humans are often believed to learn motor control through directed error-based learning. Within this learning setting, the control system is assumed to have access to exact error signals and their gradients with respect to the control signal. This is unlike reward-based learning, in which errors are assumed to be unsigned, encoding relative successes and failures. Here, we try to understand the relation between these two approaches, reward- and error- based learning, and ballistic arm reaches. To do so, we test canonical (deep) RL algorithms on a well-known sensorimotor perturbation in neuroscience: mirror-reversal of visual feedback during arm reaching. This test leads us to propose a potentially novel RL algorithm, denoted as model-based deterministic policy gradient (MB-DPG). This RL algorithm draws inspiration from error-based learning to qualitatively reproduce human reaching performance under mirror-reversal. Next, we show MB-DPG outperforms the other canonical (deep) RL algorithms on a single- and a multi- target ballistic reaching task, based on a biomechanical model of the human arm. Finally, we propose MB-DPG may provide an efficient computational framework to help explain error-based learning in neuroscience.