Machine 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.

Analogical reasoning –learning about new things by relating it to previous knowledge– lies at the heart of human intelligence and creativity and forms the core of educational practice. Children start creating and using analogies early on, making incredible progress moving from associative processes to successful analogical reasoning. For example, if we ask a four-year-old “Horse belongs to stable like chicken belongs to …?” they may use association and reply “egg”, whereas older children will likely give the intended relational response “chicken coop” (or other term to refer to a chicken’s home). Interestingly, despite state-of-the-art AI-language models having superhuman encyclopedic knowledge and superior memory and computational power, our pilot studies show that these large language models often make mistakes providing associative rather than relational responses to verbal analogies. For example, when we asked four- to eight-year-olds to solve the analogy “body is to feet as tree is to …?” they responded “roots” without hesitation, but large language models tend to provide more associative responses such as “leaves”. In this study we examine the similarities and differences between children's and six large language models' (Dutch/multilingual models: RobBERT, BERT-je, M-BERT, GPT-2, M-GPT, Word2Vec and Fasttext) responses to verbal analogies extracted from an online adaptive learning environment, where >14,000 7-12 year-olds from the Netherlands solved 20 or more items from a database of 900 Dutch language verbal analogies.

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.

A central problem in biological learning is how information about the outcome of a decision or behavior can be used to reliably guide learning across distributed neural circuits while obeying biological constraints. This “credit assignment” problem is commonly solved in artificial neural networks through supervised gradient descent and the backpropagation algorithm. In contrast, biological learning is typically modelled using unsupervised Hebbian learning rules. While these rules only use local information to update synaptic weights, and are sometimes combined with weight constraints to reflect a diversity of excitatory (only positive weights) and inhibitory (only negative weights) cell types, they do not prescribe a clear mechanism for how to coordinate learning across multiple layers and propagate error information accurately across the network. In recent years, several groups have drawn inspiration from the known dendritic non-linearities of pyramidal neurons to propose new learning rules and network architectures that enable biologically plausible multi-layer learning by processing error information in segregated dendrites. Meanwhile, recent experimental results from the hippocampus have revealed a new form of plasticity—Behavioral Timescale Synaptic Plasticity (BTSP)—in which large dendritic depolarizations rapidly reshape synaptic weights and stimulus selectivity with as little as a single stimulus presentation (“one-shot learning”). Here we explore the implications of this new learning rule through a biologically plausible implementation in a rate neuron network. We demonstrate that regulation of dendritic spiking and BTSP by top-down feedback signals can effectively coordinate plasticity across multiple network layers in a simple pattern recognition task. By analyzing hidden feature representations and weight trajectories during learning, we show the differences between networks trained with standard backpropagation, Hebbian learning rules, and BTSP.

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.

Neural computations are currently investigated using two separate approaches: sorting neurons into functional subpopulations or examining the low-dimensional dynamics of collective activity. Whether and how these two aspects interact to shape computations is currently unclear. Using a novel approach to extract computational mechanisms from networks trained on neuroscience tasks, here we show that the dimensionality of the dynamics and subpopulation structure play fundamentally com- plementary roles. Although various tasks can be implemented by increasing the dimensionality in networks with fully random population structure, flexible input–output mappings instead require a non-random population structure that can be described in terms of multiple subpopulations. Our analyses revealed that such a subpopulation structure enables flexible computations through a mechanism based on gain-controlled modulations that flexibly shape the collective dynamics. Our results lead to task-specific predictions for the structure of neural selectivity, for inactivation experiments and for the implication of different neurons in multi-tasking.

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.

Humans perceive the world in terms of objects and relations between them. In fact, for any given pair of objects, there is a myriad of relations that apply to them. How does the cognitive system learn which relations are useful to characterize the task at hand? And how can it use these representations to build a relational policy to interact effectively with the environment? In this paper we propose that this problem can be understood through the lens of a sub-field of symbolic machine learning called relational reinforcement learning (RRL). To demonstrate the potential of our approach, we build a simple model of relational policy learning based on a function approximator developed in RRL. We trained and tested our model in three Atari games that required to consider an increasingly number of potential relations: Breakout, Pong and Demon Attack. In each game, our model was able to select adequate relational representations and build a relational policy incrementally. We discuss the relationship between our model with models of relational and analogical reasoning, as well as its limitations and future directions of research.

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

Animals exhibit remarkable learning capabilities with little direct supervision. Likewise, self-supervised learning is an emergent paradigm in artificial intelligence, closing the performance gap to supervised learning. In the context of biology, self-supervised learning corresponds to a setting where one sense or specific stimulus may serve as a supervisory signal for another. After learning, the latter can be used to predict the former. On the implementation level, it has been demonstrated that such predictive learning can occur at the single neuron level, in compartmentalized neurons that separate and associate information from different streams. We demonstrate the power such self-supervised learning over unsupervised (Hebb-like) learning rules, which depend heavily on stimulus statistics, in two examples: First, in the context of animal navigation where predictive learning can associate internal self-motion information always available to the animal with external visual landmark information, leading to accurate path-integration in the dark. We focus on the well-characterized fly head direction system and show that our setting learns a connectivity strikingly similar to the one reported in experiments. The mature network is a quasi-continuous attractor and reproduces key experiments in which optogenetic stimulation controls the internal representation of heading, and where the network remaps to integrate with different gains. Second, we show that incorporating global gating by reward prediction errors allows the same setting to learn conditioning at the neuronal level with mixed selectivity. At its core, conditioning entails associating a neural activity pattern induced by an unconditioned stimulus (US) with the pattern arising in response to a conditioned stimulus (CS). Solving the generic problem of pattern-to-pattern associations naturally leads to emergent cognitive phenomena like blocking, overshadowing, saliency effects, extinction, interstimulus interval effects etc. Surprisingly, we find that the same network offers a reductionist mechanism for causal inference by resolving the post hoc, ergo propter hoc fallacy.

Which nodes in a brain network causally influence one another, and how do such interactions utilize the underlying structural connectivity? One of the fundamental goals of neuroscience is to pinpoint such causal relations. Conventionally, these relationships are established by manipulating a node while tracking changes in another node. A causal role is then assigned to the first node if this intervention led to a significant change in the state of the tracked node. In this presentation, I use a series of intuitive thought experiments to demonstrate the methodological shortcomings of the current ‘causation via manipulation’ framework. Namely, a node might causally influence another node, but how much and through which mechanistic interactions? Therefore, establishing a causal relationship, however reliable, does not provide the proper causal understanding of the system, because there often exists a wide range of causal influences that require to be adequately decomposed. To do so, I introduce a game-theoretical framework called Multi-perturbation Shapley value Analysis (MSA). Then, I present our work in which we employed MSA on an Echo State Network (ESN), quantified how much its nodes were influencing each other, and compared these measures with the underlying synaptic strength. We found that: 1. Even though the network itself was sparse, every node could causally influence other nodes. In this case, a mere elucidation of causal relationships did not provide any useful information. 2. Additionally, the full knowledge of the structural connectome did not provide a complete causal picture of the system either, since nodes frequently influenced each other indirectly, that is, via other intermediate nodes. Our results show that just elucidating causal contributions in complex networks such as the brain is not sufficient to draw mechanistic conclusions. Moreover, quantifying causal interactions requires a systematic and extensive manipulation framework. The framework put forward here benefits from employing neural network models, and in turn, provides explainability for them.

Spiking recurrent neural networks (RNNs) are a promising tool for solving a wide variety of complex cognitive and motor tasks, due to their rich temporal dynamics and sparse processing. However training spiking RNNs on dedicated neuromorphic hardware is still an open challenge. This is due mainly to the lack of local, hardware-friendly learning mechanisms that can solve the temporal credit assignment problem and ensure stable network dynamics, even when the weight resolution is limited. These challenges are further accentuated, if one resorts to using memristive devices for in-memory computing to resolve the von-Neumann bottleneck problem, at the expense of a substantial increase in variability in both the computation and the working memory of the spiking RNNs. In this talk, I will present our recent work where we introduced a PyTorch simulation framework of memristive crossbar arrays that enables accurate investigation of such challenges. I will show that recently proposed e-prop learning rule can be used to train spiking RNNs whose weights are emulated in the presented simulation framework. Although e-prop locally approximates the ideal synaptic updates, it is difficult to implement the updates on the memristive substrate due to substantial device non-idealities. I will mention several widely adapted weight update schemes that primarily aim to cope with these device non-idealities and demonstrate that accumulating gradients can enable online and efficient training of spiking RNN on memristive substrates.

In the Learning Salon, we will discuss the similarities and differences between biological and machine learning, including individuals with diverse perspectives and backgrounds, so we can all learn from one another.