Project

Back to overview

Probabilistic learning in deep cortical networks

Applicant Sacramento João
Number 186027
Funding scheme Ambizione
Research institution Institut für Neuroinformatik Universität Zürich Irchel und ETH Zürich
Institution of higher education ETH Zurich - ETHZ
Main discipline Biophysics
Start/End 01.04.2020 - 31.03.2024
Approved amount 811'352.00
Show all

Keywords (7)

theoretical neuroscience; model uncertainty; Bayesian parameter inference; neuroinformatics; deep cortical network models; deep neural networks; probabilistic learning

Lay Summary (German)

Lead
Künstliche neuronale Netzwerke spielen sowohl als Systeme für maschinelles Lernen als auch als Modelle für neurowissenschaftliche Experimente eine wichtige Rolle. Gegenwärtig stellt es für solche Netzwerke jedoch eine Herausforderung dar, Unsicherheit in ihren Vorhersagen adäquat abzubilden. Während Menschen damit keine Probleme haben, erweist sich dieses Defizit in vielen Anwendungsszenarien als problematisch. Wir werden in dieser Hinsicht verschiedene Lösungsansätze untersuchen, die auf bayesscher Inferenz aufbauen.
Lay summary

Künstliche neuronale Netzwerke beruhen auf einer stark vereinfachten Abstraktionen ihrer biologischen Vorbilder. Sie erzielen beachtliche Erfolge bei der Lösung zahlreicher Problemstellungen. Zum Beispiel erkennen sie subtile Muster in medizinischen Bildmaterial. Dennoch haben viele neuronale Netzwerke derzeit Schwierigkeiten die Verlässlichkeit ihrer Vorhersagen einzuschätzen. Werden ihnen neue Daten präsentiert, die sich stark von dem bisher erlernten unterscheiden, schneiden sie häufig schlecht ab und können nicht zum Ausdruck bringen, dass ihre Vorhersagen unpräzise sind. Im Gegensatz dazu beziehen Menschen Unsicherheit automatisch in Entscheidungsprozesse ein. In diesem Projekt untersuchen wir neuronale Netzwerkmodelle des Lernens. Zunächst ist unser Ziel, diese in die Lage zu versetzen, die Unsicherheit ihrer Einschätzungen besser abzubilden. Darauf aufbauend werden wir Theorien und Hypothesen generieren, wie biologische neuronale Netzwerke mit Unsicherheit umgehen. Wir betrachten Lernprozesse aus einer probabilistischen Perspektive, aufbauend auf bayesscher Wahrscheinlichkeitstheorie. Unsere Arbeit wird durch die neurowissenschaftlichen Experimente unserer Kollegen am Institut für Neuroinformatik in Zürich inspiriert.

Direct link to Lay Summary Last update: 08.04.2020

Responsible applicant and co-applicants

Employees

Publications

Publication
A contrastive rule for meta-learning
Zucchet Nicolas, Schug Simon, von Oswald Johannes, Zhao Dominic, Sacramento João (2021), A contrastive rule for meta-learning, arXiv, n/a.
Credit assignment in neural networks through deep feedback control
Meulemans Alexander, Tristany Farinha Matilde, Ordóñez Javier Garcia, Vilimelis Aceituno Pau, Sacramento João, Grewe Benjamin F. (2021), Credit assignment in neural networks through deep feedback control, in Advances in Neural Information Processing Systems, 34, Curran Associates, Inc., United States 34.
Learning where to learn: gradient sparsity in meta and continual learning
von Oswald Johannes, Zhao Dominic, Kobayashi Seijin, Schug Simon, Caccia Massimo, Zucchet Nicolas, Sacramento João (2021), Learning where to learn: gradient sparsity in meta and continual learning, in Advances in Neural Information Processing Systems, 34, Curran Associates, Inc., United States 34.
Neural networks with late-phase weights
von Oswald Johannes, Kobayashi Seijin, Meulemans Alexander, Henning Christian, Grewe Benjamin F., SacramentoJoão (2021), Neural networks with late-phase weights, in Proceedings of the 9th International Conference on Learning Representations, ICLR 2021, OpenReview.net, n/a.
Posterior meta-replay for continual learning
Henning Christian, R. Cervera Maria, D'Angelo Francesco, von Oswald Johannes, Traber Regina, Ehret Benjamin, Kobayashi Seijin, Grewe Benjamin F., Sacramento João (2021), Posterior meta-replay for continual learning, in Advances in Neural Information Processing Systems, 34, Curran Associates, Inc., United States 34.
Presynaptic stochasticity improves energy efficiency and helps alleviate the stability-plasticity dilemma
Schug Simon, Benzing Frederik, Steger Angelika (2021), Presynaptic stochasticity improves energy efficiency and helps alleviate the stability-plasticity dilemma, in eLife, 10, 69884-69884.
A theoretical framework for target propagation
MeulemansAlexander, CarzanigaFrancesco, SuykensJohan, SacramentoJoão, GreweBenjamin F. (2020), A theoretical framework for target propagation, in Advances in Neural Information Processing Systems, Curran Associates, Inc., United States.
Meta-learning via hypernetworks
Zhao Dominic, Kobayashi Seijin, Sacramento João, von OswaldJohannes (2020), Meta-learning via hypernetworks, n/a, n/a.

Collaboration

Group / person Country
Types of collaboration
Angelika Steger (ETHZ D-INFK) Switzerland (Europe)
- in-depth/constructive exchanges on approaches, methods or results

Scientific events

Active participation

Title Type of contribution Title of article or contribution Date Place Persons involved
2021 Champalimaud Research Symposium Poster A contrastive rule for meta-learning 13.10.2021 Lisbon, Portugal Schug Simon; Sacramento João;
Max Planck Institute CBS CoCoNUT Seminar Individual talk A contrastive rule for meta-learning 17.09.2021 Virtual, Germany Sacramento João;


Abstract

To survive and thrive in continuously changing environments animals must constantly translate the information provided by the external world into appropriate behavior. In practice, however, this proves difficult because sensory information is almost always incomplete, sparse and noisy. Furthermore, their ability to hold and transform information is limited. Thus, the internal representations of information acquired by an organism - and, ultimately, the decisions that it can make - are fundamentally uncertain.Uncertainty can be dealt with in a principled manner within the mathematical framework of Bayesian probability theory, which offers a method of formalizing and combining prior beliefs with available data in a rational and, in some sense, optimal way. The solid theoretical grounding of the probabilistic approach and its practical appeal have led to countless applications in almost every field of engineering and science. The ubiquity of Bayesian inference is however not limited to engineering and scientific endeavor: mounting evidence from behavioral and neuroscience studies supports that humans and primates, but also simpler organisms naturally process information according to the rules of probability theory.In line with such empirical findings, the idea that the brain performs probabilistic inference is highly influential in theoretical neuroscience, that has inspired models for a wide range of tasks, from sensory perception all the way down to behavior and decision-making. However, the majority of neuroscience theories do not approach learning as a problem of probabilistic inference. This is striking, as learning plays a crucial and natural role in probabilistic modelling and learning can itself be understood as a problem of probabilistic inference, with the goal of seeking plausible models that explain observed data. Such models are then used as the basis to make predictions. Importantly, unlike in non-probabilistic approaches, predictions are explicitly (and quantitatively) informed by model uncertainty. Model uncertainty arises naturally because data is limited and noisy: a set of observations is therefore expected to be consistent with many different models. Neglecting this will often result in overconfident predictions, poor generalization, as well as the inability to communicate prediction uncertainty to downstream neural circuits.The aim of my research is to address this major problem and develop models of probabilistic learning in cortical neural circuits that represent parameter uncertainty and leverage such knowledge when making predictions. I will model cortical learning as an inference process in which synaptic weights correspond to model parameters. Because exact Bayesian inference is intractable in neural networks, my first objective will be to develop a biologically-plausible approximate inference method, resorting to a type of approximation known as stochastic variational inference. This leads me to consider the hypothesis that synapses evolve stochastically by sampling from distributions of approximating models that are consistent with data, and that their temporal variability encodes uncertainty. As a second objective, I will then develop a unified model which takes into account both synaptic weight and neural activity uncertainty, which plays a pivotal role in probabilistic theories of perception and decision-making. Finally, for my third objective, I will increase the level of biological realism of my models. This will allow me to better capture existing experimental data and propose novel experiments.I will focus on a class of models that I termed deep cortical networks (DCNs), that lie at the intersection of neuroscience and machine learning while keeping a balance between biological detail and computational capability. This will allow me to connect with ongoing research in artificial deep neural networks (DNNs; where Bayesian learning is an active and important current research direction) while advancing our understanding of if, and how, cortical circuits adjust and acquire information probabilistically. If my research is successful, it will result in novel, experimentally-falsifiable neural circuit learning models, as well as artificial neural networks with improved performance.
-