Cette page dédiée présente l’un des thème / domaine stratégique actuellement en discussion dans le cadre de notre Programme de financement de recherche stratégique. L’ensemble des thèmes / domaines en discussion est indiqué sur la page du programme. Chaque page dédiée (y compris celle-ci) peut être à un niveau de détail et de maturité variable. Pour participer à la discussion sur ce thème ou en proposer de nouveaux, veuillez utiliser ce formulaire. Si vous souhaitez être tenu.e au courant des développements autour de ce thème / domaine, inscrivez-vous ci-dessous.

Next-generation neuro-digital intelligence

Description et justification du domaine

Despite the undeniable power of recent advances in digital intelligence, artificial systems still exhibit several limitations that prevent them from maximizing their economic and social impact. Specifically, current digital intelligence is lacking five essential attributes, all of which are intimately related to neuroscience and psychology:

(1) Out-of-distribution systematic generalization and reasoning in real-world settings

The human brain is capable of learning models of the world, logical rules, and causal relationships and using those to make inferences and engage in planning. Current AI systems are only capable of engaging in such systematic reasoning in highly constrained toy environments (such as a board game) and within the same data distribution as that in which they were trained. Achieving systematic generalization in out-of-distribution and real-world environments will likely require taking inspiration from the one system that displays these abilities and thus provides a proof-of-principle solution: the human brain.

(2) Self-supervised learning without rewards or labels

Animals learn useful representations with very few external rewards or instructions. Indeed, the mammalian brain can learn how to structure its representations of the world simply from exposure to the world. This has inspired self-supervised learning in AI, and self-supervised learning has recently seen great success,but it remains limited to simple contexts. To achieve the rich representation learning powers of the real brain, we would likely benefit greatly from direct neural inspiration. How does the brain learn to identify important latent variables? What are the inductive biases that promote the learning of an understanding of objects, part-whole structures and other relational knowledge, physical interactions, etc.?

(3) Continual learning from brains to machines

Humans and animals accumulate knowledge and skills by learning continually throughout their lives. Current AI systems on the other hand, are generally trained once on a comprehensive dataset and then deployed. While it is natural for biological brains to iteratively refine previously learned tasks with new data in a streaming manner, current AI approaches suffer from catastrophic forgetting in such settings. Addressing this shortcoming could dramatically expand the applicability of digital intelligence in adaptive contexts. By studying the biological mechanisms of learning, such as synaptic plasticity and the gating of learning by internal prediction, we aim to understand how brains achieve continual learning, and devise adaptive AI systems which share this capability.

(4) Abstracting from action-perception loops in interactive environments

As first articulated by ecological psychologists, animals are not passive recipients of information from the world. Our sensations are shaped not only by the external world, but also by our actions within it. These interactive action-perception loops (or “affordances”) are likely key to our ability to learn more abstract, causal representations for intelligent decision making and adaptive motor control. However, current AI systems lack the ability to form abstract understandings based on affordances, leaving many unanswered questions. How do abstract (i.e., verbalizable) actions and motor signals shape sensory predictions and sensory interpretation? How does learning proceed from action-perception loops to more abstract knowledge? What inductive biases allow animals to learn causal relationships from interactions with the world and with other agents? These questions are relevant for BMI research as well as for the next topic: human-in-the-loop hybrid AI.

(5) Human-in-the-loop hybrid artificial intelligence

It is becoming increasingly clear that fully autonomous AI systems are likely many years away. For example, estimates of the arrival of fully self-driving cars have been continuously pushed back, and are now many decades in the future. In all likelihood, AI will have its biggest impact on the economy and our lives via hybrid, interactive, human-in-the-loop systems. This mandates a focus on AI that is designed to interact well with humans and their cognitive proclivities. Human-centric AI that takes into consideration the natural operating modes of the human brain will require insights from psychology and neuroscience. Moreover, brain-machine interface systems may open the door to seamless human-AI interaction, and should be a focus of the next generation of human-in-the-loop, hybrid AI systems.

We propose that one of IVADO’s themes should be the development of “neuro-digital intelligence”. This includes both brain-inspired AI systems, and human-in-the-loop AI focussed on intimate interactions with living brains. Ultimately, our proposal is to have a theme devoted to developing the next-generation of digital intelligence via interdisciplinary interactions with the brain sciences. In particular, we envision a theme focussed on achieving the five attributes listed above (systematic generalization, self-supervised learning, continual learning, action-perception abstraction, and human-in-the-loop AI) via an explicit focus on collaborations between Montreal’s flourishing machine learning and brain sciences communities. We believe that IVADO is uniquely well-positioned to help fill these current gaps in digital intelligence, and that the research ecosystem in Montreal is perfectly suited to help make this happen, thanks to the existence of a critical mass of researchers in brain-inspired AI and other strong networks, such as UNIQUE and HBHL. Applications of such a neuro-digital intelligence can impact many domains, and even feedback into the investigation of the human brain itself. Indeed, understanding brain functions and dysfunctions requires the combination of large amounts of data, across scales and modalities, with fragmented prior knowledge about the biological mechanisms underlying our cognitive abilities.

Contexte

Mots-clefs : artificial intelligence, neuroscience, psychology, generalization, self-supervised learning, continual learning, action-perception, human-in-the-loop AI

Organisations pertinentes :

Mila, UNIQUE, Universite de Montreal, McGill University,  Polytechnique Montreal (https://www.polymtl.ca/), HBHL, CIFAR program on « Learning in Machines and Brains »

IVADO-UNIQUE team grant on « Continual Learning from Brains to Machines », CERC, Google Deepmind, Microsoft Research, Facebook AI, IBM Research.

Personnes pertinentes suggérées durant la consultation :

Les noms suivants ont été proposés par la communauté et les personnes mentionnées ci-dessous ont accepté d’afficher publiquement leur nom. Notez cependant que tous les noms des professeur.e.s (qu’ils soient affichés publiquement ou non sur notre site web) seront transmis au comité conseil pour l’étape d’identification et de sélection des thèmes stratégiques. Notez également que les personnes identifiées durant l’étape de consultation n’ont pas la garantie de recevoir une partie du financement. Cette étape sert avant tout à présenter un panorama du domaine, incluant les personnes pertinentes et non à monter des équipes pour les programmes-cadres.

  • Andrea Green
  • Anne Gallagher
  • Blake Richards
  • Eilif Muller
  • Frédéric Gosselin
  • Guillaume Dumas
  • Guillaume Lajoie
  • Julien Cohen-Adad
  • Pouya Bashivan
  • Karim Jerbi
  • Yoshua Bengio
  • Numa Dancause
  • Paul Cisek
  • Pierre Rainville
  • Tomas Paus

Programmes-cadres potentiels

(pas de programmes-cadres potentiels pour le moment)

Documentation complémentaire

(pas de documentation complémentaire pour le moment)

Intéressé.e? Entrez votre courriel pour recevoir les mises à jour en lien avec cette page :

Historique

13 juillet 2021 : Première version

15 juillet 2021 : Ajout de personnes pertinentes

22 juillet 2021 : Ajout de personnes pertinentes

17 août 2021 : Ajout de personnes pertinentes