Speakers

DeepMind, London, UK

Reverse-Engineering Human Evolution with Multi-Agent Reinforcement Learning

In both cognitive science and computer science, intelligence has traditionally been viewed solipsistically, as a property of individual agents and devoid of social context. Yet converging evidence from multiple fields shows that human intelligence emerged in interdependent networks of interacting agents. For those of us interested in reverse-engineering human intelligence, such findings suggest constraints on the space of workable algorithms. In this talk I'll discuss how methods based on multi-agent reinforcement learning can be used as part of a research program that aims to catalog the great diversity of social challenges faced by hominins over the last two million years. This reverse-engineering approach aims to discover how the need to overcome such challenges may have facilitated the emergence of our most important social-cognitive capacities, representations, and motivations.

University of Warsaw, Poland

Nested timings of coaction in mother-infant interaction

In our recent work, we illustrate attunements to several timing orders that are essential to the emergence of language: i) the timing of participation within social routines; ii) the timing of vocalizations and iii) the timing of "action arcs" or "vitality contours", which are larger interactive structures. The goal of this work is to show how embodied interactions gain their linguistic character for the infant. Highlighting the importance of timing for learning to participate should improve our understanding of the progressive saturation of language with interactive structures for a child on the one hand and our understanding of the structuring of language as an interactive control on different timescales on the other.

Trinity College Dublin & Trinity College Institute of Neuroscience, Ireland

Developing development: Of babies and machines

The aims of developmental robotics are two-fold: (1) Using AI to assist developmental science in uncovering the nuts and bolts of human cognitive development, and (2) using insights from human child development to foster AI’s research goals. In this talk, I will first briefly have a look at the first of the two mentioned directions. I will shed light on how AI can help developmental scientists think about the underlying mechanisms and building blocks of cognitive capacities, taking previous work on the sense of agency as concrete example. I will then turn to the second direction – i.e., developmental science supporting advances in AI – for the main part of my talk. I will show how insights from child development can be used to improve AI, both in the context of the sense of agency as well as in machine learning more generally. Especially unsupervised machine learning can be greatly facilitated by lessons from infant cognition as, e.g., infants too do not learn from curated and labelled data or with vast amounts of prior knowledge. I will discuss three crucial factors enabling infants' quality and speed of learning, assess the extent to which these insights have already been exploited in machine learning, and provide concrete suggestions for how they can be further adopted in next-generation systems so that unsupervised machine learning can finally set the next step in growing up.

Neuro-Cognitive Modeling, Department of Computer Science and Department of Psychology, Faculty of Science, University of Tübingen, Germany (MB & CG)
& Max Planck Institute for Intelligent Systems, Tübingen, Germany (CG)

Towards Linking Event-Predictive, Sensorimotor Grounded Structures to Language

Hierarchical, compositional models and behavioral primitives constitute crucial components for interacting with our world in a versatile and adaptive manner. Event-predictive cognition offers a theoretical framework on how such models may develop and may be invoked in a self-motivated manner. Interestingly, evidence is accumulating that these models offer themselves as direct links to language. We selectively introduce some of our recent recurrent artificial neural network models, sketching-out a pathway on how to develop event-predictive Gestalten and how their anticipatory, self-motivated activation can model human-like behavior. Moreover, we show that latent, symbol-like structures develop, which strongly resemble typical event-descriptive language structures. Research thus will soon close the gap between current language models and sensorimotor-grounded, embodied symbol-like structures – with all consequent implications for strong AI.

Microsoft Research NYC, USA

The Graph Structure of Collective Cognition and Innovation in Humans and Machine

Cognition and innovation do not take place in a vacuum. From what we remember and forget, to hierarchical discovery and invention, both human and machine intelligence rely on collective phenomena, including social norms, collective beliefs, and group structures within which humans and machines are embedded. Can we quantify how the structure of social networks shapes collective memory, cognition, and innovation? I will first discuss a series of our human behavioral studies showing how the graph structure of social networks shape collective memories. I will then share behavioral and neuroimaging studies from colleagues supporting the view that human brains are sensitive to the graph structure of social networks and that graph structure of social groups shapes problem solving. I will then show SAPIENS, where we built different social network structures, according to which deep RL agents shared their experience (the content of their replay buffer) with one another, toward the goal of hierarchical innovation. I will conclude that the graph structure of social networks shape collective cognition and hierarchical innovation in humans and machines, and discuss implications for understanding and designing systemic interventions and for evolution-inspired approaches in machine learning.

CNRS, Pascal Institute, Clermont-Ferrand University, France

Toddler-inspired learning induces hierarchical object representations

Humans learn to both visually recognize individual objects and categorize them at different levels of abstraction. Such multi-semantic representation is crucial to efficiently reason about the world. However, it is currently unclear how such representations could be learned with the very sparse labeling available to human learners. To answer this question we let an artificial agent play with objects while occasionally “hearing” their category label. Our agent assigns similar representations to a) similarly labelled and b) close-in-time visual inputs. We show that our agent learns a 2-level hierarchical representation that first aggregates different views of objects and then brings together different objects to form categories. Interestingly, we do not observe a trade-off between each semantic content. Our work suggests that the temporal structure of visual experience during object play together with occasional labeling suffice for learning a hierarchically structured object/category representation.

PPSP team, CHU Sainte Justine Research Center,
Université de Montréal, Canada

Modeling sensorimotor dynamics of embodied collective decision making

Models of collective decision-making have so far focused mostly on dynamics at the collective level, while the individual agents followed well-defined decision rules. Yet, living agents also have complex internal dynamics that determine how they will react to social information. We introduce an approach for simultaneously modeling both intra-agent and inter- agent dynamics during collective decision-making. We apply this approach to a simple task where a group of agents has to collectively move towards one of several possible targets. Our aim is to show that, by allowing the agents to use sensorimotor feedback to change their internal dynamics, agents can efficiently coordinate their movements and reach a consensus.