Cognitively Informed Artificial Intelligence

Overview

The goal of this workshop is to bring together cognitive scientists, neuroscientists, and machine-learning researchers to discuss opportunities for improving AI by leveraging our scientific understanding of human perception and cognition. There is a history of making these connections: artificial neural networks were originally motivated by the massively parallel, layered architecture of the brain; considerations of biological plausibility have driven the development of learning procedures; and architectures for computer vision draw parallels to the connectivity and physiology of mammalian visual cortex. However, beyond these celebrated examples, cognitive science and neuroscience has fallen short of its potential to influence the next generation of AI systems. Areas such as memory, attention, and development have rich theoretical and experimental histories, yet these concepts, as applied to AI systems so far, only bear a superficial resemblance to their biological counterparts.

The premise of this workshop is that there are valuable data and models from cognitive science that can inform the development of intelligent adaptive machines, and can endow learning architectures with the strength and flexibility of the human cognitive architecture. The structures and mechanisms of the mind and brain can provide the sort of strong inductive bias needed for machine-learning systems to attain human-like performance. This inductive bias will become more important as researchers move from domain-specific tasks such as object and speech recognition toward tackling general intelligence and the human-like ability to dynamically reconfigure cognition in service of changing goals. For ML researchers, the workshop will provide access to a wealth of data and concepts situated in the context of contemporary ML. For cognitive scientists, the workshop will suggest research questions that are of critical interest to ML researchers.

The workshop will focus on three interconnected topics of particular relevance to ML:

(1) Learning and development. Cognitive capabilities expressed early in a child’s development are likely to be crucial for bootstrapping adult learning and intelligence. Intuitive physics and intuitive psychology allow the developing organism to build an understanding of the world and of other agents. Additionally, children and adults often demonstrate “learning-to-learn,” where previous concepts and skills form a compositional basis for learning new concepts and skills.

(2) Memory. Human memory operates on multiple time scales, from memories that literally persist for the blink of an eye to those that persist for a lifetime. These different forms of memory serve different computational purposes. Although forgetting is typically thought of as a disadvantage, the ability to selectively forget/override irrelevant knowledge in nonstationary environments is highly desirable.

(3) Attention and Decision Making. These refer to relatively high-level cognitive functions that allow task demands to purposefully control an agent’s external environment and sensory data stream, dynamically reconfigure internal representation and architecture, and devise action plans that strategically trade off multiple, oft-conflicting behavioral objectives.

Our long-term goals are:

• to incorporate insights from human cognition to suggest novel and improved AI architectures;
• to facilitate the development of ML methods that can better predict human behavior; and
• to support the development of a field of ‘cognitive computing’ that is more than a marketing slogan一a field that improves on both natural and artificial cognition by synergistically advancing each and integrating their strengths in complementary manners.

Organizers

Brenden Lake, NYU

Mike Mozer, U. Colorado Boulder

Angela Yu, UCSD

Invited speakers

Peter Battaglia (Deep Mind), Object-oriented intelligence

Yoshua Bengio (U. Montreal), From deep learning of disentangled representations to higher-level cognition

Matt Botvinick (Deep Mind), Meta-reinforcement learning in brains and machines

Alison Gopnik (UC Berkeley), Life history and learning: Extended human childhood as a way to resolve explore/exploit trade-offs and improve hypothesis search

Tom Griffiths (UC Berkeley), Revealing human inductive biases and metacognitive processes with rational models

Marc Howard (Boston University), Scale invariant temporal memory in AI

Robert Jacobs (U. Rochester), People infer object shape in a 3D, object-centered coordinate system

Gary Marcus (NYU), Representational primitives, in minds and machines

Aude Oliva (MIT), Mapping the spatio-temporal dynamics of cognition in the human brain

Contributing to the workshop

The goal of this workshop is to bring together cognitive scientists, neuroscientists, and AI researchers to discuss opportunities for improving machine learning, by leveraging our scientific understanding of human perception and cognition. We have reserved time for contributed papers and posters. We welcome submissions that present at least preliminary results. We are specifically aiming to identify work showing that cognitively-informed models and learning systems outperform standard AI/ML approaches.

We will select based on (1) the depth to which cognitive principles, theories, and models inform the system, and (2) the performance advantage of the cognitively informed system. We encourage submissions making contact with any area of cognition―attention, perception, development, memory, learning from experience, judgment and decision making―which elucidate the computational principles or mechanisms that allow people to outperform machines, and which suggest novel approaches to solving AI challenges such as: flexible and generalizable learning, task-dependent information acquisition and processing, avoidance of catastrophic forgetting, and operating subject to energy (computational efficiency) constraints.

We prefer brief submissions of up to four pages, excluding references, formatted in NIPS style. No need to anonymize submissions. If you have a longer manuscript already submitted and under review, you may submit the manuscript instead. Accepted submissions will be posted on the workshop page if the authors wish, but otherwise the submissions will be used only for reviewing contributions.

Submit your contribution (in PDF format) to cognitivelyinformedAI@gmail.com (indicate whether you have already registered for the workshops). Feel free to contact the organizers if you have questions about the relevance of your research for the workshop.