Systems Neuroscience Approach to General Intelligence (SynAGI)

Systems Neuroscience Approach to General Intelligence (SynAGI) - AAAI 2023 Workshop

AAAI Website Link: https://aaai.org/Conferences/AAAI-23/ws23workshops/#ws29

Dates and Times:

Day 1 — Mon. Feb. 13
Morning Session : 8:30 AM-12:30 PM
Afternoon Session : 2:00 PM - 6:00 PM

Day 2 — Tues. Feb. 14
Morning Session : 8:30 AM-12:30 PM
Afternoon Session : 2:00 PM - 6:00 PM

Workshop Title: Systems Neuroscience and Deep Learning Approach to General Intelligence

Discussion: Early in the study of AI, researchers believed the task of emulating a human brain might be feasible because it included a constructive proof. Just as people thousands of years ago believed engineered human flight would one day be possible because of birds, Norbert Weiner and others believed machine-based thought would be feasible because of brains. This feasibility has been elusive, just as the mythical Daedalus never actually created a flying machine (although in AD 875, Abbas Ibn Firnas created and flew a real glider). Of course, science and engineering progressed to the era of the Wright brothers (ca. 1903), and today, AI technology and Neuroscience have progressed such that it’s again prudent to look to the brain as a model for AI. By examining the broader picture and overlap of current artificial neural networks, theoretical computer science and systems neuroscience, this workshop will uncover important gaps in our current knowledge about the brain and models of intelligence and thereby advance both fields.

Bernard Baars modeled the brain and its cognitive processes as a Global Workspace — a theater where content is globally broadcast from the stage to an audience of many powerful specialized processors. This model was then elaborated in terms of network neuroscience as the Global Neuronal Workspace (GNW) by Jean-Pierre Changeux, Stanislas Dehaene and others, and more recently in the terms of theoretical computer science as the Conscious Turing Machine (CTM) [1]. The CTM is a substrate independent model for consciousness. It was inspired both by the Global Workspace model model and by Alan Turing’s simple yet powerful model of computation. We believe it also provides an excellent departure from past neuroscience models of the Global Workspace in that it includes a specific architecture that can guide analysis of global brain anatomy. Progress in the GNW has aimed to incorporate new mechanisms that remain to be translated into machine learning, such as oscillations and neurotransmission [2]. Some AI researchers have proposed variations and extensions of the Global Workspace, implementing the theoretical CTM in much the way early computers implemented the theoretical Turing Machine, or connecting the CTM to advances in neural nets like Transformers and VAEs. In related work, Transformer- [3] or VAE-based [4] architectures have been proposed, though they lack the CTM’s specialized processors.

Meanwhile, neuroscience is discovering global circuits in the brain that bear a striking resemblance to the patterns found in contemporary AI architectures such as Transformers. The idea of using global brain anatomy to inspire better AI systems is novel. This workshop will aim to map GWT to AI systems via the brain’s global architecture. We hypothesize that the actions of AI models of brain regions when coupled by a brain-like architecture and implementing something like the CTM can achieve general intelligence. High resolution recordings from global brain circuits are now feasible. These recordings can be used to validate such models.

The goal of this workshop is to bring together a multi-disciplinary group comprising AI researchers, systems neuroscientists, algorithmic information theorists, and physicists to compare notes and understand gaps in the larger agenda. Our goal is that gaps identified by some but not all are communicated, and the ability to span gaps by some disciplines but not others discussed. Our goal is to determine what we know about what’s needed to build a thinking machine in the same way that the Wright brothers built a flying machine.

References:

[1] https://doi.org/10.1073/pnas.2115934119

[2] https://doi.org/10.1101/2022.01.24.477526

[3] https://arxiv.org/abs/2103.01197

[4] https://doi.org/10.1016/j.tins.2021.04.005

Speakers will include Turing Award winners, members of the National Academy and Royal Society, a Canada Research Chair holder, and a winner of the Presidential Award for Excellence

Workshop Schedule:

Day 1 – Morning Session – Mon. Feb. 13 — "Towards generalized principles of computation in the mind"

Moderated by James Kozloski, IBM Research

Start	End	Speaker/Activity	Title/Description	Q&A	Link
8:30	9:00	Welcome and Introductions, James Kozloski, IBM Research	Systems Neuroscience Approach to General Intelligence
9:00	9:45	Professors Lenore and Manuel Blum, CMU emeriti, U.C. Berkeley, Peking U.	Insights from the Conscious Turing Machine (CTM)	:15	R
10:00	10:30	Guillaume Dumas, Associate Professor of Computational Psychiatry (CHUSJ/Mila, University of Montreal); Jean-Pierre Changeux, Emeritus Professor at College de France and Institut Pasteur, Paris France	Multilevel development of cognitive abilities in an artificial neural network	:10	R
10:40	11:00	Break
11:00	11:30	Joscha Bach, PhD., Principal Research Scientist, Intel Labs	Moving beyond the neural network paradigm – minds as self-organizing systems	:10	R
11:40	12:10	Grady Booch, Chief Scientist for Software Engineering, IBM	Cognitive Architectures	:10	R
12:20	12:30	Pre-Lunch Discussion	How do we know if we’re making progress?
12:30	2:00	Lunch

Day 1 –Afternoon Session – Mon. Feb 13 — "Brain-based architectures I."

Moderated by Irina Rish, Mila – Quebec AI Institute

Start	End	Speaker/Activity	Title/Description	Q&A	Link
2:00	2:25	James Whittington, Stanford & Oxford U.	Organizing knowledge for flexible behavior	:05	R
2:30	2:50	Taku Ito, Research Scientist, IBM Research	Multitask Representations in the Human Cortex Transform along a Sensory-to-Motor Hierarchy	:05	R
2:55	3:15	Yangfan Peng, MRC Brain Network Dynamics Unit, U. Oxford (Sharott group)	Directed and acyclic synaptic connectivity in the human layer 2-3 cortical microcircuit	:05	R
3:20	3:40	Break
3:40	3:55	Samuel Schmidgall and Maryam Parsa, George Mason U., Dept. of Electrical and Computer Engineering	Biological connectomes as a representation for the architecture of artificial neural networks	SYD	R
3:55	4:10	Omid Madani, Cisco	Advancing Prediction Games for Learning Networks of Hierarchical Patterns	SYD	R
4:10	4:25	Boris Galitsky, Founder, Knowledge-Trail Inc, San Jose CA	Reasoning and learning of people with autism, control subjects and machines	SYD	R, P
4:25	4:40	Chris Rourk, Citizen Scientist and Patent Attorney	The Hard(ware) Problem of (Machine) Consciousness – Why a special purpose processor and associated processing architecture is needed for cognitive process/conscious action selection mechanisms	SYD	R
4:40	5:10	Arlindo Oliveira, Professor of Computer Science, INESC-ID / Instituto Superior Técnico, Lisbon, Portugal	Connecting metrics for shape-texture knowledge in computer vision	:10	R
5:20	6:00	Workshop Discussion Moderator Murray Campbell, IBM Research	What problem are we solving?

Day 2 – Morning Session – Tue. Feb. 14— "Brain-based architectures II."

Moderated by Lenore Blum, Berkeley U., CMU

Start	End	Speaker/Activity	Title/Description	Q&A	Link
8:30	9:00	Networking
9:00	9:30	James Kozloski, Principal Research Scientist, IBM Research	Decomposition of a Global Brain Circuit for SynAGI Architecture	:10	R
9:40	10:10	Mark Wegman, IBM Fellow, IBM Research	Overview of the SynAGI Architecture	:10	R
10:20	10:40	Break
10:40	11:25	Yoshua Bengio, Professor of computer science, University of Montreal, Mila, IVADO, CIFAR	Towards Implementing Global Workspace Inductive Biases and Probabilistic Reasoning with GFlowNets	:15	R
11:40	12:10	Ryota Kanai, CEO, Araya, Inc.	On the potential link between consciousness and intelligence	:10	R
12:20	12:30	Pre-Lunch Discussion	The collaboration between Neuroscience and AI
12:30	2:00	Lunch

Day 2 –Afternoon Session – Tue. Feb. 14 — "Towards generalized principles of computation in neural circuits"

Moderated by Andrew Sharott, Oxford U.

Start	End	Speaker/Activity	Title/Description	Q&A	Link
2:00	2:45	Wilten Nicola, Assistant Professor, University of Calgary, Hotchkiss Brain Institute	Disk-Drive like operations in the Hippocampus	:10	R
2:55	3:40	Michael Levin,Professor of Biology, Director of the Allen Discovery Allen Center at Tufts University	Neuroscience beyond neurons: the multiscale general intelligence of the body	:10	R
3:50	4:10	Break
4:10	4:40	Konrad Koerding, U. Pennsylvania, Depts. of Neuroscience and Bioengineering, Member of CIFAR LMB group	Scaling up systems identification in Neuroscience towards entire simulated brains	:10
4:50	5:20	Adam Safron, Research Fellow: Center for Psychedelic and Consciousness Research, Johns Hopkins U. School of Medicine and Cognitive Science Program, Indiana U.; Research Consultant, Institute for Advanced Consciousness Studies	On the neurocomputational varieties of conscious experiences: psychedelics as means of understanding and enhancing intelligence and agency in biological and artificial systems	:10	R
5:20	6:00	Panel Joscha Bach, Yoshua Bengio, Lenore Blum, Guillaume Dumas, Ryota Kanai, and James Kozloski Moderator Irina Rish, Mila – Quebec AI Institute	Next Steps for SynAGI

R – Speaker recommended reading

P – Recommended reading purchase

Speaker Bios:

Issues To Be Covered:

Plausible comparisons between existing AI architectures and brain regions
Architectures for coupling AI components according to brain anatomy and physiology
Micro-benchmarks to determine whether those architectures achieve their goals
Validation of synthetic AI brain models against large-scale brain recordings

Organizing Committee:

Co-chair: Mark Wegman, ACM, IEEE and IBM Fellow, Member of NAE. Expertise information theory, algorithms. wegman@us.ibm.com.

Co-chair: James Kozloski, IBM Master Inventor, Principal Research Scientist. Expertise computational neuroscience, neuroanatomy, and simulation. kozloski@us.ibm.com.

Lenore Blum, Distinguished-Professor-in-Residence EECS, Berkeley; Distinguished Career Professor of Computer Science, Emerita, CMU; President, Association for Mathematical Consciousness Science (AMCS). Co-author of the CTM. Expertise, theoretical CS, both discrete and real, and theories of consciousness. lblum@cs.cmu.edu.

Irina Rish, Professor, Computer Science and Operations Research, Université de Montréal (UdeM), Mila – Quebec AI Institute. Canada CIFAR AI Chair and the Canadian Excellence Research Chair in Autonomous AI. irina.rish@mila.quebec.

Andrew Sharott, Associate Professor, Oxford University Medical Research Council, Brain Network Dynamics Unit. Expertise neuroscience, global brain recordings. andrew.sharott@bndu.ox.ac.uk.

Format and Logistics: The workshop will take place over 2 days, and the organizing committee will select 40 presenters/participants. These include special guests for invited talks, and applicants who submit an up to two-page vision paper, or a previously published paper, on the workshop topic.

To join the workshop please register for it through the AAAI website (https://aaai.org/Conferences/AAAI-23/registration) and for our planning purposes, provide us a bit more information about your research and interests here: https://docs.google.com/forms/d/e/1FAIpQLSf5xuuqUggzeNesfJ8Km-rpj__NLhhojQXspxAK446KtCbr6g/viewform. You may want to conduct some prereading from the links provided by each speaker on the schedule, and consider some of the issues raised here, then add to these points during discussion, or you may want to explain why the points of view expressed need to be modified in a fundamental way. You are welcome to point to some of your published work during this workshop.

Most participants will be expected to join the workshop, and we will limit participation so that we can form a group that has good discussion and hopefully plan joint work. In exceptional cases, participation by remote access will be possible. Within 2 months of the workshop, participants should have submitted follow up articles, which we anticipate “publishing” in a forum to be determined. We might produce a video summary of some of the proceedings.

Speaker-Recommended Reading

https://doi.org/10.1073/pnas.2115934119

https://doi.org/10.3389/fnana.2016.00003

https://www.biorxiv.org/content/10.1101/2022.10.05.511000v1

https://arxiv.org/abs/2112.04035

Upcoming events of the Synagi group

AAAI Workshop: Systems Neuroscience and Deep Learning Approach to General Intelligence

Systems Neuroscience Approach to General Intelligence (SynAGI) gen_feedback_link(leftr, rightr);

links