Connectionists: RS meeting on beyond symbols and signals

[Peter Dayan]

Royal Society Discussion Meeting entitled Beyond the Symbols vs Signals
Debate, which will take place on 28-29th October 2024 at the Royal
Society in London:

https://royalsociety.org/science-events-and-lectures/2024/10/symbols-vs-signals/

There will be a poster session on Monday 28 October 2024. Attendees can
apply at the above link by 17th September 2024 to present a poster.

Building artificial intelligence systems that can emulate human
intelligence will need to draw on the complementary strengths of machine
learning and compositional reasoning, and take advantage of symbolic
knowledge as well as raw signals. This meeting aims to advance our
scientific understanding of how human learning, reasoning and cognition
can be brought to bear on the engineering foundations of AI.

The rapid progress of artificial intelligence in the past decade is
founded on deep learning, which excels at learning statistical patterns
from raw signal data. Neural networks are loosely based on neural
processes in the brain but the way they learn, requiring large amounts
of training data and relying on parameter optimisation via gradient
descent, is unlike the way the human brain learns, with humans being
able to learn effectively from few examples and utilise prior knowledge
in a flexible and adaptive manner. Conventional symbolic reasoning
founded on logic, on the other hand, has the advantage of being able to
reason compositionally, with robustness and rigorous guarantees, but is
often not compatible with raw data and lacks scalability. The
complementarity of the two approaches has long been recognised, leading
to scientific debates about their respective strengths and weaknesses
that featured under multiple guises: signals vs symbols, neural vs
symbolic, and statistical vs reasoning. Elements of this debate also
appear in the psychological literature on dual systems, such as the
notion of “System 1” and “System 2” popularised by Kahneman. Aspects
that are actively discussed in the AI community include whether either
approach is enough on its own to emulate human intelligence, and, if
not, how they can be combined to the best of their advantage. At the
same time, emerging developments in experimental neuroscience and
progress in computational modelling of human cognition offer valuable
new perspectives on engineered AI systems.

We believe the time is ripe to move beyond the signals vs symbols debate
and, building on the tremendous success of deep learning and recent
advances in experimental and computational cognitive science, address
fundamental questions about the computational nature of human
intelligence that can, in turn, provide sound scientific foundations for
methodologies and software technologies that are capable of engineering
human-like AI.

Several distinguished speakers from across the scientific spectrum have
agreed to present their views to stimulate the discussions.

This meeting is free to attend and is intended for researchers in
relevant fields. Advance registration is essential. In-person and online
attendance is available.  We hope to see you in London!

Marta Kwiatkowska FRS, University of Oxford
Doina Precup, McGill and Google
Peter Dayan, Max Planck Institute for Biological Cybernetics
Tom Griffiths, Princeton University

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/x-pkcs7-signature
Size: 5505 bytes
Desc: not available
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20240905/ecc9b009/attachment.bin>