ActInf GuestStream 065.1 ~ Phoebe Klett, "Towards Bayesian World Models"

Abstract At Normal Computing, we believe that achieving System-2 thinking in
artificial systems will necessarily involve going beyond the auto-regressive
language models we all know and love. Leveraging said systems in high-stakes
applications will require step changes in explainability and reliability. The
foundations of probabilistic machine learning and Bayesian decision theory
provide rich tool-kits to augment today’s powerful and unwieldy LLMs. In this
talk, we discuss the Bayesian world models we’re developing, highlighting
connections to the Free Energy Principle. In particular, we discuss using LLMs
as likelihood machines, hierarchical world models from message passing, and
using world models as recommendation engines. Normal Computing:
https://normalcomputing.ai/ “When it comes to robust reasoning, an Achilles’
heel of current large language models is that the world model and the
inference machine are one and the same.” \- Yoshua Bengio [BH23] “I submit
that devising learning paradigms and architectures that would allow machines
to learn world models in an unsupervised (or self-supervised) fashion, and to
use those models to predict, to reason, and to plan is one of the main
challenges of AI and ML today. One major technical hurdle is how to devise
trainable world models that can deal with complex uncertainty in the
predictions.” \- Yann LeCun [LeC22] [BH23] Yoshua Bengio and Edward Hu.
Scaling in the service of reasoning model-based ml. 2023. [LeC22] Yann LeCun.
A path towards autonomous machine intelligence. 2022. Active Inference
Institute information: Website: https://activeinference.org/ Twitter: /
inferenceactive Discord: / discord YouTube: / activeinference Active Inference
Livestreams: https://coda.io/@active-inference-ins...

CSID: 12c18c8e6be51f20

Content Managed by ContentSafe.co