Noumenal Labs

Designing machine intelligence that can reason by analogy Inspired by neural reuse and reassembly, we consider key design features that unlock the ability for AI to reason by analogy and by metaphor

New blog post by @noumenal-labs.bsky.social : “Designing machine intelligence that can reason by analogy”:
www.noumenal.ai/post/designi...

Dynamic Markov Blanket Detection for Macroscopic Physics Discovery The free energy principle (FEP), along with the associated constructs of Markov blankets and ontological potentials, have recently been presented as the core components of a generalized modeling metho...

I’m happy to share the new @noumenal-labs.bsky.social preprint, “Dynamic Markov blanket detection for macroscopic physics discovery,” by Jeff Beck and me
arxiv.org/abs/2502.21217
Here’s a thread!
1/16

Noumenal Labs’s approach is to build models that are explicitly structured — in a manner that is consistent both with the means by which we intuitively understand and conceptualize the world, and the means by which our best scientific models describe it 7/7

Noumenal Labs is developing a new approach to macroscopic physics discovery, enabling machine intelligences to discover the different objects/concepts and object/concept types that generate time series data, as well as the type-specific rules that govern their behavior directly from data 6/7

Generically, these models lack the requisite structure to represent the world as we understand it, i.e., as composed of distinct objects and object types, as well as object-type-specific interactions, unless hand-specified by the user 5/7

This gap in performance is largely due to the relative simplicity of the generative models, underlying state space, and relationships typically used in the active inference literature 4/7

While theoretically well grounded, active inference has yet to yield state of the art performance in practice 3/7

Active inference is a form of Bayesian machine learning that replaces the objective function used in traditional machine learning — the cost or reward function — with an information theoretic surprise minimization objective, which is used for both inference and learning 2/7

Filling the gaps in active inference Here we discuss key gaps in SOTA applications of active inference in AI - and how Noumenal Labs is working to fill them.

New blog post by @noumenal_labs: “Filling the gaps in active inference”: www.noumenal.ai/post/filling...

Here’s the tl;dr:

In this blog post, we discuss key gaps in state-of-the-art applications of active inference in artificial intelligence — and how Noumenal Labs is working to fill them 1/7

Why AI should reason like a scientist by Noumenal LabsThe tl;dr • Generative AI and large language models (LLMs) feature feedforward architectures based on the diffusion model and on the transformer architecture. They have yielded excitin...

We discuss data prediction versus data explanation in the design of artificial intelligence — and why we think AI agents should be designed to behave like curious scientists: www.noumenal.ai/post/why-ai-...

First blog post by Noumenal Labs on the “DeepSeek moment”: www.noumenal.ai/post/the-dee...

If you have the skills and interest to join us, please apply on our website or directly here: jobs.gusto.com/postings/nou...
If you have any questions, please email: talent at noumenal dot ai 5/5

State space modeling
Maximum entropy inverse reinforcement learning
Structured 4D Gaussian splatting
Expert-level programming in Python and JAX 4/5

Energy-based models for inference and learning
Bayesian machine learning
Probabilistic programming languages
Neural partial differential equations
Graph neural networks 3/5

We are expanding our world-class research and development team to build the next generation of machine intelligence. We are looking for candidates with with expertise in the following domains: 2/5

Now hiring! Introducing Noumenal Labs (www.noumenal.ai), a new deep tech company focusing on multiscale, multimodal time series modeling, object-centered causal physics discovery, continual learning, and energy-based architectures 1/5