Inferring Inference
summary
The gist
Patterns of microcircuitry in the cerebral cortex suggest that the brain has an array of repeated elementary or “canonical” computational units.
In short
The framework infers canonical distributed computations from large neural activity patterns by combining normative and algorithmic theories of computation. It aims to simultaneously find the neural representation of latent variables, their interactions defining a world model, and the specific message-passing functions that govern the brain's inference algorithm.
Key concepts
- Message-Passing Dynamics
- This describes how complex computations are broken down into simple local steps across a network. It models neural activity as nodes sending messages to neighbors using fixed rules (message function, aggregation, update function). This structure is key because it assumes brain computations follow these shared dynamics.
- Hidden Markov Model (HMM)
- This probabilistic model is used to link measured neural activity to the underlying latent variables. The HMM helps find the best combination of neural encoding, world interactions, and message functions that most likely generated the observed brain signals.
- Canonical Message-Passing
- This hypothesis posits that brain computations rely on shared or 'canonical' operations across different parts of the network. This constraint allows researchers to statistically distinguish these specific types of computations from other possible neural models.
Terminology used across episodes
This episode discusses
- Inferring Inference · Paper Radio
- Dynamic Routing Between Capsules
- Modularity benefits reinforcement learning agents with competing homeostatic drives
- Neural Relational Inference for Interacting Systems
- Relational inductive biases, deep learning, and graph networks
- Inference in Probabilistic Graphical Models by Graph Neural Networks
- Generalization of graph network inferences in higher-order graphical models
- GraphLIME: Local Interpretable Model Explanations for Graph Neural Networks
- Physics-informed dynamic mode decomposition (piDMD)
- Capsule Network Performance on Complex Data
- Recurrent Independent Mechanisms
- Discovering Governing Equations from Partial Measurements with Deep Delay Autoencoders
- Sparse Identification of Nonlinear Dynamics with Control (SINDYc)
- Shallow Univariate ReLu Networks as Splines: Initialization, Loss Surface, Hessian, & Gradient Flow Dynamics
- Black box variational inference for state space models
- A Survey on The Expressive Power of Graph Neural Networks
- On the Principles of Parsimony and Self-Consistency for the Emergence of Intelligence
- Adam: A Method for Stochastic Optimization
The paper
Inferring Inference · Read on arXiv
Rajkumar Vasudeva Raju, Zhe Li, Scott Linderman, ×aq Pitkow
Department of ECE, Rice University · Department of Neuroscience, Baylor College of Medicine · Department of Statistics, Stanford University · ×Neuroscience Institute, Carnegie Mellon University · Department of Machine Learning, Carnegie Mellon University · Center for Neuroscience and Artificial Intelligence, Baylor College of Medicine
Patterns of microcircuitry in the cerebral cortex suggest that the brain has an array of repeated canonical computational units. However, neural representations are distributed, so the relevant computations may only be related indirectly to single-neuron transformations. It thus remains an open challenge how to define canonical distributed computations. Here we integrate normative and algorithmic theories of neural computation to present a mathematical framework for inferring canonical distributed computations from large-scale neural activity patterns. At the normative level, we hypothesize that the brain learns a structured internal model of its environment and its latent causes. At the algorithmic level, we propose that this inference process is a nonlinear message-passing algorithm on a graph-structured model of the world. Given a time series of neural activity during a perceptual inference task, our analysis framework finds (i) the neural representation of the relevant latent variables, (ii) interactions between these latent variables that define the brain's internal model of the world, and (iii) message functions that specify the inference algorithm. With enough data, key computational properties are identifiable due to constraints from the symmetries inherent in any canonical computation, up to a joint global transformation of all interactions and the message-passing functions. As a demonstration of this framework, we analyze artificial neural recordings generated by a simulated brain that implicitly implements an approximate inference algorithm. Given simulated inputs and neural activity we successfully recover the latent variables, their neural representation and dynamics, and canonical message functions that govern the dynamics. Analysis of these models reveals features of experimental design required to extract canonical computations from neural data.
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We're looking at the paper: "Inferring Inference".
Marcus: Here's the gist of the paper. Patterns of microcircuitry in the cerebral cortex suggest that the brain has an array of repeated elementary or “canonical” computational units.
Ines: Here is a mathematical framework for inferring canonical distributed computations from large-scale neural activity patterns.
Marcus: Conceptual Framework and Hypotheses The paper proposes a framework that integrates normative and algorithmic theories of neural computation to infer canonical distributed computations from large-scale neural activity patterns.
Ines: At the normative level, the hypothesis posits that the brain creates a structured internal model of its environment, positing latent causes that explain its sensory inputs, and using those sensory inputs to infer the states of the latent causes.
Title and authors: Ines: Let’s talk a bit about the title and who wrote this paper, "Inferring Inference". It immediately tells you the goal is inference from inference. It’s not just observing activity; it’s trying to reverse-engineer the brain's reasoning process.
Marcus: And it brings together these two theoretical approaches we mentioned earlier, normative and algorithmic theories of neural computation. That's a lot of conceptual heavy lifting right there for a framework that aims to pull out actual computational motifs from data.
Yuki: I wonder if this "inference" they’re talking about is just statistical inference, or if it’s something more like genuine cognitive reasoning happening in the system.
Ines: The paper suggests they are aiming for both—the normative level hypothesizes that the brain creates a structured internal model of its environment, positing latent causes that explain sensory inputs, and then uses those inputs to infer the states of those latent causes.
Marcus: And at the algorithmic level, they frame this inference as a nonlinear message-passing algorithm on a graph-structured model of the world. It’s trying to treat the brain's operation like a specific type of communication process across interconnected parts.
Yuki: So it’s proposing that we can map complex neural patterns onto these simple, structured message-passing dynamics, which sounds like a big step toward making sense of distributed cognition.
Ines: Right. The core idea is finding the neural representation of those latent variables and the interactions between them that define the brain’s internal model.
Marcus: And they also want to discover those message-functions, which are essentially the specific rules for how information flows and gets processed in this inference algorithm.
The paper's summary: Ines: So, when we look at the actual summary of "Inferring Inference", what they’re saying is that they developed a structured statistical model to fit neural data and reveal these distributed computational motifs. It’s about moving beyond just looking at local transformations in neurons.
Marcus: They are focusing on how these latent variables interact, and they want to see if the dynamics of those interactions match what a message-passing algorithm predicts across different brain regions or neuronal populations.
Yuki: It sounds like they are trying to find the underlying probabilistic graphical model that governs how the brain updates its beliefs about the world state over time. That’s a deep connection to population genetics, thinking about how information propagates through a network of interacting entities.
Ines: They set up this framework where you have an encoding function for neural activity, coupling parameters defining interactions, and message-passing functions specifying the algorithm itself. All these parts are parameterized into one big vector theta that they want to fit to the measured neural activity.
Marcus: The inference procedure itself uses a Particle EM algorithm to get the maximum likelihood estimate of these TAP brain parameters. They use a particle filter in an E-step and gradient ascent on the Q function in an M-step.
Yuki: That sounds like they are tackling a very complex estimation problem, trying to find the posterior distribution of those latent world states given the observations. It’s ambitious for any single experiment to do that.
Ines: The paper shows how this procedure works through a schematic: you have time k, encoding k, true world state wt, observations ot at time t, and then the probabilistic model D and message-passing algorithm are used to infer the brain’s posterior over the assumed world state s >.
The paper's improvements: Marcus: Now that we know what they did, there are some suggestions for how this framework can be improved. They point out that their current assumption is just canonical message-passing on graphs, which is a bit too restrictive for real brain dynamics.
Ines: They suggest generalizing to richer classes of graphical models, specifically accounting for continuous variables and directed acyclic graphs. This would allow them to capture causal structure more accurately than just undirected connections.
Yuki: Causal structure is key here; if you can model the world as a directed graph, you’re moving closer to understanding how actions cause effects in the environment, which is vital for evolutionary thinking.
Marcus: They also propose augmenting this framework to account for world dynamics by allowing an internal model based on a spatiotemporal graph. That would mean modeling how the environment itself evolves over time, not just assuming static world states.
Ines: The authors also mention that the core of their work is finding canonical operations—shared message-passing functions across the graph. They suggest future work should compare this model against other models that don't rely on these shared functions to really establish what makes a computation canonical.
Yuki: So, they are looking at testing if their specific assumption about shared operations actually yields something distinct from other possible ways the brain could be computing things. That’s a good way to test the necessity of their model structure.
Conclusion: Ines: To wrap this up, "Inferring Inference" suggests that by integrating normative and algorithmic theories, we can uncover latent causes and world models within neural activity patterns using message-passing dynamics as the mechanism for inference.
Marcus: The paper shows a path forward through their Particle EM algorithm to estimate these brain parameters, linking the neural encoding to the underlying probabilistic model of reality.
Yuki: For me, this work opens up a way to look at the brain not just as a processor, but as an active system that builds and tests models of its surroundings over time.
Ines: And it gives us tools to test if those internal world models are consistent with the actual sensory inputs we observe. It’s about moving from correlation to causal modeling of cognition.
Marcus: The authors flag that their current model assumes canonical message-passing on graphs, and they suggest generalizing this to directed acyclic graphs and spatiotemporal world dynamics for future work.
Yuki: So the paper provides a concrete mathematical structure, but the next step is testing how flexible that structure is when we introduce more realistic complexities like causality and temporal evolution.
Ines: And that’s where this leaves us, looking at the full scope of what these systems can actually model. That’s what "Inferring Inference" sets up for everyone else working in this area.
More episodes
- 2610.33881-Sensorimotor behavior from an untrained brain-and-nerve-cord connectome
- 2610.40344-Morphological Features of Murine Pulmonary Arterial Networks for Diagnosing Vascular Disease
- Daily Summary for 2026-11-19
- 2609.10193-Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
- 2610.38236-Generative discovery of chemical reaction networks
- 2507.11176-An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2504.03590-Towards a Unified Framework for Determining Conformational Ensembles of Disordered Proteins
- 2603.15217-A multiscale discrete-to-continuum framework for structured population models