Theory & Data Analysis Team

Scientific Overview

The Theory & Data Analysis team is the core "hub" team in SCENE's hub-and-spokes model, formed by merging the former Theory and Data Science teams. Its mission is twofold: to develop the mathematical and computational framework that unifies ecological neuroscience across species, tasks, and brain areas, and to handle, interpret, and openly share the large-scale, multi-modal datasets produced across SCENE's experimental spokes.

SCENE's theoretical approach is grounded in Partially Observable Markov Decision Processes (POMDPs) — the natural language to mathematically formalise the closed loop between sensory observations, subjective beliefs implemented via neural dynamics, actions, and the environment. In this framework, animals use their observations to guide a sequence of decisions or actions to maximise their total expected value, even when the current world state must be inferred from partial observations.

The team is formalising SCENE's three overarching hypotheses into the language of POMDPs:

H0 — Encode all causal latent variables: Internal models support flexible prediction, generalisation, and decision-making, creating action- and value-agnostic representations.
H1 — Encode controllable latent variables (affordances): Brains may prioritise representing the opportunities for action that the environment provides, generalising Gibson's categorical "affordance" into a continuous measure of "controllability."
H2 — Encode reward-predictive variables: Organisms form direct connections between stimuli, actions, and rewards without necessarily constructing an internal model of the world.

These hypotheses are not mutually exclusive — they represent orthogonal axes in the space of all possible internal representations.

On the data side, the team develops robust pipelines for data processing, analysis, and visualisation, enabling all experimental teams to work with standardised tools and reproducible methods. This includes machine learning approaches for pose estimation (building on tools such as DeepLabCut), dimensionality reduction, and inferring latent neural dynamics from sparse spiking observations. A core objective is building cloud-based storage solutions for seamless data sharing and standardised frameworks to ensure reproducibility across studies and laboratories.

Key Questions

What principles dictate how H0-like, H1-like, and H2-like a state representation and algorithm will be, given an organism's ecology, computational challenges, and bodily constraints?
How does splitting world-states and belief-states into controllable and uncontrollable components impact an agent's ability to perform a given task?
Can we derive, for a given neural architecture, the optimal balance between optimising for information (H0), reward (H2), or empowerment (H1)?
How do different brain areas occupy different positions along the H0, H1, and H2 axes?
How can we build unified data pipelines that handle multi-species, multi-modality neural and behavioural recordings?
What machine learning approaches best capture the latent dynamics underlying ecological behaviour?
How do we standardise data formats and analysis tools to ensure reproducibility across laboratories and species?