Measuring what neural networks encode
Principled measures, scaled with diffusion models, of which inputs and features a population encodes.
Steeve Laquitaine, PhD
Research Scientist · Neural representations, information theory and inference
I develop mathematical and computational methods to understand what neural networks represent and how they use uncertain information.
Research scientist, Paris Vision Institute (Sorbonne Université) · Affiliate, Stanford University · Previously EPFL Blue Brain Project, Stanford and RIKEN, and Lead Data Scientist at Capgemini Engineering
NeurIPS 2026 · Oral
A multi-scale information geometry of neural population codes
Co-author
NeurIPS 2025
Decomposing neural information with diffusion models
Co-first author
Nature Communications
Spike sorting biases and information loss (accepted)
First and corresponding author
Neuron 2018
A switching observer for human perceptual estimation
First author
Selected research
I look for cases where a convincing measurement supports the wrong explanation, then build the method and the experiment that tell the explanations apart.
Telling apart mechanisms behind near-optimal behaviour
Human estimates matched Bayesian averages, yet their full distributions revealed a switching strategy.
Evaluating spike sorting against ground truth
With every spike known, a detailed cortical simulation shows that sorting biases nearly halve stimulus information.
Research influence
-
Chetverikov & JeheeNature Communications, 2023Test the switching observer as an alternative account of bimodal motion-direction reports.
-
Bévalot & MeynielCommunications Psychology, 2024Cite it when noting that quantifying how Bayesian a computation is requires more model comparisons.
-
Garcia et al.eLife, 2026Identify full-network simulations like ours as a source of complete ground truth for spike-sorting benchmarks.
Also cited in Trends in Cognitive Sciences, Nature Neuroscience, Nature Human Behaviour, Nature Reviews Neuroscience, PNAS and eLife. Google Scholar
Industry experience
From research to production
Lead Data Scientist, Capgemini Engineering
- Supervised four teams for three years through the development and deployment of machine-learning software in production.
- Designed and taught a course on reproducible, maintainable machine-learning engineering for Société Générale practitioners (2020–2021).
Research directions
I aim to extend this work toward understanding how learned representations support inference under changing conditions: when systems combine prior knowledge with new evidence, when they rely on shortcuts, and which experiments distinguish these strategies.