Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Decomposing stimulus-specific sensory neural information via diffusion models
Authors: Steeve Laquitaine, Simone Azeglio, Carlo Paris, Ulisse Ferrari, Matthew Chalk
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Applied to a model of visual neurons, our method quantifies how specific stimuli and features contribute to encoded information. Our approach provides a scalable, interpretable tool for probing representations in both biological and artificial neural systems. We demonstrate the power of our method by quantifying the information encoded by a model of visual neurons about individual images and pixels. |
| Researcher Affiliation | Academia | Steeve Laquitaine Institut de la Vision, INSERM, CNRS Sorbonne Université 17 Rue Moreau, Paris 75012 Simone Azeglio Institut de la Vision, INSERM, CNRS, & Laboratoire des Systèmes Perceptifs Sorbonne University & École Normale Supérieure PSL 17 Rue Moreau, Paris 75012 & 29 Rue d Ulm Paris 75005 Carlo Paris Institut de la Vision, INSERM, CNRS Sorbonne Université 17 Rue Moreau, Paris 75012 Ulisse Ferrari Institut de la Vision, INSERM, CNRS Sorbonne Université 17 Rue Moreau, Paris 75012 Matthew Chalk Institut de la Vision, INSERM, CNRS Sorbonne Université 17 Rue Moreau, Paris 75012 EMAIL |
| Pseudocode | No | The paper provides detailed mathematical derivations and descriptions of methods, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps in a code-like format. |
| Open Source Code | Yes | We are sharing our code on an anonymised github repository, here: https://anonymous.4open.science/r/neural-info-decompo-5BAF/. |
| Open Datasets | Yes | We utilized the MNIST dataset [19], which consists of 60,000 grayscale images of handwritten digits for training and 10,000 for testing. To demonstrate that our approach generalizes to more complex and biologically realistic neural encoder models as well as natural images, we trained the diffusion models on 60,000 64 64 natural images from Hugging Face s Tiny Image Net dataset |
| Dataset Splits | Yes | We utilized the MNIST dataset [19], which consists of 60,000 grayscale images of handwritten digits for training and 10,000 for testing. |
| Hardware Specification | Yes | On average, approximating Ilocal(x) for one image took approximately 1 minute on an NVIDIA Ge Force RTX 4080 GPU. Both models were implemented in Py Torch (version 2.7.0) with Cuda 12.6 and trained using the Hugging Face Diffusers library [27] and Accelerate library [11] for distributed training on 7 NVIDIA A100 GPUs (40 GB VRAM). |
| Software Dependencies | Yes | Both models were implemented in Py Torch (version 2.7.0) with Cuda 12.6 and trained using the Hugging Face Diffusers library [27] and Accelerate library [11] for distributed training on 7 NVIDIA A100 GPUs (40 GB VRAM). |
| Experiment Setup | Yes | Optimizer: Adam W with learning rate 1 10 4, β1 = 0.9, β2 = 0.999 Learning rate scheduler: Cosine schedule with warmup (500 steps) Batch size: 256 Training duration: 150 epochs Loss function: Mean squared error (MSE) between predicted and true noise Precision: Mixed precision training with bfloat16 Training Time: 25/30 minutes per model We used a linear noise schedule with βt increasing from 1 10 4 to 0.02 over 1,000 time steps. |