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..
Convolution Goes Higher-Order: A Biologically Inspired Mechanism Empowers Image Classification
Authors: Simone Azeglio, Olivier Marre, Peter Neri, Ulisse Ferrari
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive evaluation on standard benchmarks and synthetic datasets, we demonstrate that our architecture consistently outperforms traditional CNN baselines, achieving optimal performance with 3rd/4th order expansions. |
| Researcher Affiliation | Academia | Simone Azeglio Institut de la Vision & Laboratoire des Systèmes Perceptifs INSERM, CNRS, Sorbonne University & École Normale Supérieure PSL 17 Rue Moreau, Paris 75012 & 29 Rue d Ulm Paris 75005 EMAIL Olivier Marre Institut de la Vision INSERM, CNRS, Sorbonne University 17 Rue Moreau, Paris 75012 Peter Neri CHT Erzelli & Laboratoire des Systèmes Perceptifs IIT & École Normale Supérieure PSL Via Enrico Melen 83, Genova 16152 & 29 Rue d Ulm Paris 75005 Ulisse Ferrari Institut de la Vision INSERM, CNRS, Sorbonne University 17 Rue Moreau, Paris 75012 |
| Pseudocode | No | The paper describes the higher-order convolution operation using mathematical equations and descriptive text, supplemented by architectural diagrams. It does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://github.com/sazio/Higher Order Conv. |
| Open Datasets | Yes | We evaluated our models on a range of datasets, beginning with synthetic images... Subsequently, we extended our experiments to compare our model against CNN baselines on widely-used datasets including MNIST, Fashion MNIST, CIFAR10, CIFAR100, Imagenette and Imagenet. IMAGENET To investigate the scalability and real-world applicability of our approach, we implemented Higherorder Convolution in a Res Net-18 [4] architecture (Ho Res Net-18) and evaluated it on Image Net [40]. |
| Dataset Splits | Yes | Our synthetic dataset comprises 2000 training images, 1000 validation images, and 2000 testing images. All reported results are based on the test set. (A.9.2 Training Setup: Imagenette) Training images: Random resized crop to 224 224, random horizontal flip, normalization Test images: Resize to 256 256, center crop to 224 224, normalization |
| Hardware Specification | Yes | Experiments used an NVIDIA RTX 4080 Ti GPU, with training times of 10-15 minutes for MNIST/Fashion MNIST and 20-30 minutes for CIFAR-10/CIFAR-100 (Ho CNN requiring 1.5-2 longer). The Image Net experiments were run on an NVIDIA A100 GPU, requiring approximately 2-3 days per run. |
| Software Dependencies | No | For all experiments, we used Adam W optimizer [39] with learning rate 0.001, weight decay 5e-4, batch size 64, cross-entropy loss and early stopping. |
| Experiment Setup | Yes | For all experiments, we used Adam W optimizer [39] with learning rate 0.001, weight decay 5e-4, batch size 64, cross-entropy loss and early stopping. Images were normalized using z-score standardization without data augmentation. (A.9.2 Training Setup: Imagenette) Batch size: 64 Loss function: Cross-entropy Optimizer: Adam W with learning rate 0.001 and weight decay 5e-4 Learning rate scheduling: Reduce LROn Plateau (halves LR after 5 epochs without improvement) Early stopping: Implemented with 12 epochs patience |