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..
T-REGS: Minimum Spanning Tree Regularization for Self-Supervised Learning
Authors: Julie Mordacq, David Loiseaux, Vicky Kalogeiton, Steve OUDOT
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Several experiments on synthetic data and on classical SSL benchmarks validate the effectiveness of our approach at enhancing representation quality. |
| Researcher Affiliation | Academia | Julie Mordacq1,2 David Loiseaux1,2 Vicky Kalogeiton2 Steve Oudot1,2 1 Inria Saclay 2 LIX, CNRS, École Polytechnique, IP Paris |
| Pseudocode | Yes | Algorithm 1 T-REGS combined with view invariance using Py Torch pseudocode |
| Open Source Code | No | The full code will be made available. The experiments use standard datasets (Image Net and CIFAR) which are publicly available. |
| Open Datasets | Yes | Our experiments are performed on 1. Image Net dataset [18], and a subset Image Net-100 which are subject to the Image Net terms of access 2. CIFAR-10, CIFAR-100 |
| Dataset Splits | Yes | We evaluate the representations obtained after training with T-REGS, either directly combined with view invariance or integrated with existing methods (i.e., BYOL, and Barlow Twins) on CIFAR-10/100 [37], Image Net-100 [59], and Image Net [18]. Our implementation is based on solo-learn [16], and we use torchph [7] for the MST computations. For T-REGS as a standalone regularizer, we use β = 10, γ = 0.2, λ = 8e 4. ... We evaluate our model on Image Net-100 and Image Net-1k using Res Net-18 and Res Net-50, respectively, following the standard linear evaluation protocol on Image Net and comparing with the state of the art. |
| Hardware Specification | Yes | We conducted our experiments using NVIDIA H100 and V100 GPUs. |
| Software Dependencies | No | Our implementation is based on solo-learn [16], and we use torchph [7] for the MST computations. ... To compute the length of the minimum spanning tree, we rely on torchph and Gudhi [50], both released under MIT Licenses. |
| Experiment Setup | Yes | For T-REGS as a standalone regularizer, we use β = 10, γ = 0.2, λ = 8e 4. |