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..
Contrastive Self-Supervised Learning As Neural Manifold Packing
Authors: Guanming Zhang, David Heeger, Stefano Martiniani
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | CLAMP achieves competitive performance with state-of-the-art self-supervised models. Furthermore, our analysis reveals that neural manifolds corresponding to different categories emerge naturally and are effectively separated in the learned representation space... 4 Evaluation 4.1 Linear evaluation 4.2 Semi-supervised learning 4.3 Transfer to object detection tasks |
| Researcher Affiliation | Academia | Guanming Zhang1,2 David J. Heeger2,3 Stefano Martiniani1,2,4,5 1 Center for Soft Matter Research, Department of Physics, New York University 2 Center for Neural Science, New York University 3 Department of Psychology, New York University 4 Simons Center for Computational Physical Chemistry, New York University 5 Courant Institute of Mathematical Sciences, New York University EMAIL |
| Pseudocode | Yes | A Algorithm Algorithm 1 CLAMP algorithm |
| Open Source Code | Yes | The code for CLAMP is available at https://github.com/guanming-zhang/clamp |
| Open Datasets | Yes | CIFAR10 For the CIFAR10 dataset [38], we used the Res Net-18 [39] network... Image Net-100/1K For the Image Net dataset [40]... We selected 10 images from the MNIST dataset [37]... The model is finetuned on the VOC2007+2012 training set and evaluated on the VOC2007 test set (Table. 2). |
| Dataset Splits | Yes | Training was conducted for 100 epochs on Image Net-1K and 200 epochs on Image Net-100. Classification accuracies are reported on the corresponding validation sets (Table 1)... We evaluated the model using a semi-supervised setup with 1% and 10% split (the same as [1]) of the Image Net-1K training dataset... At each epoch, we compute the average number of neighbors (for m = 4 views) and the average manifold size, Ei data[ Λi/m], using a 1% validation split. |
| Hardware Specification | Yes | Training on 8 A100 GPUs with batch size 1024 for 4 views with distributed data parallelization for 100 epochs took approximately 17 hours. |
| Software Dependencies | No | We used the LARS optimizer [41] for training the network... We use the Adam optimizer... We use SGD optimizer... We use Albumentaions library for fast image augmentation [53]. (No specific version numbers are provided for these software components in the text.) |
| Experiment Setup | Yes | We used the LARS optimizer [41] for training the network. For CIFAR10, we trained the model for 1000 epochs using the warmup-decay learning rate scheduler, and with 10 warmup steps and a cosine decay for rest of the steps. For Image Net-100, we trained the model for 200 epochs with 10 steps of warmup and 190 steps of cosine decay for the learning rate. For Image Net-1K, we trained the model for 100 epochs with 10 steps of warmup and 90 steps of cosine decay for the learning rate... rs = 8.5 is applied for CIFAR-10 and Image Net-100/1K. |