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..
Stochastic Forward-Forward Learning through Representational Dimensionality Compression
Authors: Zhichao Zhu, YANG QI, Hengyuan Ma, Wenlian Lu, Jianfeng Feng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The numerical experiments on standard datasets demonstrate that the proposed method can achieve performance comparable to that of other non-BP methods. Moreover, we show that noise plays a constructive role that can enhance generalization and improve inference when predictions are derived from the mean of squared output, which is equivalent to making predictions based on an energy term. |
| Researcher Affiliation | Academia | 1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai 200433, China.2Key Laboratory of Computational Neuroscience and Brain-Inspired Intelligence (Fudan University), Ministry of Education, China. 3MOE Frontiers Center for Brain Science, Fudan University, Shanghai 200433, China. 4Ji Hua Laboratory, Foshan 528200, China. *Corresponding author. |
| Pseudocode | No | The paper includes a network architecture diagram (Figure 2) and describes procedures in text, but it does not contain a dedicated pseudocode or algorithm block with structured steps. |
| Open Source Code | Yes | The code is available at https://github.com/Zhichao Zhu/Stochastic Forward Forward. |
| Open Datasets | Yes | We use the MNIST[32], CIFAR-10 and CIFAR-100 datasets [33] to verify the effectiveness of the proposed goodness function. |
| Dataset Splits | Yes | We use the MNIST[32], CIFAR-10 and CIFAR-100 datasets [33] to verify the effectiveness of the proposed goodness function. For MNIST, we only use random crop for data augmentation. For CIFAR-10 and CIFAR-100, we first apply zero phase component whitening and then random crop and random horizontal flip for data augmentation. The best validation accuracy is reported as we care more about the upper bound of the performance. |
| Hardware Specification | Yes | All experiments utilize a NVIDIA RTX 3090 GPU and an Intel Xeon(R) Gold 6226R CPU, as detailed in the Appendix A1.2. |
| Software Dependencies | No | All models were trained using the Adam W optimizer with a learning rate of 0.001 and a weight decay of 0.01. The output of each block was projected onto a predefined lower-dimensional space using randomly generated orthogonal basis vectors sampled from the Haar distribution (via Sci Py). No specific version numbers for these software dependencies are provided. |
| Experiment Setup | Yes | All models were trained using the Adam W optimizer with a learning rate of 0.001 and a weight decay of 0.01. A cosine annealing learning rate schedule was applied, with a maximum of 3 and 60 iterations for phase 1 and phase 2 training, respectively. The batch size was fixed at 128 across all experiments. During phase 1 training, given an input batch X RB C H W , after passing through the batch normalization layer of the first convolutional block, we applied dropout with a probability of 0.2 to generate N = 20 noisy variants per sample... Under the default Graded setting, the projection dimensions were 30-20-10 for MNIST and CIFAR-10 and 90-150-100 for CIFAR-100. The block was then optimized using the proposed L objective computed on the projected output. By default, a trade-off factor α = 0.5 was used. |