A systematic comparison of Backpropagation (BP) and Feedback Alignment (FA) — a biologically plausible learning rule — on the MNIST digit classification task.
Team NeuroAligners — Neuromatch Academy Microlearning Project (Summer 2025)
Team Lead: Alireza AlaeiFaradonbeh — Certificate
Backpropagation (BP) drives modern deep learning but requires symmetric weight transport — a biologically implausible mechanism. Feedback Alignment (FA) bypasses this by using fixed random feedback weights, offering a more brain-like alternative (Lillicrap et al., 2016).
This project investigates:
- Performance: Accuracy, convergence speed, noise robustness
- Learning dynamics: Weight update alignment (cosine similarity), signal-to-noise ratio (SNR)
- Representations: Latent feature similarity, weight norm evolution
| Metric | Backpropagation | Feedback Alignment |
|---|---|---|
| Test Accuracy | 97.0% | 83.5% |
| Test Loss | 0.107 | 0.685 |
| Epochs to 90% Val Acc | 1 | Not reached |
FA demonstrates more graceful degradation under extreme noise:
| Gaussian Noise σ | BP Acc | FA Acc |
|---|---|---|
| 0.0 | 97.0% | 83.5% |
| 1.0 | 66.2% | 57.2% |
| 2.0 | 32.2% | 33.9% |
| 3.0 | 21.4% | 25.6% |
At σ ≥ 1.8, FA matches or exceeds BP, suggesting its less specialized representations are more robust to severe input corruption.
├── README.md ← This file
├── requirements.txt ← Python dependencies
├── notebook/
│ └── main.ipynb ← Full experiment notebook (runnable)
├── proposal/
│ ├── NeuroAI-Final Proposal.docx
│ └── README.md ← Proposal summary
├── poster/
│ └── NeuroAligners-Poster.pptx
├── report/
│ ├── Feedback Alignment in Deep Networks.md ← Polished final report
│ ├── Feedback Alignment in Deep Networks.docx
│ ├── METHODS & ARCHITECTURE.docx
│ ├── project code + Progress Report.docx
│ └── workflow of code.docx
├── references/
│ ├── Papers.docx ← Annotated bibliography
│ ├── README.md ← Reference paper list
│ └── *.pdf ← Key reference papers
├── results/
│ └── NeuroMatch_Result.zip ← All experiment outputs (252 files)
└── team/
└── members.md ← Team roster with roles & contributions
git clone https://github.com/88888Z/bioinspired-learning-rules.git
cd bioinspired-learning-rules
pip install -r requirements.txt
jupyter notebook notebook/main.ipynb- Python 3.8+
- PyTorch (CUDA recommended)
- torchvision, matplotlib, seaborn, numpy
The notebook generates 6 key plots covering accuracy/loss curves, noise robustness (Gaussian & salt-and-pepper), per-layer bias/SNR analysis, latent representation similarity, weight norms, and batch-level dynamics. Archived results for 9 hyperparameter configurations are in results/.
| Role | Name |
|---|---|
| Team Lead & Lead Developer | Alireza AlaeiFaradonbeh |
| Core Implementation | Morteza Hashemi |
| Data & Architecture | Yasaman Malahzadeh |
| Research & Literature | Avisa Fallah |
| Progress & Reporting | Melina Mostafa |
| Literature | Mobina Golsefatan |
| Literature | Sina Pirmoradian |
See team/members.md for full details.
- Lillicrap, T. P., et al. (2016). Random synaptic feedback weights support error backpropagation for deep learning. Nature Communications, 7, 13276.
- Konishi & Miura (2024). Biologically Plausible CNNs and Autoencoders with Feedback Alignment.
- LeCun, Y., et al. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324.
- Nayebi, A., et al. (2020). Identifying learning rules from neural network observables. NeurIPS.