Remembering to Learn: A Brain-Inspired Approach to Catastrophic Forgetting in AI

This project extends the FlyModel framework—originally proposed by Shen et al.—to evaluate and improve biologically inspired continual learning systems. The FlyModel uses sparse representations and local learning rules based on the fruit fly mushroom body to support class-incremental learning.

As part of a final project for Columbia Engineering, three biologically grounded extensions were implemented and tested:

• Brain-Inspired Replay (BI-R): class-conditional VAE for internal generative replay
• Elastic Weight Consolidation (EWC): synaptic regularization based on Fisher Information
• Cascade Synapses: multi-state synaptic model simulating long-term potentiation

All variants were evaluated on CIFAR-100 under a class-incremental setup. While BI-R and EWC offered modest gains, cascade synapses significantly reduced forgetting without additional memory or replay.

---

Citations

@article{shen2021algorithmic,
  title={Algorithmic insights on continual learning from fruit flies},
  author={Shen, Yang and Dasgupta, Sanjoy and Navlakha, Saket},
  journal={arXiv preprint arXiv:2107.07617},
  year={2021}
}

@article{wang2023comprehensive,
  title={A Comprehensive Survey of Forgetting in Deep Learning Beyond Continual Learning},
  author={Wang, Zhenyi and Yang, Enneng and Shen, Li and Huang, Heng},
  journal={arXiv preprint arXiv:2307.09218},
  year={2023}
}

@article{mccloskey1989catastrophic,
  title={Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem},
  author={McCloskey, Michael and Cohen, Neal J.},
  journal={Psychology of Learning and Motivation},
  volume={24},
  pages={109--165},
  year={1989},
  publisher={Elsevier}
}

@article{kirkpatrick2017overcoming,
  title={Overcoming catastrophic forgetting in neural networks},
  author={Kirkpatrick, James and Pascanu, Razvan and Rabinowitz, Neil and Veness, Joel and Desjardins, Guillaume and Rusu, Andrei A and Milan, Kieran and Quan, John and Ramalho, Tiago and Grabska-Barwinska, Agnieszka and others},
  journal={Proceedings of the National Academy of Sciences},
  volume={114},
  number={13},
  pages={3521--3526},
  year={2017},
  publisher={National Acad Sciences}
}

@article{gonzalez2020sleep,
  title={Can sleep protect memories from catastrophic forgetting?},
  author={Gonzalez, Oscar C and Sokolov, Yury and Krishnan, Giri P and Delanois, Jean Erik and Bazhenov, Maxim},
  journal={eLife},
  volume={9},
  pages={e51005},
  year={2020},
  publisher={eLife Sciences Publications Limited}
}

@inproceedings{kemker2018measuring,
  title={Measuring catastrophic forgetting in neural networks},
  author={Kemker, Ronald and Kanan, Christopher},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={32},
  number={1},
  year={2018}
}

@techreport{krizhevsky2009learning,
  author       = {Alex Krizhevsky and Geoffrey Hinton},
  title        = {Learning Multiple Layers of Features from Tiny Images},
  institution  = {University of Toronto},
  year         = {2009},
  number       = {UTML TR 2009-001},
  address      = {Toronto, Ontario},
  url          = {https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf}
}

@article{sun2020boovae,
  title={BooVAE: Boosting Approach for Continual Learning of VAE},
  author={Sun, Yibo and Wu, Jianing and Zhang, Cheng and others},
  journal={arXiv preprint arXiv:1908.11853},
  year={2020}
}

@article{bliss2011ltp,
  title     = {Long-term potentiation and long-term depression: a clinical perspective},
  author    = {Bliss, Timothy V. P. and Cooke, Sam F.},
  journal   = {Clinics (Sao Paulo)},
  volume    = {66 Suppl 1},
  pages     = {3--17},
  year      = {2011},
  doi       = {10.1590/s1807-59322011001300002},
  pmid      = {21779718},
  pmcid     = {PMC3118435},
  publisher = {Clinics},
  issn      = {1807-5932},
  url       = {https://doi.org/10.1590/s1807-59322011001300002}
}

@article{vandevenbir,
  title={Brain-inspired replay for continual learning with artificial neural networks},
  author={van de Ven, Gido M and Siegelmann, Hava T and Tolias, Andreas S},
  journal={Nature Communications},
  volume={11},
  number={1},
  pages={4069},
  year={2020},
  publisher={Nature Publishing Group}
}

@inproceedings{lopezpaz2017gradient,
  title={Gradient Episodic Memory for Continual Learning},
  author={Lopez-Paz, David and Ranzato, Marc'Aurelio},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  volume={30},
  year={2017},
  publisher={Curran Associates, Inc.},
  url={https://proceedings.neurips.cc/paper_files/paper/2017/file/f87522788a2be2d171666752f97ddebb-Paper.pdf}
}

@article{fusi2005cascade,
  title={Cascade models of synaptically stored memories},
  author={Fusi, Stefano and Drew, Patrick J and Abbott, L F},
  journal={Neuron},
  volume={45},
  number={4},
  pages={599--611},
  year={2005},
  publisher={Elsevier},
  doi={10.1016/j.neuron.2005.01.010}
}