Biologically inspired heterogeneous learning for accurate, efficient and low-latency neural network
2024

Biologically Inspired Learning for Neural Networks

publication Evidence: high

Author Information

Author(s): Wang Bo, Zhang Yuxuan, Li Hongjue, Dou Hongkun, Guo Yuchen, Deng Yue

Primary Institution: School of Astronautics, Beihang University, Beijing, China

Hypothesis

Can a spiking neural network (SNN) be enhanced by incorporating self-inhibiting autapse and neuron heterogeneity to improve learning and memorizing capacities?

Conclusion

The proposed HIFI model outperforms traditional SNNs in accuracy, efficiency, and latency while accurately identifying rare cell types in scRNA-seq data.

Supporting Evidence

  • The HIFI model demonstrated a 1%–10% improvement in accuracy and a 17.83-fold reduction in energy consumption.
  • HIFI was able to identify rare cell types associated with severe brain diseases that traditional SNNs could not.
  • The model achieved superior performance on multiple AI benchmarks compared to existing SNNs.

Takeaway

This study created a new type of neural network that learns better by mimicking how real brain cells work, making it faster and more efficient.

Methodology

The study developed a Heterogeneous spIking Framework with self-Inhibiting neurons (HIFI) using a bi-level programming paradigm to optimize neuron-level and network-level parameters.

Limitations

The model's computational complexity may limit its application in resource-constrained environments.

Digital Object Identifier (DOI)

10.1093/nsr/nwae301

Want to read the original?

Access the complete publication on the publisher's website

View Original Publication