Realistic neuron model boosts AI performance without extra parameters
A new study replaces the standard point neuron model in artificial neural networks (ANNs) with a more realistic cortical cell model, yielding gains in expressivity, robustness, and learning speed while reducing memorization and data requirements. The point neuron model, adopted from 1950s neuroscience, has been known for decades to be oversimplified, yet remained the default in ANNs. The substitution, detailed in a preprint on arXiv, does not increase the number of parameters. The work is categorized under Neural and Evolutionary Computing.
Key facts
- Standard point neuron model in ANNs dates to the 1950s.
- Neuroscience has long shown the point neuron model is too simplistic.
- The new model uses a recent cortical cell model.
- No increase in the number of parameters.
- Advantages include increased expressivity, robustness, and learning speed.
- Reduction in memorization and training data needed.
- Theoretical analyses and experimental results support the findings.
- Preprint available on arXiv.
Entities
Institutions
- arXiv