Putting a bug in ML: The moth olfactory network learns to read MNIST

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Biological neural networks exhibit efficient learning mechanisms, particularly in low-data environments.
• The moth olfactory network, a simple yet effective biological system, utilizes features like cascaded networks, inhibition, noise, sparsity, and reward-based plasticity for rapid learning.

Purpose of the Study:

• To characterize learning architectures in biological neural networks for few-sample training.
• To translate these biological structures into a machine learning context, creating efficient artificial neural networks.

Main Methods:

• Developed MothNet, a computational model of the moth olfactory network, integrating biophysical properties and in vivo data.
• Utilized MothNet to perform few-sample learning on the MNIST digit recognition task.

Main Results:

• MothNet successfully learned to recognize MNIST digits with only 1-10 training samples per class.
• Outperformed standard machine learning methods (k-NN, SVM, NNs) in the few-samples regime.
• Matched specialized one-shot transfer-learning methods without requiring pre-training.

Conclusions:

• Biological brain structures offer effective algorithmic solutions for efficient, low-data learning.
• MothNet demonstrates a viable approach to developing artificial neural networks that overcome the data-hungry limitations of current models.