A model of neurobiologically plausible least-squares learning in visual cortex
Lists
Only you see this
Abstract
We show mathematically how a pseudo-inverse approach to training an artificial neural network for classification using a mean square error loss function and large batches of training data can be recast as an iterative method that circuits of real neurons in visual cortex could plausibly learn in an online manner. We argue the case for neurobiological plausibility based on our illustration that the iterative method can be reformulated as an unsupervised stage that learns to decorrelate using an anti-Hebbian learning rule, and a supervised stage that follows simple Hebbian learning while retaining mean square error optimality. Importantly, we modify the baseline method to ensure the learning rules rely on information that would be available locally at synapses if the method were instantiated in a network of cortical neurons. We demonstrate results comparable on par or better than similar iterative methods applied to the MNIST hand-written digits image classification benchmark, and the related but more challenging EMNIST database. The proposed learning algorithm learns quickly, reaching good accuracy with a small amount of training data.
BibTeX
@inproceedings{bagchi2018model,
author = {Samya Bagchi and Mark D. McDonnell},
title = {{A model of neurobiologically plausible least-squares learning in visual cortex}},
booktitle = {IJCNN},
year = {2018},
url = {https://openreview.net/forum?id=PPh_4epdcpO},
}Loading related content...