bioRxivJClub
Toward defining loss functions in neuroscience: an XOR-based neuronal mechanism for local learning
bioRxiv · · Preprint · Open access
Pena Fernandez, Lloret Iglesias + 1 more
Abstract ↗AI summary
The abstract is read at the publisher; the summary is JClub's.
A binary autoencoder trained for five epochs on binarized MNIST using a signed XOR mismatch error reaches 97.7% bit-wise and 94.8% balanced reconstruction accuracy without backpropagation.
- Why it matters: Loss functions are the least mapped biological counterpart of deep learning, and it is unclear whether a locally computable signed mismatch can act as an effective learning signal.
- What they did: The study trains a binary autoencoder with one XOR comparator per output unit as the only error signal, using fixed random sign feedback to the hidden layer, for five epochs.
- The result: With the signed error, learning succeeds (96.3% bit-wise and 89.2% balanced for a convolutional model with 533 parameters), while removing the error sign collapses learning.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.
- Open access
- 1 cites