Intermittent noisy gradients can leave learning variance unbounded
Original title: On the stability of learning
Average dynamics may still reach the correct optimum while learning variance grows without bound.
The paper argues that learning can be unstable when gradient information is noisy and only intermittently available, as with sparse biological learning signals. Even if average dynamics converge to the correct optimum, variance may be asymptotically infinite. Explicit stability conditions are derived for single-layer and multi-layer linear networks. The authors say unbiased credit assignment is not enough without sufficiently frequent, informative feedback.
- Conditions are described as distinct from existing SGD stability criteria.
- Analysis covers analytically tractable single-layer and multi-layer linear networks.
- The setting is motivated by sparse, local, intermittent synaptic signals.
From the original article
1 min read at arxiv.orgLearning algorithms are usually analyzed under the assumption that informative gradients are continuously available. But this is not always true---learning signals in biological neural networks, for example, may be sparse, local, and only intermittently available at any given synapse. In this work, we show that if gradient information is noisy and only intermittently available, learning may be unstable.
By John J. Vastola · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.