Metric Space Magnitude and Generalisation in Neural Networks — mentioned by Pseudomanifold

Mon, 24 Jul 2023 07:13:15 GMT

Amazing to see 'magnitude' on the front page of Hacker News! If you are interested in a direct application of this invariant beyond ecology, check out our recent pre-print in which we study the generalisation behaviour of neural networks: https://arxiv.org/abs/2305.05611 My personal approach to magnitude is not based on category theory but rather based on weightings of a metric space. If your metric satisfies certain properties, you can obtain a measure of the 'effective number of points' of a metric space. This is particularly relevant when looking at the metric space from different…

Paper and discussion history. A mention is not an endorsement.

Metric Space Magnitude and Generalisation in Neural Networks · HackerPapers discussions

Metric Space Magnitude and Generalisation in Neural Networks — mentioned by Pseudomanifold