Learning to See by Looking at Noise — mentioned by kd5bjo

Sat, 19 Feb 2022 09:13:57 GMT

Random initialization is already starting to give way. One of the more promising approaches right now is pretraining with synthetic data to obtain “good” initial weights. Once that converges, you’ve got a network that internally recognizes interesting features which is a bettter-than-random starting point for training with real-world data. cf. https://arxiv.org/abs/2106.05963

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

Learning to See by Looking at Noise · HackerPapers discussions

Learning to See by Looking at Noise — mentioned by kd5bjo