4 ms·
A recipe for the typical media hype "AI learns physics" discovery 1. Find a well-known classical problem with a battle-tested deterministic model representatio
by geoalchimista 7y ago
A recipe for the typical media hype "AI learns physics" discovery
1. Find a well-known classical problem with a battle-tested deterministic model representation
2. Generate synthetic data using the model
3. Train a 1-layer neural network to score 99% in predicting the synthetic data (even if the deterministic model can score 99.999% but let's ignore the minor difference)
4. Claim AI "learns" the underlying physics and does a better job than classical models
5. World dominance
- knzhou 7y agoThis is exactly it. Earlier this year there was a hyped up paper [0] going around claiming that they could derive physics with AI. On inspection I found that what they actually did was this: 1. Take one simple physics equation, such as F = ma 2. Generate 10^6 tuples exactly satisfying this equation (F = 6, m = 2, a = 3), with absolutely no redundant, erroneous, or extraneous information 3. Feed these to a primitive "AI" system which brute-force tries the simplest relations that could relate the three provided variables (m = Fa, a = Fm, F = ma, ...) 4. Recover F = ma, declare physicists obsolete It's such a poor imitation of the real process of discovery in science that it's frankly insulting. When people hear AI is involved with something they turn their brains off. 0: https://arxiv.org/abs/1905.11481 https://arxiv.org/abs/1905.11481