BI 120 James Fitzgerald, Andrew Saxe, Weinan Sun: Optimizing Memories
Brain Inspired
00:00
Complementary Learning Systems - What's New About a Generalization?
In the presence of noise, it's not always ideal for a slow learning system to learn forever. At some point you actually have to stop learning to avoid over fitting to this noise. So if the amount of experiences roughly match the number of degrees of freedom in your model, then that's the point where a lot of benefit can be gained. If it's just playing data free,you could determine what the whole map should be like. But if it's also noisy, as badly as possible, there is no such thing as 'perfect' generalization.
Transcript
Play full episode
