3 ms·
Because the solution space is convex if you've chosen your representation well.
by robotresearcher 11y ago
Because the solution space is convex if you've chosen your representation well.
- gipp 11y agoThis is definitely not the case for general neural networks, though.
- robotresearcher 11y agoIf gradient descent is working reliably, the problem is convex. See the sibling comments for the intuition for large dimensional spaces.
- gipp 11y ago"The problem is convex" and "the algorithm is unlikely to get stuck in a local minimum on realistic problems" are very different things.
- robotresearcher 11y agoRight. Hence the word 'reliably'.