4 ms·
I suppose it depends on the activation function, but many of them are bounded. Polynomials are unbounded at the extremes, and as such should really only be tru
by excessive 7y ago
I suppose it depends on the activation function, but many of them are bounded. Polynomials are unbounded at the extremes, and as such should really only be trusted (sometimes) for interpolation, not extrapolation. If you get a new input which is outside the radius of your training set, the answer could be pretty extreme and not supported by the training data.
- thanatropism 7y ago> Polynomials are unbounded at the extremes, and as such should really only be trusted for interpolation, not extrapolation. Very well noted. There's also this, which is usually taught in second year engineer "numerical calculus" and again extensively studied in graduate programs involving [something numbers inside computers something]: https://en.wikipedia.org/wiki/Runge's_phenomenon https://en.wikipedia.org/wiki/Runge's_phenomenon
- excessive 7y agoIt was exactly that topic which is why I included the "sometimes" qualification :-)