2 ms·
Just want to add that biological neurons are stochastic and that artificial neural nets rely on noise injection, random connection dropout, random initialisatio
by visarga 9y ago
Just want to add that biological neurons are stochastic and that artificial neural nets rely on noise injection, random connection dropout, random initialisation and random mini-batch reshuffling to work well. GANs start from a random number in generating images or other types of data. Alternative optimisation strategies to gradient descent - such as evolutionary strategies, use randomness in combining genes for the next generation of agents.
Without randomness we would have no intelligent artificial agents. It's like intelligence can only happen at the demarcation line between chaos and order.
- thomasahle 9y ago> Without randomness we would have no intelligent artificial agents. It's like intelligence can only happen at the demarcation line between chaos and order. I doubt most NN's are trained using true randomness. Most likely a fairly simple deterministic generator works just fine.