4 ms·
> As Ng explained, "The remarkable thing was that [the system] had discovered the concept of a cat itself. No one had ever told it what a cat is. That was a mil
by strangeworld1 11y ago
> As Ng explained, "The remarkable thing was that [the system] had discovered the concept of a cat itself. No one had ever told it what a cat is. That was a milestone in machine learning."
If you dont mind, let me call this bullshit out. No one in machine / deep learning thinks that this was anything more than PR fluff combined with a very very weak paper that people had to approximately cheat to find the actual cats. (You had to initialize your random vector very close to an actual cat image, and do gradient descent, before it "figured out" for itself about a cat).
- asgard1024 11y agoI am not sure why you're saying they cheated. I am a layman, but my understanding of that quote is that he refers to a rather large deep neural network, which was trained unsupervised to encode general images (not just cats), had a neuron representing the concept of "cat". You can run the network in generative mode to see how the general concept of cat would look like. So I am puzzled by your comment - you seem to be talking about supervised learning, but I think the network was trained unsupervised.
- conceit 11y ago> You had to initialize your random vector very close to an actual cat image, and do gradient descent, before it "figured out" for itself about a cat and > a rather large deep neural network, which was trained unsupervised to encode general images (not just cats) seem to be in contradiction, afaict, or was general vector initialized rather close general images?
- pizza 11y agoThe initial value of the vector determines what it will encode/converge to.