An arbitrary data set in a cat that reveals Google deep learning work?

Question

I decide if I want to study Deep Learning. Can you help with a dumb question? At Google, I understand that their ANN uncontrollably learns that the “cat face function” is best activated, given that the cat’s face meets with rhinestones. Is the set of exercises absolutely random frames for youtube, or are they YouTube frames with cats in them, maybe? It is intriguing. Thanks!

Answer 1

Yes, a set of workouts are random frames from youtube, one per video, unbiased for any searches. The sample was random, it so happened that if you want to encode YouTube frames well, you will need a human face detector / encoder and a code encoder.

Answer 2

Their ANN is a classifier that answers the question "Does this frame include the face of a cat?". Like any other classifier, it needs a set of training materials for uncontrolled training, which is somewhat balanced. However, using random Youtube frames will probably give you a very distorted dataset (too many negative samples). To get a more balanced set of workouts, they probably use keywords in the video title or manual video selection to get more positive patterns and less negative ones.

Answer 3

From what I understand, a deep dream network learns a lot more cat photographs than anything else. The concomitant bias for cat faces is associated with “retraining,” which is a common problem for neural networks. Overflow occurs when the weight of the network becomes excessively changed and concrete. In a sense, the network is becoming too good at classifying training data and therefore has problems generalizing to new inputs.