3 ms·

The one generated after looking at completely random noise on the bottom row, second from the right: http://googleresearch.blogspot.co.uk/2015/06/inceptionism-

by IanCal 11y ago

The one generated after looking at completely random noise on the bottom row, second from the right: http://googleresearch.blogspot.co.uk/2015/06/inceptionism-going-deeper-into-neural.html http://googleresearch.blogspot.co.uk/2015/06/inceptionism-go... Reminds me very heavily of The Starry Night https://www.google.com/culturalinstitute/asset-viewer/the-starry-night/bgEuwDxel93-Pg?utm_source=google&utm_medium=kp&hl=en-GB&projectId=art-project https://www.google.com/culturalinstitute/asset-viewer/the-st... Lovely imagery. I never had much luck with generative networks. I did some work putting RBMs on a GPU partly because I'd seen Hinton talk showing starting with a low level description and feeding it forwards, but always ended up with highly unstable networks myself.

Lewton 11y ago
https://lh3.googleusercontent.com/4jaIlDI1xXGOhxNejib833qA5yVMAvQLaKmmuaQ0PGY=w1200-h800-no https://lh3.googleusercontent.com/4jaIlDI1xXGOhxNejib833qA5y... full resolution image
- IanCal 11y ago
  That's great, thank you.
abrichr 11y ago
Neural networks are notoriously difficult to train due to the large number of hyper-parameters that need to be tuned. If your network never converged, it's possible your learning rate was too high, so it kept overshooting the minima of the loss function.
- IanCal 11y ago
  Quite possibly. The classification results were great, just wasn't good when trying to run things back through the network repeatedly. I did have issues that the learning rates reported in some of the original papers didn't match the ones in the released code.