As regular readers know, I’m interested in methodviz, and so I took a look at this visualisation by Adam Harley, a doctoral student at Carnegie Mellon. You draw a number in the top right and it guesses the digit using a convolutional neural network. The interesting thing is the attempt to show what’s going on.
The number I drew in this instance, a deliberately bad 2, is the input layer at the bottom. This goes into a convolutional layer of 5 grids, then downsampled to another layer, and so on until you get the 10-category multinomial output, which in this case gets it wrong. That’s CNNs for ya, not Adam’s fault, to be clear.
This is like Tensorflow Playground but with image input rather than two continuous variables. It’s a difficult thing to show all in one place, and this does a good job.
- The colours!
- The lines on mouseover from a pixel in layer i back to those in layer i-1 that influence it
- You can zoom in and out
- The obvious control buttons and input zone
I don’t like:
- No explanation of the colours
- No explanation of the lines and how they relate to weights
- No scrolling once you’ve zoomed
But all in all, a nice page and a great springboard to doing more methodviz.