context: Lossy mnist encoder - #6 by JarvisGoBrr
Just wanted to share this because I think it looks cool.
its the same encoder as before but with binarized synapses and thresholded reconstruction.
I had to actually implement a new type of additive boosting for the binarized synapses otherwise it just gets stuck at local minima and normal multiplicative boosting is not enough to get it going.
t
You called it dreaming because those are just random sdrs and their corresponding decoded 28x28 image?
PS1 what accuracy value means in the pictures?
PS2 any code available? It would be interesting to see classifier results on these encodings.
For obvious reasons one can think of the “value” of an encoder as how well performs a downstream classification (or regression but not with MNIST)
Could you elaborate on the additive boosting you implemented? Currently trying to get something similar to work and I would really appreciate learning about how you overcame the challenges with binary weights
@JarvisGoBrr is this open source? Do you have a repo were we can look how you implemented this?