What is an Autoencoder and How Does it Compress Data?

I continue my study of artificial intelligence in the 'Russian school'. It's a tough but good school. It took me 10 minutes to solve the autoencoder task with an error rate of 6.7%, but it took three days of experimentation to tame the error size down to 5%.

The first task involves training a network called an encoder to encode an image of a human face as efficiently as possible into a few digits (in this case, 32 digits). Then, I train a second network called a decoder to attempt to reconstruct the image from those few digits. Essentially, it's a clever and lossy super-compression of the image.

32 digits is quite a small amount to describe an image. The network must focus only on what is important. How do you like the results?

This one turned out quite nicely. The face reconstructed from just 32 digits resembles the original and appears only slightly blurred.

A somewhat poorer result is likely due to the uneven distribution of the dataset. The network inaccurately reconstructs a wider nose and full, slightly open mouth.

A fairly nice reconstruction. It seems to have somewhat softened the sharp features.

Here, the glasses took quite a hit. Their structure is too complex for such strong compression. The expression of the mouth was also translated into something more common.