NeuralCompression

Hey there!

I have been working on a neural algorithm to compress images. It compresses a 32 x 32 pixel image down to just 64 numbers. These 64 numbers act as a choke point that the image must pass through. Next, it brings these numbers back up to the 32 x 32 pixel image. By training the model to output exactly the same as its input, this chokepoint can be thought of as a compressed version of the image.

After doing all of this, the model was able to output a very blurry version of the original image. While this may seem bad, it is actually a success! This just means that the image has to pass through a much too small choke point. Unfortunately, this idea is wrong, and I do not yet know exactly what is wrong.