But this new kind of transcoder gives us a string of text, in English (or another language of your choice), which we can simply read.
And this string of text simply… tells us what the sub-block is doing.
Natural-Language Autoencoders sound a lot like this, although they interpret activations rather than transcoding, and of course they don’t have the second property mention in the post (zero reconstruction error).
Natural-Language Autoencoders sound a lot like this, although they interpret activations rather than transcoding, and of course they don’t have the second property mention in the post (zero reconstruction error).