I have zero experience with machine learning, AI, etc. Is it possible to explain, in layman's terms, the differences between the different types of WUs?
For example, what to rand_automata units do as opposed to ParityModified ParityMachine or EightBitModified?

There's a little discussion over in the "science" forum if you want some background.

Basically the "*Machine" workunits are building neural networks that model the behavior of simple machines to a sequence of input instructions (e.g, ParityMachine computes the 8-bit running parity of a machines with a sequence of commands "set bit 3","clear bit 7", etc...). You can see more about these machine in the paper: https://arxiv.org/abs/1805.07869 .

The idea of this project is to train a whole much of similar networks, and so *what* we're training on doesn't matter *that* much. We could just as easily train on cat pictures, but I'm the author of the above paper and had those datasets available.

The "*Modified" WUs (dataset 2) are the same as the above, but with minor changes to the way the "machine" works, so the training data learns a slightly different output. The goal is to see if we can detect networks that have been trained with the modified data as opposed to the original data based just on the trained weights of the network.

The "Rand" WU (dataset 3) ramp up the complexity of the machines we modify a lot. We randomly generated 100 different automata, and then each WU trains a network to mimic the behavior of that automata. The goal is to see if we can classify which of the 100 original automata a network was trained to mimic.

The next dataset will switch from "sequences of commands to a specific machine" to image classification. Stay tuned.