I’m super geeked ➡️ this is video of @MatthewGuz playing a game generated by a ML algorithm trained on video of Super Mario Bros., Kirby, and Mega Man Everything is learned from scratch: level design and mechanics/rules.
Paper: arxiv.org/abs/1809.02232
Paper: arxiv.org/abs/1809.02232
What I like about this video is that the game is very different from any of the training examples.
The conceptual expansion algorithm is able to extrapolate beyond the training data, hypothesizing the existence of models that aren’t directly supported by the training data
The conceptual expansion algorithm is able to extrapolate beyond the training data, hypothesizing the existence of models that aren’t directly supported by the training data
.@MatthewGuz and I think the ability to extrapolate beyond training data is a very important quality for ML-based computational creativity.
We think conceptual expansion may be potentially important for ML generalization and domain transfer too.
We think conceptual expansion may be potentially important for ML generalization and domain transfer too.
Conceptual expansion works on deep neural nets too. Here are some experiments on few-shot learning (CIFAR) and GANs showing that the extrapolative power of conceptual expansion gives good results when data is scarce: arxiv.org/abs/1802.03605
(Might be the first few-shot GAN?)
(Might be the first few-shot GAN?)
