Our most recent research on unsupervised representation learning, titled “Neural Expectation Maximization”, will be featured at the 5th International Conference on Learning Representations (ICLR) as a workshop paper.
In our paper we introduce a novel framework for clustering that combines generalized EM with neural networks and can be implemented as an end-to-end differentiable recurrent neural network titled RNN-EM. Our method groups the individual components of an input tensor that share statistical regularities across the dataset. It learns its statistical model directly from the data and can represent complex non-linear dependencies between individual components of the input.
We apply RNN-EM to a perceptual grouping task in which the data consists of images each composed of three objects. Each object shares its structure across the data and we expect RNN-EM to learn to group the pixels belonging to each object separately and independently for each image. Indeed we empirically find that it yields the intended behaviour.
Aside from the resulting grouping our method learns a distributed representation for each group, that captures all information required to reproduce the members of that group, while ignoring all others. RNN-EM shares weights across groups and each representation learned therefore shares the same semantics. This suggests that these learned representations may be symbol-like and are most likely useful in upstream supervised learning tasks.
Code will be available upon publication.
We recently published our new paper  titled “A wavelet-based Encoding for Neuroevolution” in the Proceedings of the 2016 Genetic and Evolutionary Computation Conference (GECCO).
In our paper we introduce a novel indirect encoding scheme that encodes neural network connection weights as low-frequency wavelet-domain coefficients. A lossy Discrete Inverse Wavelet Transform (IDWT) maps these coefficients back to the neural network phenotype. This Wavelet-based Encoding (WBE) builds on top of a Discrete Cosine Transform (DCT) encoding  and similarly satisfies continuity in the genotype-phenotype mapping. However, unlike the DCT encoding, the WBE satisfies a variable degree of gene-locality.
In our experiments we observe that the WBE yields superior performance on the Octopus-arm Control task (a relevant benchmark used previously for a Reinforcement Learning competition) compared to the DCT encoding. We argue that this is due to the added gene-locality of the WBE, which positively affects the efficiency of training neural networks by means of evolutionary search. A more general theoretically motivated intuition underlying our reasoning is presented in the paper.
Other novelties in our approach arise from the flexibility of the WBE. Being able to freely choose the wavelet basis function to compute the IDWT we were able to augment the WBE with a dynamic basis function, which can be optimised alongside the low-frequency wavelet coefficients by means of the same evolutionary algorithm. The resulting approach is able to optimise the mapping and coefficients simultaneously.
The conference presentation slides and code are available online. A more in depth discussion of our approach can be found on this page (which is currently under construction).
 van Steenkiste, S., Koutník, J., Driessens, K. and Schmidhuber, J., 2016, July. A Wavelet-based Encoding for Neuroevolution. In Proceedings of the 2016 on Genetic and Evolutionary Computation Conference (pp. 517-524). ACM.
 Koutnik, J., Gomez, F. and Schmidhuber, J., 2010, July. Evolving neural networks in compressed weight space. In Proceedings of the 12th annual conference on Genetic and evolutionary computation (pp. 619-626). ACM.