Vector Quantized-Variational AutoEncoders (VQ-VAE) are generative models based on discrete latent representations of the data, where inputs are mapped to a finite set of learned embeddings. To generate new samples, an autoregressive prior distribution over the discrete states must be trained separately. This prior is generally very complex and leads to slow generation. In this work, we propose a new model to train the prior and the encoder/decoder networks simultaneously. We build a diffusion bridge between a continuous coded vector and a non-informative prior distribution. The latent discrete states are then given as random functions of these continuous vectors. We show that our model is competitive with the autoregressive prior on the mini-Imagenet and CIFAR dataset and is efficient in both optimization and sampling. Our framework also extends the standard VQ-VAE and enables end-to-end training.

Alice Martin, Guillaume Quispe, Charles Ollion, Sylvain Le Corff, Florian Strub, Olivier Pietquin. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Sampling from a complex distribution $\pi$ and approximating its intractable normalizing constant $\mathrm{Z}$ are challenging problems. In this paper, a novel family of importance samplers (IS) and Markov chain Monte Carlo (MCMC) samplers is derived. Given an invertible map $\mathrm{T}$, these schemes combine (with weights) elements from the forward and backward Orbits through points sampled from a proposal distribution $\rho$. The map $\mathrm{T}$ does not leave the target $\pi$ invariant, hence the name NEO, standing for Non-Equilibrium Orbits. NEO-IS provides unbiased estimators of the normalizing constant and self-normalized IS estimators of expectations under $\pi$ while NEO-MCMC combines multiple NEO-IS estimates of the normalizing constant and an iterated sampling-importance resampling mechanism to sample from $\pi$. For $\mathrm{T}$ chosen as a discrete-time integrator of a conformal Hamiltonian system, NEO-IS achieves state-of-the art performance on difficult benchmarks and NEO-MCMC is able to explore highly multimodal targets. Additionally, we provide detailed theoretical results for both methods. In particular, we show that NEO-MCMC is uniformly geometrically ergodic and establish explicit mixing time estimates under mild conditions.

Evolutionary robotics (ER) have successfully built robot controllers presenting a reactive behavior. However, the evolution of cognitive controllers is still a challenge. We hypothesize here that a fitness function which rewards the fulfillment of a task requiring cognitive abilities does not necessarily reward the stepping stones that lead to cognitive controllers. In other words, our hypothesis is that evolving cognitive abilities is a deceptive problem, and that the selective pressures driving the evolutionary search are of critical importance. This paper presents some experiments to confirm this hypothesis and addresses this selective pressure problem by introducing a new helper-objective that rewards controllers with a memory. This is potentially useful for the design of controllers in which an internal representation of some data is required to solve a task. It does not assume how the memory is stored in the controller, therefore reducing the bias towards a particular solution. The new objective is tested in a multi-objective scheme on a T-maze ER task — a task involving both navigation and working memory. The efficiency of the helperobjective is studied, as well as its effects on the overall performance and generalization ability of the controller.

Exploration and exploitation are two complementary aspects of Evolutionary Algorithms. Exploration, in particular, is promoted by specific diversity keeping mechanisms generally relying on the genotype or the fitness value. Recent works suggest that, in the case of Evolutionary Robotics or more generally behavioral system evolution, promoting exploration directly in the behavioral space is of critical importance. In this work an exploration indicator is proposed, based on the sparseness of the population in the behavioral space. This exploration measure is used on two challenging neuro-evolution experiments and validated by showing the dependence of the fitness at the end of the run on the exploration measure during the very first generations. Such a prediction ability could be used to design parameter settings algorithms or selection algorithms dedicated to the evolution of behavioral systems. Several other potential uses of this measure are also proposed and discussed.