3IA PhD/Postdoc Seminar #4

Conditional image generation using structural priors
Abstract: High quality image generation has now been around for a while. However, restricting the process through the structural data still has a long way to go. Among the applications, novel view synthesis of human images is one of the most challenging tasks that simultaneously demands for transferring the pose and textures through the same pathway of networks, causing the information to be washed off during their consecutive interactions. In this context, I will review some novel strategies that try to solve the problem via recently proposed data-driven prior estimations and discuss some downsides that open up new design spaces to further improve the quality of this kind of synthesis models.

Randomizing gradient descents on the space of probability measures
Abstract: Numerous machine learning problems may be recast as an optimization over the space of probability measures. For example, the stochastic gradient descents associated to a certain class of over-parameterised neural networks can be approximated by gradient descents in the space of probability measures, providing theoretical justification for convergence. Thus far, these descents are non-random as a consequence of assuming each neuron is subject to its own source of randomness. By introducing an appropriate systemic noise to the neural network, the corresponding descents in the space of probability measures are expected to become stochastic, consequently enjoying explorative properties. The aim is to introduce practically implementable systemic noises for which theoretical guarantees may be established.

Exploration noise for Mean Field Games
Abstract: From their beginning Mean Field Games (MFGs) have attracted considerable attention and interest from disciplines ranging from economics to ecology. In this talk we will present some recent results we have obtained for learning MFGs equilibria in presence of a common noise and use the noise itself for exploration in a reinforcement learning paradigm.

Additional questions and discussions