Recent & Upcoming Talks

In order to leave the factory floors and research labs, future robots must abandon their stiff and pre-programmed movements and be capable to learn complex policies. These control policies must be adaptive to the inevitable changes of the physical world and must only select feasible actions. Current black-box function approximation approaches are not sufficient for such tasks as these approaches overfit to the training domain and commonly perform damaging actions. In this talk, I want to introduce my research focusing on robot learning for physical robots by using inductive biases and deep learning. This approach combines the representational power of deep networks and directly includes the feasibility and robustness as inductive bias within the learning problem to enable the application to the physical system. Using this approach we showed that deep networks can be constrained to learn physically plausible models and optimal feedback controller by incorporating domain knowledge and demonstrated the applicability to the physical system.

The use of Gaussian process models is typically limited to datasets with a few tens of thousands of observations due to their complexity and memory footprint. The two most commonly used methods to overcome this limitation are 1) the variational sparse approximation which relies on inducing points and 2) the state-space equivalent formulation of Gaussian processes, which can be seen as exploiting some sparsity in the precision matrix. We propose to take the best of both worlds: we show that the inducing point framework is still valid for state space models and that it can bring further computational and memory savings. This formulation allows for minibatching and can be applied to deep Gaussian process models.

I will discuss multiple-data-source prediction and modelling problems arising in a number of fields, for instance in omics-based precision medicine. What is less common is that some of the data sources are experts, whose time is costly, changing the problem to active learning for prediction. We have addressed this setup as a probabilistic modelling problem, where different types of sources need different modelling assumptions, expert user models ultimately drawing from cognitive science. This brings links to other lines of work on interactive intent modelling and likelihood-free inference to infer the user models.

We derive a family of Monte Carlo estimators for gradients of expectations, which is related to the log-derivative trick, but involves pairwise interactions between samples. The first of these comes from either a) introducing and approximating an integral representation based on the fundamental theorem of calculus, or b) applying the reparameterisation trick to an implicit parameterisation under infinitesimal perturbation of the parameters. From the former perspective we generalise to a reproducing kernel Hilbert space representation, giving rise to locality parameter in the pairwise interactions mentioned above. The resulting estimators are unbiased and shown to offer an independent component of useful information in comparison with the log-derivative estimator. Promising analytical and numerical examples confirm the intuitions behind the new estimators.

In machine learning, new learning algorithms are designed by borrowing ideas from optimization and statistics followed by an extensive empirical efforts to make them practical. However, there is a lack of underlying principles to guide this process. I will present a stochastic learning algorithm derived from Bayesian principle. Using this algorithm, we can obtain a range of existing algorithms: from classical methods such as least-squares, Newton’s method, and Kalman filter to new deep-learning algorithms such as RMSprop and Adam. Surprisingly, using the same principles, new algorithms can be naturally obtained even for the challenging learning tasks such as online learning, continual learning, and reinforcement learning. This talk will summarize recent works and outline future directions on how this principle can be used to make algorithms that mimic the learning behaviour of living beings.