Publications

Fine-tuning vision-language foundation models has emerged as a powerful approach to leveraging internet-scale data for generalization …

This paper introduces a novel method for enhancing the effectiveness of the Asynchronous Advantage Actor-Critic (A3C) algorithm by …

In this work we explore a VQGAN-transfomer based world-and-action model trained on several years of gameplay data in a team based game with complex goals requiring adversarial play and map navigation.

Causal confusion is a phenomenon where an agent learns a policy that reflects imperfect spurious correlations in the data. Such a …

In this work we develop a gradient-based meta-learning algorithm for efficient, online continual learning, that is robust and scalable to real-world visual benchmarks.

This work surveys recent work in the very nascent field of probabilistic detection and pesents insights and promising avenues for future research in this area.

This work proposes to train models with imitation learning in such that they come with a PAC-bayes bound as a performance guarantee for the model. It connects the expressions for the PAC bayes bound and the ELBO of a stochastic predictive policy learnt through likelihood maximisation. The model is then trained by minimising the pac-bayes bound as the objective that also doubles as an error bound.

This work introduces and formalises the problem of recognising intersections from drastically different viewpoints to enable place recognition for SLAM using a siamese convolutional recurrent architecture trained to classify pairs of short video streams.

CTCNet proposed using the compositional property of transformations to self-supervise learning of visual odometry from images.