Vancouver Vision & Learning Workshop @ ICML 2025

The Bayes error represents the lowest possible error achievable by any classifier. Accurately estimating the Bayes error provides valuable insight into the intrinsic difficulty of a classification task—revealing whether there is room for performance improvement or if overfitting to the test data is occurring. In this talk, I will present an overview of our recent advances in Bayes error estimation. Our proposed methods estimate the Bayes error directly from soft labels, without relying on data instances, features, or classifiers.

Over the last decade, we've seen unprecedented progress in AI across many disciplines and applications. However, autonomous vehicles are still far from mainstream even after billions of dollars of investment. In this talk we’ll explore what’s been holding progress back, and how by adopting a modern embodied AI approach to the problem, Wayve is finally unlocking the potential of scalable autonomous driving across the globe.

We’ll also explore some of our latest research in multimodal learning to combine the power of large language models with the driving problem, and in controllable generative world models as learned simulators.

Diffusion models have achieved remarkable success in generating high-quality and diverse outputs, supported by stable and scalable training procedures. However, when applied to real-world scenarios, they continue to exhibit key limitations. In this talk, I will examine several of these fundamental challenges, including slow sampling, inadequate modeling of distributional tails, and inefficiencies in training. I will also present recent advances aimed at mitigating these issues, focusing on techniques such as accelerated sampling via trajectory and distribution matching objectives, as well as improved diffusion processes designed specifically for video generation. The talk will conclude with a discussion of emerging requirements for generative models, particularly their ability to capture rare events and heavy-tailed data distributions.

TBD

TBD

Large-scale pretraining has become the standard solution to automated reasoning over text and/or visual perception. But how far does this approach get us to systems that generalize to language use in realistic multi-agent situated interactions? First, I will talk about existing work in evaluating the spatial and compositional reasoning capabilities of current multimodal language models. Then I will talk about how these benchmarks miss a key aspect of real-world situated interactions: joint embodiment. I will discuss how joint embodiment in a shared world supports perspective-taking, an underlooked aspect of situated reasoning, and introduce a new environment and benchmark for studying the influence of perspective-taking on language use in interaction. I will also describe experiments in learning from communicative success in jointly-embodied interactions with human users.

TBD

TBD