*Abstract*

Nowadays, collecting the right dataset for machine learning is often more challenging than choosing the model. We address this with photorealistic synthetic training data – labelled images of humans made using computer graphics. With synthetics we can generate clean labels without annotation noise or error, produce labels otherwise impossible to annotate by hand, and easily control variation and diversity in our datasets. I will show you how synthetics underpins our work on understanding humans, including how it enables fast and accurate 3D face reconstruction, in the wild.

*Bio*

Dr. Erroll Wood is a Staff Software Engineer at Google, working on Digital Humans. Previously, he was a member of Microsoft's Mixed Reality AI Lab, where he worked on hand tracking for HoloLens 2, avatars for Microsoft Mesh, synthetic data for face tracking, and Holoportation. He did his PhD at the University of Cambridge, working on gaze estimation.

Google Calendar for Future Seminars: https://calendar.google.com/calendar/u/0?cid=c2pjcHN0YXM2N3QyMWU3c2FqNjB...

Abstract:

We present Claim-Dissector: a novel latent variable model for fact-checking and fact-analysis, which given a claim and a set of retrieved provenances allows learning jointly (i) what are the provenances relevant to this claim (ii) what is the veracity of this claim. We show that our system achieves state-of-the-art results on FEVER comparable to two-stage systems often used in traditional fact-checking pipelines, while using significantly less parameters and computation.
Our analysis shows that proposed approach further allows to learn not just which provenances are relevant, but also which provenances lead to supporting and which toward denying the claim, without direct supervision. This not only adds interpretability, but also allows to detect claims with conflicting evidence automatically. Furthermore, we study whether our model can learn fine-grained relevance cues while using coarse-grained supervision. We show that our model can achieve competitive sentence-recall while using only paragraph-level relevance supervision. Finally, traversing towards the finest granularity of relevance, we show that our framework is capable of achieving strong token-level interpretability. To do this, we present a new benchmark focusing on token-level interpretability ― humans annotate tokens in relevant provenances they considered essential when making their judgement. Then we measure how similar are these annotations to tokens our model is focusing on. Our code, dataset and demo will be released online.

Bio:

Martin Fajčík (read as Fay-Cheek) is a PhD candidate in Natural Language Processing from Knowledge Technology Research Group active at FIT-BUT in Brno, Czech Republic, advised by prof. Pavel Smrž (ž is read like j in french "Jean"). From 2021, he also works as a research assistant in IDIAP research institute based in Martigny, Switzerland. His PhD work is focusing on open-domain knowledge processing, mainly in question answering and fact-checking. He enjoys a good hikes and an informal discussions over tea.