Cian Eastwood

Profile photo

I'm a Senior Research Scientist at Valence Labs (part of Recursion) working on AI for drug discovery.

I did my PhD jointly at the University of Edinburgh (advised by Chris Williams) and the Max Planck Institute for Intelligent Systems (advised by Bernhard Schölkopf). During my PhD, I spent time at Google DeepMind, Meta AI (FAIR) and Spotify.

Broadly, my research interests involve "preparing models for new scenarios/contexts". This includes out-of-distribution generalization and representation learning (causal, multimodal, self-supervised), as well as their application to scientific discovery.

                       

profile photo
Recent News
Selected Publications
GIVT: Generative Infinite-Vocabulary Transformers

M Tschannen, C Eastwood, F Mentzer

ECCV 2024

Code

We introduce generative transformers for real-valued vector sequences, rather than discrete tokens from a finite vocabulary.

Spuriosity Didn't Kill the Classifier: Using Invariant Predictions to Harness Spurious Features

C Eastwood*, S Singh*, A Nicolicioiu, M Vlastelica, J von Kügelgen, B Schölkopf

NeurIPS 2023

Code

We show how a weak-but-stable training signal can be used to harness complementary spurious features, boosting performance.

(Previously @ ICML 2023 Spurious Correlations Workshop)

Probable Domain Generalization via Quantile Risk Minimization

C Eastwood*, A Robey*, S Singh, J von Kügelgen, H Hassani, G J Pappas, B Schölkopf

NeurIPS 2022

Code / Video

We learn predictors that generalize with a desired probability and argue for better evaluation protocols in domain generalization.

Source-Free Adaptation to Measurement Shift via Bottom-Up Feature Restoration

C Eastwood*, I Mason*, CKI Williams, B Schölkopf

ICLR 2022 (Spotlight)

Code

We address "measurement shift" (e.g., a new hospital scanner) by restoring the same features rather than learning new ones.

A Framework for the Quantitative Evaluation of Disentangled Representations

C Eastwood, CKI Williams

ICLR 2018

Code

We propose the DCI framework for evaluating "disentangled" representations.

(Previously Spotlight @ NeurIPS 2017 Disentanglement workshop)

Other Publications
Self-Supervised Disentanglement by Leveraging Structure in Data Augmentations

C Eastwood, J von Kügelgen, L Ericsson, D Bouchacourt, P Vincent, B Schölkopf, M Ibrahim

Preprint 2023

Code

We use data augmentations to disentangle rather than discard.

(Previously @ NeurIPS 2023 workshops on Self-Supervised Learning and Causal Representation Learning)

DCI-ES: An Extended Disentanglement Framework with Connections to Identifiability

C Eastwood*, A Nicolicioiu*, J von Kügelgen*, A Kekić, F Träuble, A Dittadi, B Schölkopf

ICLR 2023

Code

We extend the DCI framework by quantifying the ease-of-use or explicitness of a representation.

(Previously @ UAI 2022 Causal Repr. Learning Workshop)

Align-Deform-Subtract: An Interventional Framework for Explaining Object Differences

C Eastwood*, N Li*, CKI Williams

ICLR 2022 Workshop: Objects, Structure and Causality

We propose a framework for explaining object-image differences in terms of the underlying object properties.

Unit-Level Surprise in Neural Networks

C Eastwood*, I Mason*, CKI Williams

NeurIPS 2021 Workshop: I Can't Believe it's Not Better (Spotlight & Didactic Award) and PMLR

Code / Video

We use surprising unit-level activations to determine which parameters to adapt for a given distribution shift.

Learning Object-Centric Representations of Multi-Object Scenes from Multiple Views

N Li, C Eastwood, B Fisher

NeurIPS 2020 (Spotlight)

Code / Video

We learn accurate, object-centric representations of 3D scenes by aggregating information from multiple 2D views/observations.

Website source