Please note: This PhD seminar will take place in DC 2314 and online.
Marvin Pafla, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Kate Larson, Mark Hancock
Unsupervised object-centric representation learning offers a promising way for machine learning models to interpret and manipulate complex scenes by decomposing them into discrete object representations, thereby enhancing out-of-distribution generalization and explainability. Most existing approaches rely on slot- and attention-based architectures that factor these representations by object, with more recent work introducing causal mechanisms to predict how they evolve in latent space. However, ensuring stability often involves freezing the modules that generate these representations and using straight-through gradients to address non-stationarity in mechanism selection.
To address these limitations, I present preliminary results of a controller-based approach in which simple causal mechanisms are dynamically selected, shaping how object-centric representations form so the correct mechanism can “explain” future sensory input. This framework aligns with philosophical perspectives on affordances—where an object’s identity is inseparable from its potential uses—enabling a more conceptually flexible understanding in AI by replacing ground-truth labels with potentially trainable causal mechanisms.
To attend this PhD seminar in person, please go to DC 2314. You can also attend virtually on MS Teams.