Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Novelty-seeking Guides Formation of Disentangled Representations

Pingsheng Li¹; ¹EPFL - EPF Lausanne

Presenter: Pingsheng Li

Disentangled representations are common in the brain, where many neurons are tuned to single factors of task variation, such as place cells or object-vector cells. Previous work has shown that neural networks trained with biological constraints can also learn disentangled representations if trained on disentangled data, i.e., data generated by independent factors. However, in real-world, open-ended environments, such neatly disentangled data may not always be available. This raises a fundamental question: how can agents collect experiences that help them form disentangled representations? Intrinsic motivations, such as novelty, efficiently guide humans and artificial agents during exploration of unfamiliar environments but it is unclear whether they also support disentangled representation learning. Using a novel method to extract representation-specific novelty signals, we compute novelty signals from the latent representations of autoencoders (AEs) and discrete variational autoencoders (D-VAEs) and use them as intrinsic exploration rewards for an artificial agent performing unsupervised learning. We show that these novelty signals favor exploration of disentangled over entangled data, and help the agent learn disentangled representations.

Topic Area: Reward, Value & Social Decision Making

Extended Abstract: Full Text PDF