Contrastive predictive coding for video representation learning

Guillaume Lorre (1) , Jaonary Rabarisoa (1) , Astrid Orcesi (1) , Samia Ainouz (2) , Stéphane Canu (2)

1 DIASI - Département Intelligence Ambiante et Systèmes Interactifs
2 LITIS - Laboratoire d'Informatique, de Traitement de l'Information et des Systèmes

Contrastive Predictive Coding (CPC) (van den Oord et al., 2018) has been successfully used to learn representations for different signals (audio, text, images). It uses an autoregressive modeling and contrastive estimation to learn long-term temporal relation inside the raw signal while remaining robust to local noise. The result is a higher level signal representation useful to solve downstream tasks. Using CPC to learn representations for videos remains challenging due to the structure and the high dimensionality of the signal. In this work, we propose different implementations of CPC for video signal. The learned representation increases the performance of an action recognition classifier.

Domains: Artificial Intelligence [cs.AI]

Guillaume Lorre, Jaonary Rabarisoa, Astrid Orcesi, Samia Ainouz, Stéphane Canu. Contrastive predictive coding for video representation learning. ICML2019 - 36th International Conference on Machine Learning - Workshop on Self-Supervised Learning, Jun 2019, Long Beach, United States. ⟨cea-03547497⟩