2019/11/14 by Matthew J. Vowels, Vowels, Matthew J., Necati Cihan Camgöz +3 · 2 citations
Computer Science · Neuroscience · #Face recognition and analysis #Face Recognition and Perception
paper · pdf · doi:10.48550/arxiv.1911.06443
Variational AutoEncoders (VAEs) provide a means to generate representational\nlatent embeddings. Previous research has highlighted the benefits of achieving\nrepresentations that are disentangled, particularly for downstream tasks.\nHowever, there is some debate about how to encourage disentanglement with VAEs\nand evidence indicates that existing implementations of VAEs do not achieve\ndisentanglement consistently. The evaluation of how well a VAE's latent space\nhas been disentangled is often evaluated against our subjective expectations of\nwhich attributes should be disentangled for a given problem. Therefore, by\ndefinition, we already have domain knowledge of what should be achieved and yet\nwe use unsupervised approaches to achieve it. We propose a weakly-supervised\napproach that incorporates any available domain knowledge into the training\nprocess to form a Gated-VAE. The process involves partitioning the\nrepresentational embedding and gating backpropagation. All partitions are\nutilised on the forward pass but gradients are backpropagated through different\npartitions according to selected image/target pairings. The approach can be\nused to modify existing VAE models such as beta-VAE, InfoVAE and DIP-VAE-II.\nExperiments demonstrate that using gated backpropagation, latent factors are\nrepresented in their intended partition. The approach is applied to images of\nfaces for the purpose of disentangling head-pose from facial expression.\nQuantitative metrics show that using Gated-VAE improves average\ndisentanglement, completeness and informativeness, as compared with un-gated\nimplementations. Qualitative assessment of latent traversals demonstrate its\ndisentanglement of head-pose from expression, even when only weak/noisy\nsupervision is available.\n