Published in arxiv, 2021
We propose an inter-domain gradient matching objective that targets domain generalization by maximising the inner product between gradients from different domains. We also derive a simpler first-order algorithm named Fish that approximates the computation of second-order derivative.
Published in ICLR, 2021
We develop a novel contrastive framework for multimodal generative model learning, allowing us to train the model not just by the commonality between modalities, but by the distinction between ‘relatedand 'unrelated multimodal data.
Published in Neurips, 2019
We propose a mixture-of-experts multimodal variational autoencoder (MMVAE) for learning of generative models on modality pairs, including image-image and language-vision dataset.
Published in ECCV, 2018
In this paper, we introduce a novel Recurrent Neural Network-based algorithm for future video feature generation and action anticipation called feature mapping RNN.