In this TechTalk, we propose a novel hierarchical generative model with a simple Markovian structure and a corresponding inference model. Both the generative and inference models are trained using the adversarial learning paradigm. We demonstrate that the hierarchical structure supports the learning of progressively more abstract representations as well as providing semantically meaningful reconstructions with different levels of fidelity.
Furthermore, we show that minimizing the Jensen-Shanon divergence between the generative and inference network is enough to minimize the reconstruction error. The resulting semantically meaningful hierarchical latent structure discovery is exemplified on the CelebA dataset. There, we show that the features learned by our model in an unsupervised way outperform the best handcrafted features. Furthermore, the extracted features remain competitive when compared to several recent deep supervised approaches on an attribute prediction task on CelebA.
Finally, we leverage the model's inference network to achieve state-of-the-art performance on a semi-supervised variant of the MNIST digit classification task.
Negar Rostamzadeh
Negar Rostamzadeh is a Research Scientist at Element AI. Her areas of interests are Machine Learning (particularity deep learning approaches) applied to multimedia problems (mainly video understanding).
Negar received her bachelor's degree in Computer Science from the University of Tehran. She started her Ph.D. at the Mhug (Multimedia and Human understanding) group, University of Trento, Italy. There she did research under the direction of Prof. Nicu Sebe. She worked as a research intern at the MMV (Multimedia and Vision) lab at the Queen Mary University of London, where she was supervised by Prof. Yiannis Patras. Negar spent more than 2 years of her Ph.D. at the MILA (Montreal Institute for Learning Algorithms) lab under the supervision of Prof. Aaron Courville. She was a Research Intern in the Research and Machine Intelligence group at Google (Seattle) in summer 2016. After her Google internship, she came back to MILA to continue her research on Video Understanding under the supervision of Prof. Aaron Courville. She finished her Ph.D. in April 2017. She is currently working at Element AI as a Research Scientist.
She was a co-founder of Women in Deep Learning (WiDL) workshop in 2016, organizer of Women in Machine Learning (WiML) workshop at NIPS 2017, Women in Computer Vision (WiCV) workshop at CVPR 2017, and Women in Deep Learning workshop at MILA deep learning summer school 2017.
