Discovery of Latent 3D Keypoints and Imposition of Prior in Deep Learning

We are back for our first meetup in 2019 ! This time we are pleased to have two speakers from VISTEC, Dr. Supasorn Suwajanakorn and Dr. Nat Dilokthanakul. Dr. Supasorn will present his NIPS paper "Discovery of Latent 3D Keypoints via End-to-end Geometric Reasoning" and Dr. Nat will give an introduction about the imposition of prior knowledge in Deep Learning.

Agenda:

6:00 PM - 6:30 PM | Registration

6:30PM - 7:30 PM | Dr. Supasorn Suwajanakorn, Lecturer at VISTEC
Discovery of Latent 3D Keypoints via End-to-end Geometric Reasoning

7:30 PM - 8:30 PM | Dr. Nat Dilokthanakul, Postdoctoral Researcher at VISTEC
Imposition of prior knowledge in Deep Learning

8:30 PM - 9:00 PM | Networking

--------------------------------------------------------
Topic 1 : Discovery of Latent 3D Keypoints via End-to-end Geometric Reasoning
Speaker: Dr. Supasorn Suwajanakorn

Abstract of the talk:
This paper presents KeypointNet, an end-to-end geometric reasoning framework to learn an optimal set of category-specific 3D keypoints, along with their detectors.
Given a single image, KeypointNet extracts 3D keypoints that are optimized for a downstream task. We demonstrate this framework on 3D pose estimation by proposing a differentiable objective that seeks the optimal set of keypoints for recovering the relative pose between two views of an object. Our model discovers geometrically and semantically consistent keypoints across viewing angles and instances of an object category. Importantly, we find that our end-to-end framework using no ground-truth keypoint annotations outperforms a fully supervised baseline using the same neural network architecture on the task of pose estimation. The discovered 3D keypoints on the car, chair, and plane categories of ShapeNet [6] are visualized at keypointnet.github.io

Speaker Bio: http://www.supasorn.com/

--------------------------------------------------------
Topic 2 : Imposition of prior knowledge in Deep Learning
Speaker: Dr. Nat Dilokthanakul

Abstract of the talk:
Deep Learning is a data-driven approach. Deep Learning practitioners often rely solely on their data to arrive at the best model’s parameters. The model architectures and the loss functions, however, are often chosen from the best results or from the best benchmark performances. Prior knowledge that is encapsulated in these chosen network architectures often are only implicitly discussed. In this talk, I want to bring the importance of prior knowledge to your attention. Especially, how data-driven approaches can be made more data efficient using prior knowledge to limit the possible parameter search-space. I will discuss some of the works that I have done in the field of deep reinforcement learning and deep generative model. Focusing on how prior knowledge imposition can be used to improve data efficiency in deep reinforcement learning and how prior knowledge imposition can be used to give structure to representation learning algorithms such as VAE to be more interpretable.

Speaker bio: https://www.doc.ic.ac.uk/~nd1214/