Stella Yu

Learning Selective Invariance Upon Parametric Density Functions of Transformations

Researchers Utkarsh Singhal (s.utkarsh@berkeley.edu) Stella Yu (stellayu@berkeley.edu) Ameesh Makadia (makadia@google.com) Carlos Esteves (machc@google.com) Overview

Symmetry is a powerful tool in the deep learning repertoire. Natural data displays structured variations due to naturally occurring symmetries, so modeling these symmetries greatly...

Scene Sketch to Photo Synthesis

Overview

Online shopping has been made easier with Augmented Reality. Through Amazon’s View in your room, you can see how products fit and look in your home before you bring them home. We propose to develop scene understanding techniques that could lead to even more advanced shopping features such as: virtually redecorating a home by removing or replacing existing objects, suggesting items based on the room style, suggesting different furniture layout styles.

Specifically, we leverage our expertise on 2D and 3D scene parsing, sketch-to-photo synthesis, and...

Hierarchical Model-Based Reinforcement Learning with Temporal Abstractions

For long-horizon tasks in the real world, humans accumulate information over time and make inference using stored memory. An abstract transition model of the world at various time resolutions allows reasoning both globally and locally. Model-based reinforcement learning (MBRL) is efficient for learning policies for robotics control tasks. However, current MBRL approaches only predict future trajectories in a limited time span. A major reason is that the model only learns at one time resolution, typically at the environment transition frequency....

Unsupervised Learning of Visual Context from Instance Segmentation

Unsupervised representation learning aims to extract latent information from data that reflects their semantic categories. Contrastive learning has emerged as a direct winning alternative to self-supervised learning. We take a step further and propose to extract latent information that also reflects their visual context.

Overview

Instead of...

Semantic Depth Extraction for Indoor Scenes

We investigated methods for combining the strengths of semantic (instance) segmentation and monocular depth estimation to predict sharp, high-quality relative depth maps for Augmented Reality applications.

Researchers Utkarsh Singhal, UC Berkeley, s.utkarsh@berkey.edu Stella Yu, UC Berkeley, stellayu@berkeley.edu Chun Kai Wang, Amazon,...

Safe and Sound: Learning Locomotion Skills Across Robot Morphology

Learning locomotion policies on real robots can be risky. Free robot exploration for policy learning in the real world is necessary but dangerous as it may cause catastrophes for the robots, especially for large, heavy and complicated robots. In this work, we aim at learning policies for robotic locomotion tasks for risky robots with the goal of minimizing the interaction of these risky robots with the environments. Though learning on large, heavy and complicated robots tends to be risky, learning policies on smaller, lighter and simpler robots has much lower cost and risk. We...