Undergraduate Deep Learning Lab (UDLL)

The undergraduate Deep Learning Lab offers students the opportunity to fine-tune their deep learning skills while engaging in research projects. Current open projects include:

  • Deep Learning in hydrology. The hydrological sciences mainly use physics based models to study the water cycle. Recently, deep learning has shown promise as a complimentary approach for supporting scientific discovery. UDLL projects in hydrology seek to improve the speed, accuracy, robustness, and interpretability of deep hydrological models. Related papers include
    Prolongation of SMAP to spatiotemporally seamless coverage of continental US using a deep learning neural network.
  • Deep Learning in Astronomy. The field of astronomy is one of the top producers of big data. Images collected by telescopes can be contaminated with various noise such as cosmic rays and ground-based interference. UDLL projects in astronomy study how to process such noisy images.
  • Deep Learning Training Algorithms. The training of novel architectures is often complicated idiosyncrasies of algorithms such as back-propagation, which may require significant and time-consuming tuning. UDLL projects also include experimentation with alternative straining algorithms that may add to the robustness and speed of learning.
Program Verification for Differential Privacy

Differential Privacy is a name given to a collection of technologies that can provably protect the privacy of individuals who data are collected by companies and statistical government agencies. It is in use in companies such as Google and Apple, as well as agencies such as the U.S. Census Bureau. One of the big challenges is in ensuring that algorithms and their software implementations really do satisfy differential privacy. The goal of this project is to develop software tools for finding bugs or verifying that programs do satisfy differential privacy, along with the development of software engineering guidelines for differential privacy. Relevant publications include LightDP: Towards Automating Differential Privacy Proofs and Toward Detecting Violations of Differential Privacy.

Text in the Wild

Text appears in natural scene images in street signs, storefronts, house numbers, etc.  Automated methods for understanding such text can be useful for digital assistants (especially for the visually impaired) and to aid in automated scene understanding. At the same time, text in such images can be distorted due to occlusions, perspective, lighting, rotations, dirt and properties of the camera/lens.  The goal of this project is to use deep learning techniques to identify and extract text from images in real-time or near real-time. Relevant publications include • Learning to Read Irregular Text with Attention Mechanism, • Learning to Extract Semantic Structure from Documents Using Multimodal Fully Convolutional Neural Networks, • Multi-scale FCN with Instance Aware Segmentation for Arbitrary Oriented Word Spotting In The Wild, • Smart Library: Identifying Books in a Library using Richly Supervised Deep Scene Text Reading, • Aggregating Local Context for Accurate Scene Text Detection, • Detecting Arbitrary Oriented Text in the Wild with a Visual Attention Model.