UCLA and Amazon Announce 2024 Science Hub Awards

Launched October 2021, the Science Hub supports academic research, doctoral students, fellowships and outreach efforts in areas of mutual interest around AI and its benefits. The Science Hub’s goal is to find solutions that benefit society with particular attention to matters of bias, fairness, accountability and responsible AI. The hub seeks to foster collaborations between Amazon scientists and academic researchers across disciplines, including computer science, electrical and computer engineering, and mechanical and aerospace engineering.

The latest cohort of project investigators and their respective projects being supported is as follows:

Xiang “Anthony” Chen, UCLA Samueli associate professor of electrical and computer engineering: “Human-Centered Design of AI Systems to Support Drug Discovery”

Drug discovery is one of the most crucial scientific endeavors for advancing human health and well-being. Unfortunately, the process of discovering a new drug has historically been slow and costly, involving multiple phases of screening, synthesis and testing that can take more than 14 years and cost more than $1 billion. Recent advances in AI have begun to automate parts of this process.  However, developing AI models alone will not address the critical question of where and how to integrate AI into scientists’ workflows. In collaboration with the UCLA Drug Discovery Lab, this project explores the following research questions: First, what are the workflows, challenges and needs of scientists during the drug discovery process? Second, how can we leverage decades of human-computer interaction research on creativity support tools to enhance these workflows? Third, how can we translate design requirements into interactive AI systems that support scientists’ work? By addressing these questions, we aim to design and build a software prototype that utilizes a suite of AI models as foundational tools to assist experts in the early stages of drug discovery.

Juliet Beni Edgcomb, assistant professor-in-residence of psychiatry and biobehavioral sciences in the David Geffen School of Medicine at UCLA: “EHR Phenotypes for Detection of Suicide Prevention Interventions Delivered to Children in Emergency Settings”

While suicide is the second-leading cause of death for children 10 to 14 years old, tracking suicide prevention interventions using medical records remains challenging due to frequent documentation in non-standardized fields or narrative text. Not knowing whom, when and where suicide prevention interventions occur impedes progress toward targeted delivery to children most at risk. 

This project aims to develop systematic search queries, or computable phenotypes, to detect suicide prevention interventions by applying natural language processing, artificial intelligence and machine learning methods to discover data signals in electronic health records. Then, the project will examine whether the likelihood of individuals receiving an indicated intervention varies by race, ethnicity, age, sex and neighborhood vulnerability.

Songwu Lu, UCLA Samueli professor of computer science, “Sign-to-English: Live, Bidirectional Communications for Sign Language Users with AR Glasses”

This project involves designing, testing and implementing a compact mobile device solution to American Sign Language (ASL) users with hearing disabilities. It offers bidirectional ASL-to-English and English-to-ASL translations between a user with hearing disabilities and an English speaker. The signers wear the augmented reality glasses as everyday wear, run Sign-to-English on their smartphone and glasses and interacts with an English speaker. The team will exploit ASL linguistic features to simplify the model structures and improve accuracy and speed. The device’s design departs from deep-learning-based large models, and adopts two-level, hybrid models to function on mobile devices. Refined hierarchical bounding boxes at the low level will be used to extract user gesture features, and traditional AI/machine learning models at the high level for sign and sentence recognition. The model does not require a large training dataset, and leverages recent component solutions from graphics, vision and natural language processing to optimize its performance. Built on top of early prototype for 911 emergency call services, the team is working with local ASL communities to deploy the approach in a pizza-ordering business, video captioning for ASL users and SignGPT applications.

Blaise-Pascal Tine, UCLA Samueli assistant professor of computer science: “OpenGPU Support for Machine Learning”

The project focuses on creating a specialized Multi-Level Intermediate Representation dialect for the OpenGPU platform, an open-source graphics-processing unit architecture based on the Reduced Instruction Set Computer-V instruction set architecture. The main goal is to enable seamless integration with popular machine learning frameworks, such as TensorFlow and PyTorch, while offering optimized execution paths tailored to OpenGPU’s unique capabilities. Current open-source GPU solutions struggle to match the performance and compatibility of proprietary platforms like NVIDIA’s Compute Unified Device Architecture, thereby severely limits accessibility and innovation in academic and research environments. By creating an MLIR dialect customized for OpenGPU, this project aims to break down these barriers, providing a robust foundation for open, high-performance GPU computing in machine learning, especially for those without access to proprietary hardware.

Guy Van den Broeck, UCLA Samueli professor of computer science: “Tractable Deep Generative Models”

Core to many tasks in artificial intelligence is the application of logical reasoning. For example, in the context of language modelling, one might want to generate text conditional on a logical constraint, such as the absence of toxic language. Despite the advanced capabilities of modern deep generative models such as large language models, they lack the ability to efficiently and reliably reason in this way.

In this project, the team will explore techniques to scale up a class of deep generative models known as probabilistic circuits, which exploits structure in probability distributions to enable tractable logical and probabilistic reasoning. The researchers’ vision is to distill large-scale probabilistic circuits as tractable surrogate models, with the aim of building more trustworthy and reliable neuro-symbolic AI systems.

George Varghese, UCLA Samueli distinguished professor of computer science: “Trustworthy LLMs using Automated Reasoning”

Currently, it is challenging to use large language models to generate programs for critical infrastructure because LLM-generated programs sometimes have egregious errors. The project will explore the use of automated reasoning to help LLMs produce correct and verified code for critical environments, specifically programs for networking because network infrastructure is crucial for e-commerce, and errors in networks often led to widespread cloud failures. The team proposes to use LLMs to synthesize configurations (currently done manually at Amazon) and to verify these configurations using a new paradigm called Verified Prompt Programming. The researchers also propose using LLMs to generate verified models for testing and simulation of protocol implementations like Amazon’s Route 53 Domain Name System code.

Lin Yang, UCLA Samueli assistant professor of electrical and computer engineering: “Adaptive Rehabilitation Assignment Design via AI-based Agent”

The project aims to innovate in the field of physical rehabilitation by developing an extended reality-mediated AI systems using safety-critical reinforcement learning. The team will design an RL-based assignment-design agent to design physical exercises patients need to perform as rehabilitative training. These interactions are rendered visually and tangibly through an XR device and a robot arm guided by the RL-based agent.

Bolei Zhou, UCLA Samueli assistant professor of computer science: “Towards Safe, Controllable and Interpretable Multi-Modal Generation”

Multi-modal generative-AI models such as DALL-E, Imagen, Stable Diffusion and Sora successfully generate photo-realistic model images and videos from text descriptions. However, large amounts of paired data and training auxiliary modules are required to enable controllable generation like ControlNet. This project aims to develop a scalable framework that utilizes the interpretable knowledge in the learned representations of multi-modal generative models to achieve safe and controllable generation. The proposed method will discover the editing direction in the hidden space and distill the interpretable knowledge for controllable generation. This approach will enable fine-grained control over various aspects of the generated content, including style, semantics and safety, and bring AI alignment with human preferences and societal values.