Events Listing

Get ready to be inspired by the creativity and ingenuity of over 300 talented students! This exciting event features an incredible range of projects from first-year explorers to senior-year trailblazers in Electrical, Computer, and Mechanical Engineering. Discover the boundless possibilities as you explore autonomous swarms, the Internet of Things, vision processing, and more engineering marvels. Don't miss this chance to witness Design Driving Discovery!

An information theory for out-of-order information: applications in DNA data storage and genomics

The recent development of DNA-based data storage prototypes has raised several questions about how to optimally encode information in these systems. A distinguishing feature of this new storage paradigm is that the stored information is read via “shotgun” sequencing technologies. This means that the channel output comprises many short fragments of the input observed out of order. Motivated by this, we study the capacity of a class of “shuffling channels” that capture this inherent need to reorder the observed channel output. We then extend our insights to other settings where lack of ordering is a key feature. We propose an information-theoretic framework for the problem of pairwise sequence alignment and use that to derive new practical algorithms with state-of-the-art performance. Finally, motivated by the problem of dataset matching, we study the fundamental limits of reference-based data reordering.

FPGA Accelerators in the Cloud

With recent dramatic advances in Field Programmable Gate Arrays (FPGAs), these devices are being used along with multi-core and novel memory technologies to realize advanced platforms to accelerate applications in the Cloud. We will review advances in reconfigurable computing over the past 25 years leading up to accelerators for data science. We will illustrate FPGA-based parallel architectures and algorithms for a variety of data analytics kernels in streaming graph processing and graph machine learning. While demonstrating algorithm-architecture co-design methodology to realize high performance accelerators for graphs and machine learning, we demonstrate the role of modeling and algorithmic optimizations to develop highly efficient Intellectual Property (IP) cores for FPGAs. We show improved performance for two broad classes of graph analytics: iterative graph algorithms with variable workload (e. g., graph traversal, shortest paths, etc.) and machine learning on graphs (e. g., graph embedding). For variable workload iterative graph algorithms, we illustrate dynamic algorithm adaptation to exploit heterogeneity in the architecture. We conclude by identifying opportunities and challenges in exploiting emerging heterogeneous architectures composed of multi-core processors, FPGAs, GPUs, NPUs and coherent memory.

Towards resource-aware federated learning with interactive debugging solutions

This talk explores two pivotal advancements in Federated Learning (FL) systems aimed at addressing resource management, client reliability, and debugging challenges. First, I present FLOAT, a framework designed to tackle resource heterogeneity and performance inconsistencies in FL by dynamically optimizing client resource utilization, reducing dropouts, and enhancing model convergence. FLOAT leverages multi-objective Reinforcement Learning with Human Feedback to automate optimization selection based on client conditions, showing a significant boost in model accuracy and efficiency. Next, I introduce FedDebug, a fault localization framework that enables effective debugging in FL environments. By integrating record-and-replay techniques for real-time inspection and adapting differential testing with neuron activation analysis, FedDebug identifies clients causing performance degradation with high accuracy, all without additional testing data or labels. 

Making Human-Centered AI for Mental Health Prediction in Social Media Data

Machine learning and AI are now at the forefront of technological interest in solving socially challenging problems, like understanding and intervening in dangerous mental illness behaviors. There is an urgent need to innovate data-driven systems to handle the volume and risk of this content in social networks and its contagion to others in the community. However, traditional approaches to prediction have mixed success, partly because technical solutions oversimplify complex behavior for technical tractability, and these problems are uniquely human and messy. This is exacerbated by recent press indicating that technology can be harmful for well-being. 

In this talk, I will argue that we need human-centered AI as a lens to make socially complex AI tech more technically rigorous and accurate, ethical, and compassionate for the people it impacts. My approach to this problem combines my disciplinary training in Media Studies and Computer Science, drawing on social science theory for more informed technological innovation. To unpack the transformative potential of human-centered AI, we’ll look at my group’s work in mental illness and social media as a case study.

Artificial intelligence for advancing cancer, nutrition and aging research

In this seminar, Professor Rui Zhang will provide an overview of his ongoing research projects aimed at advancing clinical research across various domains developing innovative artificial intelligence (AI) methods. Focusing on the cancer domain, Zhang will introduce cancer-domain language models developed to extract cancer phenotypes. This advancement holds significant implications for predicting cardiotoxicity related to cancer treatment. Shifting to the nutrition domain, the seminar will delve into Zhang's foundational work in establishing a knowledge base for dietary supplements. Additionally, he will showcase AI methodologies employed to extract efficacy and safety information from diverse sources, contributing to a comprehensive understanding of dietary supplement safety. In the aging domain, Zhang will discuss the applications of AI to repurpose Complementary and Integrative Health (CIH) approaches for Alzheimer's disease. The seminar will also highlight their recent work advancing large language models on multiple informatics tasks. In addition, Zhang will briefly introduce the Division of Computational Health Sciences in the Medical School and potential collaboration opportunities.

Imaging at the Speed of Neural Computation: An Electro-Optical Multiphoton Microscope

Neuroscientists have recently made impressive gains in developing fluorescent sensors to track the transmission of signals across neurons at millisecond timescales. However, multiphoton microscopes for in vivo imaging technologies have reached fundamental inertial limits on speed. In this talk we will examine the development of multiphoton microscopes utilizing novel electro-optical deflector (EOD) technology to achieve high resolution imaging at ultrafast speeds. While EODs with nanosecond response times have been commercially available for decades, their small angle of deflection limits them to very few resolvable points. We examine multiple novel approaches to develop an EOD that is suitable for ultrafast imaging, including large-aperture deflectors and cavity-based angle multiplication. We will also discuss our prototype microscope that is, to our knowledge, the world’s first fully electro-optical multiphoton microscope. 

Language model alignment: theory and practice

The goal of language model alignment (post-training) is to draw samples from an aligned model that improve a reward (e.g., safety or factuality) with little perturbation to the base model. A simple baseline for this task is best-of-N, where N responses are drawn from the base model, ranked based on a reward, and the highest ranking one is selected. More sophisticated techniques generally solve a KL-regularized reinforcement learning (RL) objective with the goal of maximizing expected reward subject to a KL divergence constraint between the aligned model and the reference model. 

In this talk, we give an overview of language model alignment and give an understanding of key results in this space through simplified examples. We also present a modular alignment technique, called controlled decoding, which solves the KL-regularized RL problem while keeping the reference model frozen through learning a prefix scorer, offering inference-time configurability. Finally, we also shed light on the remarkable performance of best-of-N in terms of achieving competitive or even better reward-KL tradeoffs when compared to state-of-the-art alignment baselines.

On the performance of ranking recovery algorithms with privacy consideration

Today, ranking algorithms are of fundamental importance and are used in a wide variety of applications, such as recommender systems and search engines. Broadly speaking, the goal of a ranking algorithm is to sort a dataset so that users are provided with accurate and relevant results. Although modern ranking algorithms promise efficient means of performing large-scale data processing, there are numerous privacy considerations that must not be overlooked. For instance, a user would not like to disclose their previous purchases to a recommender system. 

In this talk, we consider the private ranking recovery problem, which consists of recovering the ranking/permutation of an input data vector from a noisy version of it. We aim to establish fundamental trade-offs between the performance of the estimation task, measured in terms of probability of error, and the level of privacy that can be guaranteed when the noise mechanism consists of adding artificial noise.