Frontiers in Electrical Engineering
Monday, January 24, 2022
Event for Caltech Community Only -- Please contact Caroline Murphy for the location or Zoom information.
| Time (PST) | Speaker |
|---|---|
| 10:10 |
Welcome & Introduction Pietro Perona, Caltech |
| 10:15 |
Machine Learning on Large-Scale Graphs Luana Ruiz, University of Pennsylvania Graph neural networks (GNNs) are successful at learning representations from most types of network data but suffer from limitations in large graphs, which do not have the Euclidean structure that time and image signals have in the limit. Yet, large graphs can often be identified as being similar to each other in the sense that they share structural properties. Indeed, graphs can be grouped in families converging to a common graph limit -- the graphon. A graphon is a bounded symmetric kernel which can be interpreted as both a random graph model and a limit object of a convergent sequence of graphs. Graphs sampled from a graphon almost surely share structural properties in the limit, which implies that graphons describe families of similar graphs. We can thus expect that processing data supported on graphs associated with the same graphon should yield similar results. In my research, I formalize this intuition by showing that the error made when transferring a GNN across two graphs in a graphon family is small when the graphs are sufficiently large. This enables large-scale graph machine learning by transference: training GNNs on moderate-scale graphs and executing them on large-scale graphs. |
| 11:30 |
Computing Using Time Georgios Tzimpragos, UC Santa Barbara/Lawrence Berkeley Lab The development of computing systems able to address our ever-increasing needs, especially as we reach the end of CMOS transistor scaling, requires truly novel methods of computing. My research draws inspiration from biology, rethinks the digital/analog boundary, and challenges conventional wisdom, which typically guides how we perform computation, by reimagining the role of time. In this talk, I first introduce a computational temporal logic that sets the foundation for temporal computing. Second, I demonstrate how this foundation opens up unique ways in which we can work with sensors and design machine learning systems. Third, I describe how temporal operators provide answers to several long-lasting problems in computing with emerging devices. Finally, I touch upon future work with themes ranging from in-sensor online learning to hybrid quantum-classical computing and formally verifiable hardware. |
| 1:30 |
Causal Inference for Socio-economic and Engineering Systems *Anish Agarwal, MIT/Simons Institute UC Berkeley What will happen to Y if we do A? A variety of meaningful socio-economic and engineering questions can be formulated this way. To name a few: What will happen to a patient’s health if they are given a new therapy? What will happen to a country’s economy if policy-makers legislate a new tax? What will happen to a company’s revenue if a new discount is introduced? What will happen to a data center’s latency if a new congestion control protocol is used? In this talk, we will explore how to answer such counterfactual questions using observational data---which is increasingly available due to digitization and pervasive sensors---and/or very limited experimental data. The two key challenges in doing so are: (i) counterfactual prediction in the presence of latent confounders; (ii) estimation with modern datasets which are high-dimensional, noisy, and sparse. Towards this goal, the key framework we introduce is connecting causal inference with tensor completion, a very active area of research across a variety of fields. In particular, we show how to represent the various potential outcomes (i.e., counterfactuals) of interest through an order-3 tensor. The key theoretical results presented are: (i) Formal identification results establishing under what missingness patterns, latent confounding, and structure on the tensor is recovery of unobserved potential outcomes possible. (ii) Introducing novel estimators to recover these unobserved potential outcomes and proving they are finite-sample consistent and asymptotically normal. The efficacy of the proposed estimators is shown on high-impact real-world applications. These include working with: (i) TaurRx Therapeutics to propose novel clinical trial designs to reduce the number of patients recruited for a trial and to correct for bias from patient dropouts. (ii) Uber Technologies on evaluating the impact of certain driver engagement policies without having to run an A/B test. (iii) U.S. and Indian policy-makers to evaluate the impact of mobility restrictions on COVID-19 mortality outcomes. (iv) The Poverty Action Lab (J-PAL) at MIT to make personalized policy recommendations to improve childhood immunization rates across different villages in Haryana, India. Finally, we discuss connections between causal inference, tensor completion, and offline reinforcement learning. |
| 2:45 |
Understanding Statistical-vs-Computational Tradeoffs via Low-Degree Polynomials *Alexander Wein, UC Berkeley/Georgia Tech A central goal in modern data science is to design algorithms for statistical inference tasks such as community detection, high-dimensional clustering, sparse PCA, and many others. Ideally these algorithms would be both statistically optimal and computationally efficient. However, it often seems impossible to achieve both these goals simultaneously: for many problems, the optimal statistical procedure involves a brute force search while all known polynomial-time algorithms are statistically sub-optimal (requiring more data or higher signal strength than is information-theoretically necessary). In the quest for optimal algorithms, it is therefore important to understand the fundamental statistical limitations of computationally efficient algorithms. I will discuss an emerging theoretical framework for understanding these questions, based on studying the class of "low-degree polynomial algorithms." This is a powerful class of algorithms that captures the best known poly-time algorithms for a wide variety of statistical tasks. This perspective has led to the discovery of many new and improved algorithms, and also many matching lower bounds: we now have tools to prove failure of all low-degree algorithms, which provides concrete evidence for inherent computational hardness of statistical problems. This line of work illustrates that low-degree polynomials provide a unifying framework for understanding the computational complexity of a wide variety of statistical tasks, encompassing hypothesis testing, estimation, and optimization. |
Monday, February 14, 2022
Event for Caltech Community Only -- Please contact Caroline Murphy for the location or Zoom information.
| Time (PST) | Speaker |
|---|---|
| 8:55 |
Welcome & Introduction Pietro Perona, Caltech |
| 9:00 |
Unlocking Musical Expression with Machine Learning Chris Donahue, Stanford While we all possess sophisticated musical intuition, conventional tools for musical expression (e.g., sheet music, instruments) are inaccessible to those of us without formal training. Conventional tools thus give rise to an expertise divide which segregates us into "musicians" and "non-musicians" and prevents the latter from realizing any creative ambitions. In this talk, I will present my work on using machine learning to build assistive tools for musical expression—ones which bridge the expertise divide and allow non-musicians to quickly accomplish musical tasks like improvisation or performing their favorite song. To this end, a common theme of my work is using generative modeling to understand the relationships between music audio and other modalities like sheet music and physical gestures. Additionally, I build real-world interactive music systems which allow a broad audience to harness the power of resultant models. By using machine learning to understand music and its multimodal relationships, we stand to unlock the full potential of music as a universal language, allowing anyone to not only listen but also participate in the conversation. |
| 10:15 |
Theoretical Foundations of Pre-trained Models *Qi Lei, Princeton A pre-trained model refers to any model trained on broad data at scale and can be adapted (e.g., fine-tuned) to a wide range of downstream tasks. The rise of pre-trained models (e.g., BERT, GPT-3, CLIP, Codex, MAE) transforms applications in various domains and aligns with how humans learn. Humans and animals first establish their concepts or impressions from different data domains and data modalities. The learned concepts then help them learn specific tasks with minimal external instructions. Accordingly, we argue that a pre-trained model follows a similar procedure through the lens of deep representation learning. 1) Learn a data representation that filters out irrelevant information from the training tasks; 2) Transfer the data representation to downstream tasks with few labeled samples and simple models. This talk establishes some theoretical understanding for pre-trained models under different settings, ranging from supervised pretraining, meta-learning, and self-supervised learning to domain adaptation or domain generalization. I will discuss the sufficient (and sometimes necessary) conditions for pre-trained models to work based on the statistical relation between training and downstream tasks. The theoretical analyses partly answer how they work, when they fail, guide technical decisions for future work, and inspire new methods in pre-trained models. |
| 11:30 |
Making Sense of the Physical World with High-resolution Tactile Sensing Wenzhen Yuan, Carnegie Mellon Robotics Institute In this talk, I will introduce the development of a high-resolution robotic tactile sensor GelSight, and how it can help robots understand and interact with the physical world. GelSight is a vision-based tactile sensor that measures the geometry of the contact surface with a spatial resolution of around 25 micrometers, and it also measures the shear forces and torques at the contact surface. With the help of high-resolution information, a robot could easily detect the precise shape and texture of the object surfaces and therefore recognize them. But it can help robots get more information from contact, such as understanding different physical properties of the objects and assisting manipulation tasks. This talk will cover our previous works on object property perception and slip detection with GelSight sensors. I will also discuss our recent development in tactile sensor simulation and how we envision the work will help the community with the research in tactile sensing. |
*CMS co-applicants
Frontiers in Computing + Mathematical Sciences
February 7, 2022