MATRIX Spring Seminar Series – Dr. John Basl

An Ethics Ecosystem for AI and Big Data

John Basl

Northeastern University

https://utsa.zoom.us/j/92387759081

Friday, April 1, 2022

11 AM – 12 PM CST

Realizing the promise of AI while mitigating against ethical pitfalls requires an ethics ecosystem, an ecosystem that includes robust, interdisciplinary research programs, curricula, and training programs, informed regulatory and oversight infrastructure, broad stakeholder engagement, shared conceptual and linguistic space for overcoming challenges, and AI and data ethicists that can translate across disciplinary divides. This talk motivates the need for such an ecosystem and highlights some work attempting to contribute to it.

MATRIX Spring Seminar Series – Davis Sawyer

Creating new Opportunities for AI with Ultra Low-bit Precision Neural Networks

Davis Sawyer

Co-Founder of Deeplite Inc.

https://utsa.zoom.us/j/92387759081

Friday, April 8, 2022

11 AM – 12 PM CST

Abstract: Edge technologies have emerged as a major computing paradigm in recent years. From automated optical inspection to autonomous driving, the possibilities for edge AI to benefit daily life are growing faster than ever. Namely, the exponential growth of Deep Neural Networks (DNNs), maturing IoT deployment architectures, as well as economic and environmental benefits of AI inference at the point of data capture have driven tremendous academic and commercial interest in the field. However, the majority of high-value tasks like Natural Language Processing (NLP) and Computer Vision (CV) still require power-hungry or purpose-built hardware to achieve acceptable inference performance. This significantly limits the potential to bring AI to the billions of connected devices around us. To make edge AI more accessible and affordable, we introduce Deeplite Runtime (DeepliteRT), a novel software approach to quantize and run ultra low-bit DNNs on low-power, low-cost processors. In this talk, we will discuss the latest in training-aware quantization and share experimental results for DeepliteRT, where our ultra-low bit models achieved near-GPU level latency using a simple ARM Cortex-A CPU. We will conclude with a discussion on remaining challenges and future research directions.

MATRIX Spring Seminar Series – Dr. Rodrigo de Salvo Braz

Bean Machine

Rodrigo de Salvo Braz, PhD

Research Scientist at Meta (formerly Facebook)

https://utsa.zoom.us/j/92387759081

Friday, April 15, 2022

11 AM – 12 PM CST

Bean Machine (https://beanmachine.org/) (BM) is a new probabilistic programming system (https://research.facebook.com/blog/2021/12/introducing-bean-machine-a-probabilistic-programming-platform-built-on-pytorch/) recently beta-released as open-source by Meta (formerly Facebook). Probabilistic programming languages (PPLs) provide a way for experts to specify a Bayesian probabilistic model using a programming language, but without having to concern themselves with, or even know, how the probabilistic inference will be performed. While probabilistic programming is a growing field, Bean Machine provides a novel combination of features: experts can declaratively specify their models in Python (including the use of PyTorch tensors), use a variety of inference methods (including Hamiltonian Monte Carlo, No U-Turn Sample (NUTS) and Newtonian Monte Carlo), use composable inference to apply different methods to different parts of a model, work with hybrid models (containing both discrete and continuous variables) and, for a certain subset of models, use a very fast C++ inference layer. Bean Machine is already being used in production for Meta internal applications, but there are many exciting future plans for making it even better. In this talk I will give examples, describe applications, and discuss some of these future directions.

MATRIX Spring Seminar Series – Dr. Zsolt Kira

Handling Distribution Shift in Visual Learning

Dr. Zsolt Kira

Georgia Institute of Technology

https://utsa.zoom.us/j/92387759081

Friday, April 22, 2022

11 AM – 12 PM CST

Abstract: While deep learning has achieved remarkable computer vision successes, fundamentally both the theory and practice for these successes have relied on vanilla supervised learning where the training and testing datasets both are sampled from the same distribution. In reality, there is likely to be significant distribution shift once models are deployed, including noise/weather/illumination/modality changes (covariate shift), new categories (semantic shift), or different label distributions. In this talk, I will present our recent work focusing on a fundamental handling of several of these shifts. For label distribution shifts, we propose a posterior-recalibration of classifiers that can be applied without re-training to handle imbalanced datasets. For covariate and semantic shift, we propose a geometric decoupling of classifiers into feature norms and angles, showing that it can be used to learn more sensitive feature spaces for better calibration and out-of-distribution detection. We demonstrate state-of-art results across multiple benchmark datasets and metrics. At the end, I will present connections to a wider set of problems including continual/lifelong learning, open-set discovery, and semi-supervised learning.

LINK