ML@GT Seminar Series

Series

ML@GT Seminar Series

Permanent Link

https://hdl.handle.net/1853/70935

Series Type

Event Series

Associated Organization(s)

Organizational Unit

Machine Learning Center

Organizational Unit

College of Computing

Full item page

Publication Search Results

Now showing 1 - 10 of 17

Automated Perception in the Real World: The Problem of Scarce Data

( 2018-11-30) Ernst, Jan

Machine perception is a key step toward artificial intelligence in domains such as self-driving cars, industrial automation, and robotics. Much progress has been made in the past decade, driven by machine learning, ever-increasing computational power, and the reliance on (seemingly) vast data sets. There are however critical issues in translating academic progress into the real world: available data sets may not match real-world environments well, and even if they are abundant and matching well, then interesting samples from a real-world perspective may be exceedingly rare and thus still be too sparsely represented to learn from directly. In this talk, I illustrate how we have approached this problem strategically as an example of industrial R&D from inception to product. I will also go in-depth on an approach to automatically infer previously unseen data by learning compositional visual concepts via mutual cycle consistency.
Multimodal, Personable, and Knowledgeable Language Generation

( 2018-11-19) Bansal, Mohit

In this talk, I will discuss my group's recent work on state-of-the-art natural language generation (NLG) and dialogue models that are multimodal, personality-based, and knowledge-rich. First, we will discuss dialogue models which generate responses that are not only history-relevant and fluent, but also multimodal, e.g., relevant to dynamic video-based context. Next, we will present personality-based conversational agents, e.g., models that generate stylistic responses with varying levels of politeness and rudeness. Finally, we will describe several directions in making NLG models more knowledgeable, e.g., via adversarial robustness to user errors, via filling reasoning gaps in multi-hop generative-QA with external commonsense knowledge, and via multi-task and reinforcement learning with novel auxiliary-skill tasks such as entailment and saliency generation.
The Statistical Foundations of Learning to Control

( 2018-11-14) Recht, Benjamin

Given the dramatic successes in machine learning and reinforcement learning over the past half decade, there has been a surge of interest in applying these techniques to continuous control problems in robotics and autonomous vehicles. Though such control applications appear to be straightforward generalizations of standard reinforcement learning, few fundamental baselines have been established prescribing how well one must know a system in order to control it. In this talk, I will discuss how one might merge techniques from statistical learning theory with robust control to derive such baselines for such continuous control. I will explore several examples that balance parameter identification against controller design and demonstrate finite sample tradeoffs between estimation fidelity and desired control performance. I will describe how these simple baselines give us insights into shortcomings of existing reinforcement learning methodology. I will close by listing several exciting open problems that must be solved before we can build robust, safe learning systems that interact with an uncertain physical environment.
Data-to-Decisions for Safe Autonomous Flight

(Georgia Institute of Technology, 2018-11-07) Atkins, Ella

Traditional sensor data can be augmented with new data sources such as roadmaps and geographical information system (GIS) Lidar/video to offer emerging unmanned aircraft systems (UAS) and urban air mobility (UAM) a new level of situational awareness. This presentation will summarize my group’s research to identify, process, and utilize GIS, map, and other real-time data sources during nominal and emergency flight planning. Specific efforts have utilized machine learning to automatically map flat rooftops as urban emergency landing sites, incorporate cell phone data into an occupancy map for risk-aware flight planning, and extend airspace geofencing into a framework capable of managing all traffic types in complex airspace and land-use environments. The presentation will end with videos illustrating recent work to experimentally validate the continuum deformation cooperative control strategy in the University of Michigan’s new M-Air netted flight facility.
Practical Applications of Signal Processing and Machine Learning in a Dynamic Retail Environment

( 2018-10-31) Poliner, Graham

The retail industry is the midst of rapid change due to intensifying competition from fragmented and non-traditional sources, expansion of assortment breadth and product availability, and more transparent pricing. Evolving consumer expectations have applied stress to traditional retailing approaches and existing networks. To adapt, the application of analytical methods has become pervasive across a wide range of retailing functions including marketing, merchandising, logistics, and pricing. In this talk, we will discuss methods to overcome emerging supply chain constraints, how cross-channel demand information can be used to inform localized assortment decisions, and the use of customer insights and visual learning to enable personalized shopping experiences.
Few-shot Learning with Meta-Learning: Progress Made and Challenges Ahead

( 2018-10-15) Larochelle, Hugo

A lot of the recent progress on many AI tasks enabled in part by the availability of large quantities of labeled data. Yet, humans are able to learn concepts from as little as a handful of examples. Meta-learning is a very promising framework for addressing the problem of generalizing from small amounts of data, known as few-shot learning. In meta-learning, our model is itself a learning algorithm: it takes input as a training set and outputs a classifier. For few-shot learning, it is (meta-)trained directly to produce classifiers with good generalization performance for problems with very little labeled data. In this talk, I'll present an overview of the recent research that has made exciting progress on this topic (including my own) and will discuss the challenges as well as research opportunities that remain.
Reaching Beyond Human Accuracy With AI Datacenters

( 2018-10-03) Diamos, Gregory

Deep learning has enabled rapid progress in diverse problems in vision, speech, healthcare, and beyond. This progress has been driven by breakthroughs in algorithms that can harness massive datasets and powerful compute accelerators like GPUs. In this talk, I will combine theoretical and experiment insights to help explain why deep learning scales predictably with bigger datasets and faster computers. I will also show how some problems are relatively easier than others, and how to tell the difference. I will show examples of important open problems that cannot be solved by small-scale systems but are within reach of the largest machines in the world. I will make the case for specializing datacenters to support AI applications using deep learning efficiently. I will outline a high-level architecture for such a design, and leave you with powerful tools to reach beyond human accuracy to confront some of the hardest open problems in computing.
Understanding the limitations of AI: When Algorithms Fail

( 2018-09-05) Gebru, Timnit

Automated decision-making tools are currently used in high stakes scenarios. From natural language processing tools used to automatically determine one’s suitability for a job, to health diagnostic systems trained to determine a patient’s outcome, machine learning models are used to make decisions that can have serious consequences on people’s lives. In spite of the consequential nature of these use cases, vendors of such models are not required to perform specific tests showing the suitability of their models for a given task. Nor are they required to provide documentation describing the characteristics of their models, or disclose the results of algorithmic audits to ensure that certain groups are not unfairly treated. Gebru will show some examples to examine the dire consequences of basing decisions entirely on machine learning based systems, and discuss recent work on auditing and exposing the gender and skin tone bias found in commercial gender classification systems. She will end with the concept of an AI datasheet to standardize information for datasets and pre-trained models, in order to push the field as a whole towards transparency and accountability.
The Natural Language Decathlon: Multitask Learning as Question Answering

( 2018-08-28) McCann, Bryan

Deep learning has improved performance on many natural language processing (NLP) tasks individually. However, general NLP models cannot emerge within a paradigm that focuses on the particularities of a single metric, dataset, and task. We introduce the Natural Language Decathlon (decaNLP), a challenge that spans ten tasks: question answering, machine translation, summarization, natural language inference, sentiment analysis, semantic role labeling, zero-shot relation extraction, goal-oriented dialogue, semantic parsing, and commonsense pronoun resolution. We cast all tasks as question answering over a context. Furthermore, we present a new Multitask Question Answering Network (MQAN) jointly learns all tasks in decaNLP without any task-specific modules or parameters in the multitask setting. MQAN shows improvements in transfer learning for machine translation and named entity recognition, domain adaptation for sentiment analysis and natural language inference, and zero-shot capabilities for text classification. We demonstrate that the MQAN's multi-pointer-generator decoder is key to this success and performance further improves with an anti-curriculum training strategy. Though designed for decaNLP, MQAN also achieves state of the art results on the WikiSQL semantic parsing task in the single-task setting. We release code for procuring and processing data, training and evaluating models, and reproducing all experiments for decaNLP.
Deep Learning to Learn

(Georgia Institute of Technology, 2018-08-20) Abbeel, Pieter

Reinforcement learning and imitation learning have seen success in many domains, including autonomous helicopter flight, Atari, simulated locomotion, Go, robotic manipulation. However, sample complexity of these methods remains very high. In contrast, humans can pick up new skills far more quickly. To do so, humans might rely on a better learning algorithm or on a better prior (potentially learned from past experience), and likely on both. In this talk I will describe some recent work on meta-learning for action, where agents learn the imitation/reinforcement learning algorithms and learn the prior. This has enabled acquiring new skills from just a single demonstration or just a few trials. While designed for imitation and RL, our work is more generally applicable and also advanced the state of the art in standard few-shot classification benchmarks such as omniglot and mini-imagenet.