The increasing popularity of smart devices that continuously monitor various aspects of users' life and the prevalence of third-party services that utilize these data feeds have resulted in a serious threat to users' privacy. One-sided focus on the utility of these applications (apps) and lack of proper access control mechanism often lead to inadvertent (or deliberate) leak of sensitive information about users. At the core of protecting user data on smart devices lies the permissions framework. It arbitrates apps' accesses to resources on the device.
In classic supervised machine learning, a learning agent behaves as a passive observer: it receives examples from some external environment which it has no control over and then makes predictions. Reinforcement Learning (RL), on the other hand, is fundamentally interactive : an autonomous agent must learn how to behave in an unknown and possibly hostile environment, by actively interacting with the environment to collect useful feedback.
Online transaction processing (OLTP) is critical for applications including finance, e-commerce, social networks, and healthcare. The increasing performance demands of these applications require OLTP to scale massively. Concurrency control is a major scalability bottleneck in such systems.
The notion of distance plays and important role in many reinforcement learning (RL) techniques. This role may be explicit, as in some non-parametric approaches, or it may be implicit in the architecture of the feature space. The ability to learn distance functions tailored for RL tasks could, thus, benefit many different RL paradigms.
Transactional key-value storage is an important service offered by cloud service providers for building applications (e.g., Amazon DynamoDB, Microsoft CosmosDB, Google Spanner). This type of service is popular because it provides high-level guarantees like consistency, scalability and fault-tolerance to ease application development and deployment on the cloud.
Neural predictive coding is an enormously successful approach to unsupervised representation learning in natural language processing. In this approach, a large-scale neural language model is trained to predict the missing signal (e.g., next word, next sentence) and the trained model is used in downstream tasks to produce useful text representations. While effective, it is computationally difficult to work with and yields uninterpretable representations.
The theory of approximate optimization faces new challenges and opportunities thanks to its increasing interaction with other fields of optimization. Such growth and diversity call for a unified theory of approximate optimization, where various algorithms and complexity results can be connected to one another and explained by a few general principles. My research aims to contribute to such a theory, by building a unified theory for general classes of optimization problems and exploring connections between such classes. This talk will showcase results in both directions.
Correctness and security problems in modern computer systems can result from problematic hardware event orderings and interleavings during an application’s execution.
In public decision making, we are confronted with the problem of aggregating the conflicting preferences of many individuals about outcomes that affect the group. Examples of public decision making include allocating shared public resources and social choice or voting. We study these problems from the perspective of an algorithm designer who takes the preferences of the individuals and the constraints of the decision making problem as input and efficiently computes a solution with provable guarantees with respect to fairness and welfare, as defined on individual preferences.
Cloud computing plays a critical role in providing computing resources to many organizations. The relentless of the need for cloud service makes reliability and efficiency two primary metrics of interest. However, the existing data center system design falls short on these two goals. Specifically, (1) operating systems have significant overheads in providing virtualization support to cloud applications; (2) network infrastructure incurs excessive cost; (3) infrastructure problems are notoriously difficult to debug and mitigate.
Degenerative retinal diseases such as retinitis pigmentosa and macular degeneration cause irreversible vision loss in more than 10 million people worldwide. Analogous to cochlear implants, retinal prostheses use a grid of electrodes to stimulate surviving retinal cells in order to evoke visual percepts. However, a major outstanding challenge in the use of these devices is translating electrode stimulation into a code that the visual system can interpret.
Learning from interaction with the environment -- trying untested actions, observing successes and failures, and tying effects back to causes -- is one of the first capabilities thought of when considering intelligent agents. Reinforcement learning is the area of artificial intelligence research that has the goal of allowing autonomous agents to learn in this way. Despite many recent empirical successes, most modern reinforcement learning algorithms are still limited by the large amounts of experience required before useful skills are learned.
In this talk, I will overview our recent progress towards understanding how we learn large capacity machine learning models. In the modern practice of machine learning, especially deep learning, many successful models have far more trainable parameters compared to the number of training examples. Consequently, the optimization objective for training such models have multiple minimizers that perfectly fit the training data.
In this project, we consider the problem of determining whether the truthfulness of a claim can be inferred from a database of fact-checks done by reputable organizations such as factcheck.org, PolitiFact, and Washington Post. The problem can be modeled as a textual entailment problem in natural language inference. We study how to adapt the Transformer, a self-attention RNN model, for our task. A key challenge we need to deal with is the lack of training data in our domain.
As our media becomes increasingly flooded with false or misleading claims, fact-checking is more important than ever. To help stop misinformation right at the point of consumption, we aim to provide a service that detects, in real time, any claim heard by a user that has been previously fact-checked. One specific challenge we address in this project is building a reliable and cost-effective infrastructure to meet the stringent bandwidth and latency requirements of real-time processing. We evaluate various implementation options, including GPUs and clusters in public cloud.
How can we intelligently acquire information for decision making, when facing a large volume of data? In this talk, I will focus on learning and decision making problems that arise in robotics, scientific discovery and human-centered systems, and present how we can develop principled approaches that actively extract information, identify the most relevant data for the learning tasks and make effective decisions under uncertainty.
Natural language processing (NLP) has come of age. For example, semantic role labeling (SRL), which automatically annotates sentences with a labeled graph representing “who” did “what” to “whom,” has in the past ten years seen nearly 40% reduction in error, bringing it to useful accuracy. As a result, a myriad of practitioners now want to deploy NLP systems on billions of documents across many domains. However, state-of-the-art NLP systems are typically not optimized for cross-domain robustness nor computational efficiency.
As the proliferation of sensors rapidly make the Internet-of-Things (IoT) a reality, the devices and sensors in this ecosystem—such as smartphones, video cameras, home automation systems and autonomous vehicles—constantly map out the real-world producing unprecedented amounts of connected data that captures complex and diverse relations. Unfortunately, existing big data processing and machine learning frameworks are ill-suited for analyzing such dynamic connected data, and face several challenges when employed for this purpose.
While state-of-the-art machine learning models are deep, large-scale, sequential and highly nonconvex, the backbone of modern learning algorithms are simple algorithms such as stochastic gradient descent, or Q-learning (in the case of reinforcement learning tasks). A basic question endures---why do simple algorithms work so well even in these challenging settings?
The first camera phone was sold in 2000, when taking pictures with your phone was an oddity, and sharing pictures online was unheard-of. Today, barely twenty years later, the smartphone is more camera than phone. How did this happen? This transformation was enabled by advances in computational photography -- the science and engineering of making great images from small form factor, mobile cameras. Modern algorithmic and computing advances changed the rules of photography, bringing to it new modes of capture, post-processing, storage, and sharing, to make images ubiquitous across the world.