Graduate students who interned in 2016 will discuss their (non-NDA-blocked) research. Speakers include Olivia Simpson, Mengting Wan, Sharad Vikram, and Julaiti Alafate

When rewards are sparse and action spaces large, Q-learning with ϵ-greedy exploration can be inefficient. This poses problems for otherwise promising applications such as task-oriented dialogue systems, where the primary reward signal, indicating successful completion of a task, requires a complex sequence of appropriate actions. Under these circumstances, a randomly exploring agent might never stumble upon a successful outcome in reasonable time. We present two techniques that significantly improve the efficiency of exploration for deep Q-learning agents in dialogue systems. First, we introduce an exploration technique based on Thompson sampling, drawing Monte Carlo samples from a Bayes-by-backprop neural network, demonstrating marked improvement over common approaches such as ϵ-greedy and Boltzmann exploration. Second, we show that spiking the replay buffer with experiences from a small number of successful episodes, as are easy to harvest for dialogue tasks, can make Q-learning feasible when it might otherwise fail

For the last several years, we have witnessed the emergence of datasets of an unprecedented scale across different scientific disciplines. The large volume of such datasets presents new computational challenges as the diverse, feature-rich, and usually high-resolution data does not allow for effective data-intensive inference. In this regard, data summarization is a compelling (and sometimes the only) approach that aims at both exploiting the richness of large-scale data and being computationally tractable; Instead of operating on complex and large data directly, carefully constructed summaries not only enable the execution of various data analytics tasks but also improve their efficiency and scalability.

A systematic way for data summarization is to turn the problem into selecting a subset of data elements optimizing a utility function that quantifies “representativeness” of the selected set. Often-times, these objective functions satisfy submodularity, an intuitive notion of diminishing returns stating that selecting any given element earlier helps more than selecting it later. Thus, many problems in data summarization require maximizing submodular set functions subject to cardinality and massive data means we have to solve these problems at scale.

In this talk, I will present our recent efforts in developing practical schemes for data summarization. In particular, I will first discuss Stochastic-Greedy, currently the fastest centralized algorithm for data summarization whose query complexity is only linear in data size. However, to truly summarize data at scale we need to opt for streaming or distributed solutions. To this end, I will then present Sieve-Streaming, the first streaming algorithm for data summarization with a constant-factor approximation guarantee. Finally, I will talk about GreeDi, the first distributed approach towards data summarization. In our extensive experiments, we demonstrate the effectiveness of our approach on several applications, including sparse Gaussian process inference and exemplar-based clustering, on tens of millions of data points using Apache-Spark

The Robotics and Healthcare Engineering Lab at UCSD works to enable robots to sense, learn, and adapt to real people in the real world. We work on creating new algorithms for robots that leverage coordination dynamics and contextual processing to enable them to be robust and independent in the long term. Applications of this work include neurorehabilitation, advanced manufacturing, and home health. In this talk I will describe some of our lab's recent work.

A number of problems in statistical information processing and data analysis model the data as a Wide-Sense Stationary (WSS) time series, whose power spectrum (or equivalently, the covariance matrix) often acts as a sufficient statistic. Popular examples include speech processing, source localization in radar and neural signal processing, wireless communication and so forth. The covariance matrix of such data exhibit a Toeplitz structure, which can be leveraged to design highly efficient compressive samplers to reduce the dimension of high dimensional WSS data, without requiring it to have a sparse representation. Unlike existing results in Compressed Sensing, the goal here is to recover the high dimensional covariance matrix (or infer parameters of interest from it), instead of reconstructing the data itself.

Inspired by our past work on nested sensor arrays, I will describe a new WSS data sketching technique, known as the "Generalized Nested Sampler" (GNS), to acquire compressive measurements in such a way that it becomes possible to perfectly reconstruct the original high dimensional covariance matrix from these compressed sketches. I will focus on low-rank Toeplitz covariance matrices and develop an efficient GNS-based sampler which allows the recovery of a rank-r Toeplitz covariance matrix from a compressed sketch of size O(√r) x O(√r), where the size of the sketch has no dependence on the ambient large dimension N. Our reconstruction technique will use a regularizer-free framework, combined with the ability to extrapolate additional “unobserved” entries of the NxN covariance matrix. The algorithm has significantly lower computational complexity (that does not scale with N) compared to recent nuclear-norm based compressive covariance estimators, and is provably robust against bounded errors. Finally, I will consider the special case of rank-1 and sparse covariance matrices that arise in the important problem of “phase retrieval” in optical imaging. The role of 2nd order difference sets in complex phase retrieval will be demonstrated, inspiring the design of a new class of non-uniform Fourier based sampler, that can provably recover a complex signal (upto a global phase ambiguity) from its amplitude measurements with near-minimal number of samples. An interesting connection with the 4N-4 conjecture will also be established, which hypothesis 4N-4 to be the minimum number of generic measurements necessary to ensure injectivity in N dimensions for complex phase retrieval.

Joint work with Heng Qiao.