Special Student Colloquium

Speaker: Geoff Pleiss

Title: From n=1,000 to n=1,000,000: Scaling Up Gaussian Processes Inference with Matrix Multiplication and GPU Acceleration

Abstract:
Gaussian processes (GPs) are powerful machine learning models-offering well-calibrated uncertainty estimates, interpretable predictions, and the ability to encode prior knowledge. Despite these desirable properties, GPs are typically not applied to datasets with more than a few thousand data points-in part because of an inference procedure that requires matrix inverses, determinants, and other expensive operations. In this talk, I will discuss how my collaborators and I were able to scale GPs to datasets with over 1 million points, without making any simplifying assumptions. Taking inspiration from neural network libraries, we constrained ourselves to writing a GP inference algorithm that only used matrix multiplication and other linear operations-procedures that are extremely amenable to parallelization, GPU acceleration, and distributed computing. The resulting algorithm, Blackbox Matrix-Matrix Inference (BBMM), is up to 100x faster than existing inference procedures and scales to datasets that are 2 orders of magnitude larger than what has previously been reported.


Speaker: Vishal Shrivastav

Title: Building High-speed Datacenter Networks in the Post-Moore's Law Era

Abstract:
With the slowdown in Moore's law and the end of Dennard scaling, general-purpose CPUs and packet switches have hit a fundamental performance wall. However, the bandwidth demand within datacenters keeps growing exponentially: Applications keep getting more distributed and resources (e.g., storage) keep getting disaggregated demanding more bandwidth. In this talk, I will discuss how my research attempts to bridge this gap for next-generation datacenter networks. To overcome the limitations of general-purpose CPUs, my research argues for domain-specific architectures for network processing. I will start by briefly introducing several domain-specific hardware architectures that I have proposed over the course of my PhD, targeted at improving a wide-range of core network functions including packet scheduling, packet processing, congestion control, and network time synchronization.
To accommodate the exponential demand for network bandwidth, in the talk I will focus on how my research attempts to overcome the fundamental limitations of packet switching within datacenters. I will describe a new approach called Shoal that is the first end-to-end network design for a fast circuit-switched network. I will conclude my talk by discussing how Shoal takes us a giant step towards realizing the idealistic goal of a datacenter switching fabric that could support practically unlimited bandwidth at low power, low cost and high performance


Speaker: Laure Thompson

Title: Understanding and Directing What Models Learn

Abstract:
Machine learning and statistical methods, such as unsupervised semantic models, are popular and useful techniques for making massive digital collections more explorable and analyzable. But what underlying patterns do these models actually learn, and which patterns are they most likely to repeatedly learn? Moreover, how might we direct what these models learn so that they are useful to a wider range of scholarly inquiry? While it might be useful to organize texts by authors, learning this structure is seldom useful when already known and can be problematic if it is mischaracterized as a cross-cutting pattern. In this talk, I will focus on a specific problem: discuss my recent work on measuring and mitigating topic-metadata correlation in topic models.