SSS Abstracts  Spring 2019

Presented by Ziqiang Feng, CSD

Deep learning has become the gold standard of computer vision. The generation of high-quality labeled training data typically becomes the bottleneck of deep learning in areas such as natural science, ecology, and medical research, where domain expertise is required to correctly identify targets and thus crowdsourcing becomes non-viable. Yet it is in those areas deep learning has huge potential value. In the worst case, a single domain expert needs to sift through a large volume of unlabeled data to discover only a few positive examples.

In this talk, I will describe our ongoing work on Eureka, a system intended to improve the human expert's productivity in building a labeled training set. Eureka views a human's attention and time as the most precious resource throughout the system and helps to optimize the utilization of this critical resource. Eureka combines three techniques to achieve its goal: early discard, iterative discovery workflow, and edge computing. Experiments show that Eureka can reduce the amount of labeling effort by two orders of magnitude relative a brute force approach.

Based on joint work with Shilpa George, Jan Harkes, Padmanabhan Pillai, Roberta Klatzky, and Mahadev Satyanarayanan.

In Partial Fulfillment of the CSD Speaking Skills Requirement

Multispectral Imaging for Fine-Grained Recognition of Powders on Complex Backgrounds

Friday, March 22^nd, 2019 from 12-1 pm in GHC 6501.

Presented by Tiancheng Zhi, CSD

Hundreds of materials, such as drugs, explosives, makeup, food additives, are in the form of powder. Recognizing such powders is important for security checks, criminal identification, drug control, and quality assessment. Powders are hard to distinguish: they are amorphous, appear matte, have little color or texture variation and blend with surfaces they are deposited on in complex ways. To address these challenges, we present the first comprehensive dataset and approach for powder recognition using multi-spectral imaging. To obtain more data, we propose a blending model to synthesize images of powders of various thickness deposited on a wide range of surfaces. We conduct fine-grained recognition of 100 powders on complex backgrounds, and achieve over 40% mean intersection-over-union (IoU) without known powder location.

Joint work with Bernardo Pires, Martial Hebert, and Srinivasa Narasimhan

In Partial Fulfillment of the Speaking Requirement

Motion Synthesis of Conversations for Background Characters

Friday, April 12^th, 2019 from 12-1 pm in GHC 6501.

Presented by Yanzhe Yang, CSD

Social scenes are common in many video games and movies. To realistically recreate such scenes in computer graphics, it is critical to animate the conversing human characters. A social scene typically contains foreground characters and background characters. While foreground characters are the focus of the scene and thus are carefully created by artists, the sole purpose of background characters is to render the atmosphere and to add realism to the environment. However, with traditional content creation tools, the artists often need to spend as much time animating the background characters as the foreground characters, despite that the exact behavior of the background characters is not critical to the experience -- they only need behave naturally so that they do not detract the audience. In this talk, I will introduce a system I have developed that helps artists to rapidly generate the animations for the talking characters in the background. The system automatically generates the body motions for two talking characters from an audio recording of a conversation. In order to produce natural looking animations, the system must ensure that the characters' body motions are smooth and are synchronized with the rhythm of the audio. For example, a speaker often uses hand gestures as they are stating an important point, and a listener will nod to acknowledge what the speaker is saying.

My talk will start with how we captured/recorded a database of real conversations and how we studied the statistics of the synchrony between body motion and audio signals from the database, then focus on the key algorithm of our system that generates novel motion sequences from an input audio based on the captured data. I will conclude my talk with results from a user study that demonstrates the effectiveness of our system.

Based on joint work with Jimei Yang and Jessica Hodgins.

In Partial Fulfillment of the Speaking Requirement.

HyperLoop: NIC-Offloading for Faster Transactions in Multi-tenant Storage Systems

Monday, April 15^th, 2019 from 12-1 pm in GHC 6501.

Presented by Daehyeok Kim, CSD

Storage systems in data centers are an important component of large-scale online services. They typically perform replicated transactional operations for high data availability and integrity. Today, however, such operations suffer from high tail latency even with recent kernel bypass and storage optimizations and thus affect the predictability of end-to-end performance of these services. We observe that the root cause of the problem is the involvement of the CPU, a precious commodity in multi-tenant settings, in the critical path of replicated transactions. In this talk, I present HyperLoop, a framework that removes CPU from the critical path of replicated transactions in storage systems by offloading them to commodity RDMA NICs, with non-volatile memory as the storage medium. To achieve this, we develop new and general NIC offloading primitives that can perform memory operations on all nodes in a replication group while guaranteeing ACID properties without CPU involvement. We demonstrate that popular storage applications can be easily optimized using HyperLoop. Our evaluation results with microbenchmarks and application benchmarks show that HyperLoop can reduce 99th percentile latency ~800x with close to 0% CPU consumption on replicas.

In Partial Fulfillment of the Speaking Requirement

Gradient-based Inference for Networks with Output Constraints

Monday, April 29^th, 2019 from 12-1 pm in GHC 6501.

Presented by Jay Yoon Lee, CSD

Practitioners apply neural networks to increasingly complex problems in natural language processing, such as syntactic parsing and semantic role labelling that have rich output structures. Many such structured-prediction problems require deterministic constraints on the output values; for example, in sequence-to-sequence syntactic parsing, we require that the sequential outputs encode valid trees. While hidden units might capture such properties, the network is not always able to learn such constraints from the training data alone, and practitioners must then resort to post-processing. In this paper, we present an inference method for neural networks that enforces deterministic constraints on outputs without performing rule-based post-processing or expensive discrete search. Instead, in the spirit of gradient-based training, we enforce constraints with gradient-based inference (GBI): for each input at test-time, we nudge continuous model weights until the networks unconstrained inference procedure generates an output that satisfies the constraints. We study the efficacy of GBI on three tasks with hard constraints: semantic role labelling, syntactic parsing, and sequence transduction. In each case, the algorithm not only satisfies constraints but improves accuracy, even when the underlying network is state-of-the-art.

Presented In Partial Fulfillment of the Speaking Skills Requirement.

Software-optimized Systems in the Era of Hardware Specialization

Friday, May 3^rd, 2019 from 12-1 pm in NSH 3305.

Presented by Anuj Kalia, CSD

In the post-Moore era, using specialized hardware tuned to specific applications is a promising direction to get higher performance. To create a cohesive specialization roadmap for the future, we must ask: how far can we go by optimizing software for existing hardware, and when must we turn to the more expensive option of deploying specialized hardware? I argue that for many important systems problems for which specialized hardware---intelligent NICs, FPGAs, programmable switches, and GPUs---has been proposed, software-optimized systems can provide competitive performance.

In this talk, I will focus on this optimization-specialization tradeoff in the context of distributed systems for modern datacenter networks. By using new techniques to better use existing CPUs and networks, my work invalidates the commonly-held belief that software-based networking cannot match datacenter network speeds, and allows building fast distributed systems that run entirely in software. I show that such designs have a fundamental latency advantage over distributed systems that use specialized hardware, whose limited flexibility often increases network round trips. I will describe in detail the design of eRPC, the first networking library to provide near-network performance in commodity datacenters. eRPC aligns with the end-to-end principle, and answers long-standing networking questions about reliability and congestion control.

In Partial Fulfillment of the Speaking Requirement

3D Machine Knitting

Monday, May 6^th, 2019 from 12-1 pm in GHC 6501.

Presented by Vidya Narayanan, CSD

Industrial knitting machines are commonly used to manufacture complicated shapes from yarns. However, designing patterns for these machines requires training and machine knitting expertise.

In this talk, I will introduce machine knitting, discuss machine knittability of arbitrary 3D shapes and introduce our new design system -- a general visual programming interface for creating 3D objects with complex surface finishes on industrial knitting machines. At the core of our interface is a new, augmented stitch mesh data structure to represent machine knittable patterns. The augmented stitch mesh stores low-level knitting operations per-face and encodes the dependencies between faces using directed edge labels. Our system can generate augmented stitch meshes from 3D models, allows users to edit these meshes in a way that preserves their knittability, and can generate machine instructions by scheduling the execution order and location of each face for fabrication. I will present knitted results fabricated on an industrial knitting machine using our system to demonstrate the flexibility of our pipeline.

Based on joint work with Kui Wu, Lea Albaugh, Jessica Hodgins, Stelian Coros, Cem Yuksel, and James McCann.

In Partial Fulfillment of the Speaking Requirement

Evolving Ext4 for Shingled Disks

Friday, May 10^th, 2019 from 12-1 pm in GHC 6501.

Presented by Abutalib Aghayev, CSD

Drive-Managed SMR (Shingled Magnetic Recording) disks offer a plug-compatible higher-capacity replacement for conventional disks. For non-sequential workloads, these disks show bimodal behavior: After a short period of high throughput they enter a continuous period of low throughput. We introduce ext4-lazy, a small change to the Linux ext4 file system that significantly improves the throughput in both modes. We present benchmarks on four different drive-managed SMR disks from two vendors, showing that ext4-lazy achieves 1.7-5.4x improvement over ext4 on a metadata-light file server benchmark. On metadata-heavy benchmarks it achieves 2-13x improvement over ext4 on drive-managed SMR disks as well as on conventional disks.

Joint work with Theodore Ts'o (Google), Garth Gibson (CMU), and Peter Desnoyers (Northeastern)

Presented in Partial Fulfillment of the CSD Speaking Skills Requirement

Web contact: sss+www@cs