April 5th

10:00 – 10:20am Welcome (Breakfast/Coffee)

10:20 – 11:40am EE-related research talk (Packard 101)

(Talk abstracts can be found at the end of the page)

Improving Resource Efficiency in Cloud Computing
Christina Delimitrou (Stanford – Computer Architecture)

Adaptive Antennas in a Dynamic GPS Environment
Emily McMilin (Stanford – GPS / Electromagnetism)

Information flows in control systems
Gireeja Ranade (Berkeley – Theory)

Characterizing Parallel Graph Analysis Algorithms
Nicole Rodia (Stanford – Computer Architecture)

11:50 – 1:50pm Lunch/Panel with Professors (Packard Atrium)
Prof. Ada Poon (Stanford EE), Prof. Debbie G. Senesky (Stanford EE), Prof. Fei-Fei Li (Stanford CS) Prof. Virginia Willams (Stanford CS), Prof. Sheryl Root (CMU)

2:00 – 3:40pm CS-related research talk  (Packard 101)

(Talk abstracts can be found at the end of the page)

Regulating Vulns: Options for Controlling the Gray Market in Zero-Day Vulnerabilities
Mailyn Fidler (Stanford – Science, Technology and Society)

Detecting stealthy, distributed SSH brute-forcing
Mobin Javed (Berkeley – Network Security)

Understanding Regulation of Gene Translation using Probabilistic Modeling
Cristina Pop (Stanford – Computational Biology)

A case for nonconvex optimization
Po-Ling Loh (Berkeley – Machine Learning)

Induced Lexico-Syntactic Patterns Improve Information Extraction from Online Medical Forums
Sonal Gupta (Stanford – Natural Language Processing)

3:40 – 4:30pm Poster Session (Packard Atrium)


Text to 3D Scene Generation
Angel X. Chang (Stanford – NLP)

Downton Abbey without Hiccups: Tales from Netflix
Te-Yuan Huang
(Stanford – Computer Networking)

Mapping Match+Action Tables to Switches
Lavanya Jose / Lisa Yan (Stanford – Computer Networking)

Machine Learning to identify transmitters in real-world spectrum measurements
Aakanksha Chowdhery (Stanford Alumni/Microsoft – Communication)

Understanding Regulation of Gene Translation using Probabilistic Modeling
Cristina Pop (Stanford – Computational Biology)

Discovering Low Work Function Materials for Thermionic Energy Conversion
Sharon Chou (Stanford – Surface Engineering)

4:30 – 5:00pm Walk to Stanford Golf Course

5:00 – 7:00pm Dinner with Industrial professionals (at Stanford Golf Course)

Industrial Professionals:
Aditi Muralidharan (Google, Berkeley PhD Alumnus)
Pi-Chuan Chang (AltSchool, Stanford PhD Alumnus)
Floraine Berthouzoz (Adobe, Berkeley PhD Alumnus)
Maria Kazandjieva (Netflix, Stanford PhD Alumnus)
Kshipra Bhawalkar (Google, Stanford PhD Alumnus)
Aakanksha Chowdhery (Microsoft Research, Stanford PhD Alumnus)
Ya-Lin Huang (Yelp, Georgia Tech PhD Alumnus)

Talk Abstract:

EE talks:

Improving Resource Efficiency in Cloud Computing
Christina Delimitrou (Stanford – Computer Architecture)

Cloud computing promises flexibility and high performance for users and lower costs for operators. However, datacenters today operate at very low utilization, wasting power and posing scalability limitations. There are several reasons behind datacenter underutilization, including exaggerated resource reservations, workload interference and platform heterogeneity.
In this work we present Quasar a cluster management system that is both resource-efficient and QoS-aware. Quasar moves away from the traditional reservation-based approach in cluster management to a more performance-centric approach. It also leverages efficient data mining techniques to extract application preferences in terms of resources without resorting to exhaustive workload profiling. We have evaluated Quasar on a large EC2 cluster under various workload scenarios and have shown that it improves utilization by 47% without degrading per-application performance.

Adaptive Antennas in a Dynamic GPS Environment
Emily McMilin (Stanford – GPS / Electromagnetism)

As a GPS signal travels 20,000km+ from a satellite to us, the signal is spread thin, in order to blanket almost an entire hemisphere of earth with its faint glow. In fact this signal is, at best, about 20dB below the thermal noise floor, or in other words, its power is about 100 times weaker than the noise generated by electrons bouncing around inside of our radio receivers. Nonetheless, not only must the GPS signal fight to be heard over these noisy electrons, it must also be prepared to face accidental interference and intentional jamming and spoofing.

Despite the antenna’s privileged position as the first line of defense against interferers, jammers and spoofers, most adaptive techniques such as detection and mitigation, are realized in the radio receiver backend, in the digital domain. Processing the signal purely in the digital domain can be computationally costly and unjustifiably complex as compared to processing in the analog domain. Furthermore, these techniques tend to rely on antenna arrays spanning multiple wavelengths in dimension, or a single antenna following some trajectory through space.

In contrast, we seek to achieve adaptability by manipulating the current distribution on a single, motion-less, sub-wavelength antenna. Specifically we will exploit analog signal processing at the antenna to achieve a low-cost solution, and use a single antenna to achieve a small form-factor solution. A primary technique for achieving adaptability in single antenna systems is the use of discrete circuit components on or near the antenna. Circuit components manipulating currents in the RF analog domain can achieve “signal processing” directly on the antenna. Thus these antennas would harden the receiver by assisting other elements in the radio chain. Alternatively, they could be used alone to reduce vulnerability in otherwise unprotected consumer-friendly products.

Information flows in control systems
Gireeja Ranade (Berkeley – Theory)
As the “Internet of things” grows we have more devices embedded with sensors that have the ability to communicate and interact seamlessly with each other and their surroundings (e.g. self-driving cars or smart home devices such as thermostats or vacuum cleaners).
To design control for systems that adapt in real-time or change at high-speeds, we need to understand situations where the system model itself is changing, i.e. systems with parameter uncertainties.

Parameter uncertainty manifests as a multiplicative noise in system models and cannot be understood using traditional additive noise approaches. In contrast to previous results, we find that multiplicative noise can serve as a ‘scrambling’ bottleneck for information flow in systems and obscures our ability to observe or control the system. Surprisingly, actively controlling the system allows us to simultaneously learn the system state. Parameter uncertainty may also be mitigated with knowledge about the system parameters, and we will explore the notion of the “value of information” in control systems. These principles can help dictate design decisions, like whether a laser or a radar sensor would be more valuable for a self-driving car?

Characterizing Parallel Graph Analysis Algorithms
Nicole Rodia (Stanford – Computer Architecture)

The growing importance of large-scale graph-based applications, which encompasses areas such as social networks, bioinformatics, and web search, necessitates efficient and high-performance parallel computing for graph analysis. I use execution-driven simulation and performance counters on an x86 multicore processor to characterize social and information network analysis algorithms. This informs how to improve graph analysis performance and efficiency using micro-architectural and algorithmic techniques.

CS Talks:

Regulating Vulns: Options for Controlling the Gray Market in Zero-Day Vulnerabilities
Mailyn Fidler (Stanford – Science, Technology and Society)
The emergent “gray-market” global trade in zero-day vulnerabilities involves governments purchasing vulnerabilities for cyber attack and espionage purposes. This market poses negative security consequences and raises questions about what regulatory mechanisms might best address these consequences.

Zero-day vulnerabilities are flaws discovered in existing software programs about which neither the responsible company nor the public knows. These vulnerabilities are exploitable on the “zero-th” day of their existence, and their secrecy and immediate exploitability make them valuable tools. Zero-days were previously the domains of software security researchers, who reported zero-days to responsible companies for free or a small reward, or of hackers, who profited by selling them on the black market for largely criminal purposes.

Governments, particularly the U.S. government, have started purchasing zero-day vulnerabilities for use in cyber attack and cyber espionage, paying high prices, building a stockpile, and feeding a thriving gray market. Prices on the gray market range from about $16,000 to $250,000 per vulnerability, usually much higher than prices on the black and white markets. The NSA employs in-house researchers to find zero-days, but the U.S. government also allocated $25 million for purchase of zero-day vulnerabilities in fiscal year 2013. The U.S. government used zero-day vulnerabilities in Stuxnet, the cyber attack against Iranian centrifuges, and in programs such as NSA’s FoxAcid, which compromises targeted computers.

Although some contend the U.S. government has compelling national security reasons to participate in the zero-day market, others criticize the practice for its broader cybersecurity consequences. The current public understanding of U.S. government policy is that the government does not notify affected companies about vulnerabilities it identifies or purchases. This practice leaves companies and citizens vulnerable to exploitation if other parties discover the flaw, which undermines citizen cybersecurity in pursuit of other national security objectives. The success of government identification, purchase, and deployment of zero-day vulnerabilities depends on the continued vulnerability of everyone else. Similarly, high gray market prices divert trade from the white market, making the white market less lucrative than when it only competed with the black market. On an international level, the burgeoning gray market means U.S. adversaries with low cyber capacities can access “ready made” cyber attack tools, potentially more rapidly achieving the capability to threaten U.S. interests in cyberspace.

Given these negative consequences, my research investigates options for regulating the zero-day gray market. Examining both domestic and international approaches, I analyze a suite of tools ranging from “soft” to “hard” law. On the domestic side, I examine criminalization, U.S.-based export controls, and inter-agency transparency-building initiatives. On the international side, I analyze potential initiatives within existing international organizations, non-binding but normative restrictions on exports through the Wassenaar Agreement, and the possibility of a binding treaty. Currently, my research demonstrates that each option has significant drawbacks, but these options are part of ongoing policy discussions. Analyzing the potential and downsides of each option is intended to serve as a useful resource for policymakers.

Detecting stealthy, distributed SSH brute-forcing
Mobin Javed (Berkeley – Network Security)

A longstanding challenge for detecting malicious activity has been the problem of how to identify attacks spread across numerous sources, such that the individual activity of any given source remains modest, and thus potentially not particularly out of the ordinary. These scenarios can arise whenever a detector employs a threshold used to flag that a given candidate attack source has exhibited a suspiciously high level of activity (e.g., when conducting scanning or DoS flooding). Attackers can respond to such detection procedures by employing multiple sources in order to thin out their activity to prevent any single source from exceeding the threshold; their attack becomes distributed and therefore potentially stealthy, i.e., hard to detect based on any individualized analysis.

In this work we present a general strategy for potentially detecting such stealthy, distributed activity.  We apply our approach to the problem of detecting distributed SSH brute-forcing: attackers employing a number of systems that each try different username/password combinations against a site’s SSH login servers, hoping that one of them will stumble across a working combination made possible by a careless user. Using the detector, we find dozens of instances of such attacks in an extensive 8-year dataset collected from a site with several thousand SSH users. Many of the attacks—some of which last months—would be quite difficult to detect individually. While a number of the attacks reflect indiscriminant global probing, we also find attacks that targeted only the local site, as well as occasional attacks that succeeded.


Understanding Regulation of Gene Translation using Probabilistic Modeling
Cristina Pop (Stanford – Computational Biology)

The translation of RNA into protein is an important step in gene expression. During translation, a molecule called the ribosome scans the RNA sequence, pausing with varying rates at each set of three bases (a codon) to translate it into an amino acid that joins the protein chain being synthesized. An emerging technique called ribosomal profiling gives a noisy count of how many ribosomes are observed at each codon on each gene at a snapshot in time. We present a rigorous machine learning model, motivated by queuing theory, that can extract from this data the rates at which the ribosome pauses and the protein synthesis rate per gene. We use these quantities to examine which codons are slowly translated and which biological features imply good efficiency.

Because several codons can translate to the same amino acid, codon usage is biased, especially in highly-expressed genes, suggesting either that efficient translation imposed evolutionary pressure that formed codon usage bias, or that we can use preferred codons to enhance efficiency. Using our model and experimental validation on mutants, we suggest that a faster codon does not necessarily lead to a more efficiently-translated gene in physiological conditions, but rather that the causality is reversed. We propose that signals such as the structure of the RNA and motifs at the start of the gene influence efficiency.

A case for nonconvex optimization
Po-Ling Loh (Berkeley – Machine Learning)
Recent years have brought about a flurry of work on convex relaxations of nonconvex problems. For instance, the convex l_1 norm is used as a convex surrogate for the nonconvex l_0 penalty, which counts the number of nonzeros in a vector. Convex objectives have the attractive property that local optima are also global optima, and these optima may be found efficiently.

We present results following a line of current work that advocates the use of nonconvex regularizers. Although the resulting objective functions possess multiple local and global optima, we show that for interesting classes of functions arising from statistical estimation problems, both local and global optima are statistically consistent. Our work is the first of its kind to provide sufficient conditions under which local and global optima are clustered together, and presents a favorable case in the realm of nonconvex methods for statistical estimation.

This is joint work with Martin Wainwright.

Induced Lexico-Syntactic Patterns Improve Information Extraction from Online Medical Forums
Sonal Gupta (Stanford – Natural Language Processing)

Despite the increasing quantity of Patient Authored Text (PAT), such as online discussion threads, tools for identifying medical entities in PAT are limited. When applied to PAT, existing tools either fail to identify specific entity types or perform poorly. Identification of SC and DT terms in PAT would enable exploration of efficacy and side effects for not only pharmaceutical drugs, but also for home remedies and components of daily care. In our work, we reliably extract two entity types, symptoms & conditions (SCs) and drugs & treatments (DTs), from PAT by learning lexico-syntactic patterns from data annotated with seed dictionaries. We use SC and DT term dictionaries compiled from online sources to label several discussion forums from MedHelp ( We then iteratively induce lexico-syntactic patterns corresponding strongly with each entity type to extract new SC and DT terms. Our system is able to extract symptom descriptions and treatments absent from our original dictionaries, such as ‘LADA’, ‘stabbing pain’, and ‘Cinnamon pills’. It outperforms the dictionaries by 4-7% for identifying DT and 2-3% for identifying SC on two forums from MedHelp. We show improvements over MetaMap, OBA, a conditional random field based classifier, and a previous pattern learning approach. To the best of our knowledge, this is the first paper to extract SC and DT entities from PAT. We exhibit learning of informal terms often used in PAT but missing from typical dictionaries.

Joint work with Diana MacLean, Jeffrey Heer, and Christopher Manning.