UC Berkeley EECS Technical ReportsThe UC Berkeley EECS Technical Memorandum Series provides a dated archive of EECS research. It includes Ph.D. theses and master's reports as well as technical documents that complement traditional publication media such as journals. For example, technical reports may document work in progress, early versions of results that are eventually published in more traditional media, and supplemental information such as long proofs, software documentation, code listings, or elaborated examples.http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018-12-15T05:07:03Z2018-12-15T05:07:03ZenOptimization Everywhere: Convex, Combinatorial, and EconomicSam Wonghttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-185.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Optimization Everywhere: Convex, Combinatorial, and Economic</p>
<p>
Sam Wong</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-185<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-185.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-185.pdf</a></p>
<p>In this thesis we study fundamental problems that arise in optimization and its applications. We present provably efficient algorithms that achieve better running times or approximation guarantees than previously known. Our method draws on the toolkit from convex and combinatorial optimization as well as economics. By intertwining techniques from these disciplines, we are able to make progress on multiple old and new problems, some of which have stood open for many years. Main results of this thesis include the following:
<p>Convex Programming: We show how to solve convex programming with an expected O(n log(nR/\epsilon)) evaluations of the separation oracle and additional time O(n^3\log^{O(1)}(nR/\epsilon)). This matches the oracle complexity and improves upon the O(n^{\omega+1}\log(nR/\epsilon)) additional time of the previous fastest algorithm achieved over 25 years ago for the current value of the matrix multiplication constant when R/\epsilon=O(\poly(n)). </p>
<p>Submodular Function Minimization: We provide new weakly and strongly polynomial time algorithms with a running time of O(n^{2}\log nM\cdot\text{EO}+n^{3}\log^{O(1)}nM) and O(n^{3}\log^{2}n\cdot\text{EO}+n^{4}\log^{O(1)}n), improving upon the previous best of O((n^{4}\cdot\text{EO}+n^{5})\log M) and O(n^{5}\cdot\text{EO}+n^{6}) respectively. We also provide the first subquadratic time algorithm for computing an approximately optimal solution. </p>
<p>Matroid Intersection: We provide new algorithms with a running time of O(nr\mathcal{T_{\text{rank}}}\log n\log(nM)+n^{3}\log^{O(1)}nM) and O(n^{2}\mathcal{T_{\text{ind}}}\log(nM)+n^{3}\log^{O(1)}nM , achieving the first quadratic bound on the query complexity for the independence and rank oracles. In the unweighted case, this is the first improvement since 1986 for independence oracle. </p>
<p>Submodular Flow: We obtain a faster weakly polynomial running time of O(n^{2}\log(nCU)\cdot\EO+n^{3}\log^{O(1)}(nCU)), improving upon the previous best of O(mn^{5}\log(nU)\cdot\EO) and O\left(n^{4}h\min\left\{ \log C,\log U\right\} \right) from 15 years ago by a factor of \tilde{O}(n^{4}). </p>
<p>Semidefinite Programming: We obtain a running time of \tilde{O}(n(n^{2}+m^{\omega}+S)), improving upon the previous best of \tilde{O}(n(n^{\omega}+m^{\omega}+S)) for the regime S is small. </p>
<p>Market Equilibrium: We present the first polynomial time algorithm for computing market equilibrium in an economy with indivisible goods and general buyer valuations having only access to an aggregate demand oracle. </p>
<p>Vertex Cover with Hard Capacity: We give a f-approximation algorithm for the minimum unweighted Vertex Cover problem with Hard Capacity constraints on f-hypergraphs This improves over the previous 2f-approximation and is the best possible assuming the unique game conjecture. </p>
<p>Network Design for Effective Resistance: We initiate the study of network design for s-t effective resistance. Among other results we present a constant factor approximation by applying classic techniques to a convex quadratic programming relaxation.</p></p>
<p><strong>Advisor:</strong> Christos Papadimitriou</p>2018-12-14T08:00:00ZDomain-Specific Techniques for High-Performance Computational Image ReconstructionMichael Driscollhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-184.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Domain-Specific Techniques for High-Performance Computational Image Reconstruction</p>
<p>
Michael Driscoll</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-184<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-184.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-184.pdf</a></p>
<p>The widespread emergence of parallel computers in the last decade has created a substantial programming challenge for application developers who wish to attain peak performance for their applications. Parallel programming requires significant expertise, and programming tools---general-purpose languages, compilers, libraries, etc.---have had limited success in hiding the complexity of parallel architectures. Furthermore, the parallel programming burden is likely to increase as processor core counts grow and memory hierarchies become deeper and more complex.
<p>The challenge of delivering productive high-performance computing is especially relevant to computational imaging. One technique in particular, iterative image reconstruction, has emerged as a prominent technique in medical and scientific imaging because it offers enticing application benefits. However, it often demands high-performance implementations that can meet tight application deadlines, and the ongoing development of the iterative reconstruction techniques discourages ad-hoc performance optimization efforts. </p>
<p>This work explores productive techniques for implementing fast image reconstruction codes. We present a domain-specific programming language that is expressive enough to represent a variety of important reconstruction problems, but restrictive enough that its programs can be analyzed and transformed to attain good performance on modern multi-core, many-core and GPU platforms. We present case studies from magnetic resonance imaging (MRI), ptychography, magnetic particle imaging, and microscopy that achieve up to 90% of peak performance. We extend our work to the distributed-memory setting for an MRI reconstruction task. There, our approach gets perfect strong scaling for reasonable machine sizes, and sets the best-known reconstruction time for our particular reconstruction task. The results indicate that a domain-specific language can be successful in hiding much of the complexity of implementing fast reconstruction codes.</p></p>
<p><strong>Advisor:</strong> Katherine A. Yelick and Armando Fox</p>2018-12-14T08:00:00ZNon-Linear Stiffness Extraction & Modeling of Wineglass Disk ResonatorsAlain Antonhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-183.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Non-Linear Stiffness Extraction & Modeling of Wineglass Disk Resonators</p>
<p>
Alain Anton</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-183<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-183.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-183.pdf</a></p>
<p><strong>Advisor:</strong> Clark Nguyen</p>2018-12-14T08:00:00ZLow Dimensional Methods for High Dimensional Magnetic Resonance ImagingFrank Onghttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-181.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Low Dimensional Methods for High Dimensional Magnetic Resonance Imaging</p>
<p>
Frank Ong</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-181<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-181.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-181.pdf</a></p>
<p>Magnetic Resonance Imaging (MRI) is an amazing imaging modality in many aspects. It offers one of the best imaging contrast for visualizing soft issues. It has no ionizing radiation at all. Its flexibility has also enabled many applications, including assessing blood flow, imaging brain activity via oxygenation contrast, and measuring tissue stiffness. Since MRI was invented, this imaging technology has saved numerous lives, and has been the frontier of biomedical and engineering research.
<p>On the other hand, imaging speed remains a main limitation of MRI. Inherently, MRI takes time to collect measurements, and often requires minutes to complete a scan. In this regard, MRI is quite similar to early cameras: Subjects have to be motionless for minutes to obtain an image, which is uncomfortable to patients. This often leads to motion and motion artifacts. When severe motion artifacts occur, scans have to be repeated. </p>
<p>This dissertation aims to change that by developing techniques to reconstruct three-dimensional (3D) dynamic MRI from continuous acquisitions. An ideal 3D dynamic scan would be able to resolve all dynamics at a high spatiotemporal resolution. Subjects would not have to be motionless. The comprehensive information in the single scan would also greatly simplify clinical workflow. While this dissertation has not achieved this ideal scan yet, it proposes several innovations toward this goal. In particular, www.doi.org/10.6084/m9.figshare.7464485 shows a 3D rendering of a reconstruction result from this dissertation. Arbitrary slices at different orientation can be selected over time. Respiratory motion, contrast enhancements, and even slight bulk motion can be seen. </p>
<p>The main challenge in high resolution 3D dynamic MRI is that the reconstruction problem is inherently underdetermined and demanding of computation and memory. To overcome these challenges, this dissertation builds on top of many fundamental methods, including non-Cartesian imaging, parallel imaging and compressed sensing. In particular, this dissertation heavily relies on the compressed sensing framework, which has three components: 1) the image of interest has a compressed signal representation. 2) MRI can acquire (pseudo)-randomized samples in k-space, which provides incoherent encoding of the underlying image. 3) sparsity/compressibility can be efficiently enforced in reconstruction to recover the compressed representation from the undersampled measurements. </p>
<p>In this dissertation, I propose a multiscale low rank model that can compactly represent dynamic image sequences. The resulting representation can be applied beyond MRI, and is useful for other applications, such as motion separation in surveillance video. With the multiscale low rank representation, I propose a technique incorporating stochastic optimization to efficiently reconstruct 3D dynamic MRI. This makes it feasible to run such large-scale reconstructions on local workstations. To further speed up the reconstruction time, I propose accelerating the convergence of non-Cartesian reconstruction using a specially designed preconditioner. Finally, I leverage external undersampled datasets to further improve reconstruction quality using convolutional sparse coding.</p></p>
<p><strong>Advisor:</strong> Michael Lustig</p>2018-12-14T08:00:00ZParametric Oscillation with Wineglass Disk ResonatorsThanh-Phong Nguyenhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-179.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Parametric Oscillation with Wineglass Disk Resonators</p>
<p>
Thanh-Phong Nguyen</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-179<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-179.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-179.pdf</a></p>
<p><strong>Advisor:</strong> Clark Nguyen</p>2018-12-14T08:00:00ZDeep Networks for Equalization in CommunicationsLaura Brinkhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-177.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Deep Networks for Equalization in Communications</p>
<p>
Laura Brink</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-177<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-177.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-177.pdf</a></p>
<p>We apply the techniques from meta-learning and machine learning to the communications domain. Speciﬁcally, we explore how neural networks can learn to equalize new channel environments without training on them and how neural networks can learn to estimate and correct carrier frequency oﬀset for new rates of rotation without training on them. We show that deep neural networks can learn to learn to estimate channel taps for two tap channels. We also explore how deep recursive neural networks learn to learn to equalize for any given channel. We demonstrate that neural networks can learn to learn to estimate and correct carrier frequency oﬀset for new rates of rotation. Crucially, we do all of this without using backpropagation to re-train the networks for each new set of environmental conditions.</p>
<p><strong>Advisor:</strong> Anant Sahai</p>2018-12-14T08:00:00ZAcquiring Diverse Robot Skills via Maximum Entropy Deep Reinforcement LearningTuomas Haarnojahttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-176.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Acquiring Diverse Robot Skills via Maximum Entropy Deep Reinforcement Learning</p>
<p>
Tuomas Haarnoja</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-176<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-176.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-176.pdf</a></p>
<p>In this thesis, we study how maximum entropy framework can provide efficient deep reinforcement learning (deep RL) algorithms that solve tasks consistently and sample efficiently. This framework has several intriguing properties. First, the optimal policies are stochastic, improving exploration and preventing convergence to local optima, particularly when the objective is multimodal. Second, the entropy term provides regularization, resulting in more consistent and robust learning when compared to deterministic methods. Third, maximum entropy policies are composable, that is, two or more policies can be combined, and the resulting policy can be shown to be approximately optimal for the sum of the constituent task rewards. And fourth, the view of maximum entropy RL as probabilistic inference provides a foundation for building hierarchical policies that can solve complex and sparse reward tasks. In the first part, we will devise new algorithms based on this framework, starting from soft Q-learning that learns expressive energy-based policies, to soft actor-critic that provides simplicity and convenience of actor-critic methods, and ending with automatic temperature adjustment scheme that practically eliminates the need for hyperparameter tuning, which is a crucial feature for real-world applications where tuning of hyperparameters can be prohibitively expensive. In the second part, we will discuss extensions enabled by the inherent stochasticity of maximum entropy polices, including compositionality and hierarchical learning. We will demonstrate the effectiveness of the proposed algorithms on both simulated and real-world robotic manipulation and locomotion tasks.</p>
<p><strong>Advisor:</strong> Pieter Abbeel and Sergey Levine</p>2018-12-14T08:00:00ZAdaptive and Diverse Techniques for Generating Adversarial ExamplesWarren Hehttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-175.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Adaptive and Diverse Techniques for Generating Adversarial Examples</p>
<p>
Warren He</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-175<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-175.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-175.pdf</a></p>
<p>Deep neural networks (DNNs) have rapidly advanced the state of the art in many important, difficult problems. However, recent research has shown that they are vulnerable to adversarial examples. Small worst-case perturbations to a DNN model's input can cause it to be processed incorrectly. Subsequent work has proposed a variety of ways to defend DNN models from adversarial examples, but many defenses are not adequately evaluated on general adversaries.
<p>In this dissertation, we present techniques for generating adversarial examples in order to evaluate defenses under a threat model with an adaptive adversary, with a focus on the task of image classification. We demonstrate our techniques on four proposed defenses and identify new limitations in them. </p>
<p>Next, in order to assess the generality of a promising class of defenses based on adversarial training, we exercise defenses on a diverse set of points near benign examples, other than adversarial examples generated by well known attack methods. First, we analyze a neighborhood of examples in a large sample of directions. Second, we experiment with three new attack methods that differ from previous additive gradient based methods in important ways. We find that these defenses are less robust to these new attacks. </p>
<p>Overall, our results show that current defenses perform better on existing well known attacks, which suggests that we have yet to see a defense that can stand up to a general adversary. We hope that this work sheds light for future work on more general defenses.</p></p>
<p><strong>Advisor:</strong> Dawn Song</p>2018-12-14T08:00:00ZQuantifying the Development Value of Code ContributionsHezheng Yinhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-174.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>Quantifying the Development Value of Code Contributions</p>
<p>
Hezheng Yin</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-174<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-174.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-174.pdf</a></p>
<p>Counting the amount of source code that a developer contributes to a project does not reflect the value of the code contributions. Quantifying the value of code contributions, instead of only the amount, makes a useful tool for instructors grading students in massive online courses, managers reviewing employees' performance, developers collaborating in open source projects, and researchers measuring development activities. In this paper, we define the concept of development value and design a framework to quantify such value of code contributions. The framework consists of structural analysis and non-structural analysis. In structural analysis, we parse the code structure and construct a new PageRank-type algorithm; for non-structural analysis, we classify the impact of code changes, and take advantage of the natural-language artifacts in repositories to train machine learning models to automate the process. Our empirical study in a software engineering course with 10 group projects, a survey of 35 open source developers with 772 responses, and massive analysis of 250k commit messages demonstrate the effectiveness of our solution.</p>
<p><strong>Advisor:</strong> Armando Fox</p>2018-12-14T08:00:00ZA Platform-Based Approach to Verification and Synthesis of Linear Temporal Logic SpecificationsAntonio Iannopollohttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-173.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>A Platform-Based Approach to Verification and Synthesis of Linear Temporal Logic Specifications</p>
<p>
Antonio Iannopollo</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-173<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-173.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-173.pdf</a></p>
<p>The design of Cyber-Physical Systems (CPS) is challenging as it requires coordination across several domains (e.g., functional, temporal, mechanical). To cope with complexity, rarely a CPS is built from scratch. Instead, it is assembled by reusing available components and subsystems. If a component is not available, then it is made to order according to a specification which ensures its compatibility with the rest of the system.
<p>To achieve design goals faster while guaranteeing system safety, the correct instantiation of modules and subsystems is essential. Formal specifications, such as those expressed in Linear Temporal Logic (LTL), have the potential to drastically reduce design and implementation efforts by enabling rigorous requirement analysis and ensuring the correct composition of reusable designs. Composing formal specifications, however, is a tedious and error-prone activity, and the scalability of existing formal analysis techniques is still an issue. </p>
<p>In this dissertation, we present a set of techniques and algorithms that leverage compositional design principles to enable faster verification and correct-by-construction, platform-based synthesis of LTL specifications. In our framework, a design is a composition of several components (which could describe both hardware and software elements) represented through their specifications, expressed as LTL assume/guarantee interfaces, or contracts. The collection of all the available contracts, i.e., a library, describes the design platform. The contracts in the library are the building blocks of different possible designs, and they are simple enough that their correctness can be easily verified, yet complete enough to guarantee the correct and safe use of the components they represent. </p>
<p>Our contribution is two-fold. On the one hand, we address the verification task: given an existing composition of contracts from the library, we want to check whether it satisfies a set of desired requirements. We improve the scalability of existing verification techniques by leveraging pre-verified relations between contracts in the library. On the other hand, we enable specification synthesis: given a (possibly incomplete) set of desired system properties, we are able to automatically generate a composition of contracts, chosen from a library, that satisfies them. We do so by devising a set of algorithms based on formal inductive synthesis, where a candidate is either accepted as correct or is used to infer new constraints and guide the synthesis process towards a solution. Additionally, we show how to increase the scalability of our approach by leveraging principles from the contract framework to decompose a synthesis problem into several independent tasks, which are simpler than the original problem. We validate our work by applying it to several industrial-relevant case studies, including the problem of verification and synthesis of a controller for the electrical power system of an aircraft.</p></p>
<p><strong>Advisor:</strong> Alberto L. Sangiovanni-Vincentelli</p>2018-12-14T08:00:00Z3D Object Detection with Sparse Sampling Neural NetworksRyan GoyAvideh Zakhorhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-172.html2018-12-14T08:00:00Z2018-12-14T08:00:00Z<p>3D Object Detection with Sparse Sampling Neural Networks</p>
<p>
Ryan Goy and Avideh Zakhor</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-172<br>
December 14, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-172.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-172.pdf</a></p>
<p>The advent of inexpensive 3D sensors has resulted in an abundance of 3D point- clouds and datasets. For instance, RGB-D sensors such as Kinect can result in 3D point clouds by projecting 2D pixels into 3D world coordinate using depth and pose information. Recent advancements in deep learning techniques appear to result in promising solutions to 2D and 3D recognition problems including 3D object detection. Unlike 3D classification, 3D object detection has received less attention in the research community. In this thesis, we propose a novel approach to 3D object detection, the Sparse Sampling Neural Network (SSNN), which takes large, unordered point clouds as input. We overcome the challenges of processing three dimensional data by convolving a collection of ”probes” across a point cloud input which then feeds into a 3D convolutional neural network. This approach allows us to efficiently and ac- curately infer bounding boxes and their associated classes without discritizing the volumetric space into voxels. We demonstrate that our network performs well on indoor scenes, achiev- ing mean Average Precision (mAP) of 54.48% on the Matterport3D dataset, 62.93% on the Stanford Large-Scale 3D Indoor Spaces Dataset, and 48.4% on the SUN RGB-D dataset.</p>
<p><strong>Advisor:</strong> Avideh Zakhor</p>2018-12-14T08:00:00ZTowards Practical Privacy-Preserving Data AnalyticsNoah Johnsonhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-171.html2018-12-13T08:00:00Z2018-12-13T08:00:00Z<p>Towards Practical Privacy-Preserving Data Analytics</p>
<p>
Noah Johnson</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-171<br>
December 13, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-171.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-171.pdf</a></p>
<p>Organizations are increasingly collecting sensitive information about individuals. Extracting value from this data requires providing analysts with flexible access, typically in the form of databases that support SQL queries. Unfortunately, allowing access to data has been a major cause of privacy breaches.
<p>Traditional approaches for data security cannot protect privacy of individuals while providing flexible access for analytics. This presents a difficult trade-off. Overly restrictive policies result in underutilization and data siloing, while insufficient restrictions can lead to privacy breaches and data leaks. </p>
<p>Differential privacy is widely recognized by experts as the most rigorous theoretical solution to this problem. Differential privacy provides a formal guarantee of privacy for individuals while allowing general statistical analysis of the data. Despite extensive academic research, differential privacy has not been widely adopted in practice. Additional work is needed to address performance and usability issues that arise when applying differential privacy to real-world environments. </p>
<p>In this dissertation we develop empirical and theoretical advances towards practical differential privacy. We conduct a study using 8.1 million real-world queries to determine the requirements for practical differential privacy, and identify limitations of previous approaches in light of these requirements. We then propose a novel method for differential privacy that addresses key limitations of previous approaches. </p>
<p>We present Chorus, an open-source system that automatically enforces differential privacy for statistical SQL queries. Chorus is the first system for differential privacy that is compatible with real databases, supports queries expressed in standard SQL, and integrates easily into existing data environments. </p>
<p>Our evaluation demonstrates that Chorus supports 93.9% of real-world statistical queries, integrates with production databases without modifications to the database, and scales to hundreds of millions of records. Chorus is currently deployed at a large technology company for internal analytics and GDPR compliance. In this capacity, Chorus processes more than 10,000 queries per day.</p></p>
<p><strong>Advisor:</strong> Dawn Song</p>2018-12-13T08:00:00ZLearning to Navigate in Visual EnvironmentsPeter Jinhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-169.html2018-12-13T08:00:00Z2018-12-13T08:00:00Z<p>Learning to Navigate in Visual Environments</p>
<p>
Peter Jin</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-169<br>
December 13, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-169.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-169.pdf</a></p>
<p>Artificially intelligent agents with some degree of autonomy in the real world must learn to complete visual navigation tasks. In this dissertation, we consider the learning problem of visual navigation, as well as implementation issues facing agents that utilize learned visual perception systems. We begin by formulating visual navigation tasks in the setting of deep reinforcement learning under partial observation. Previous approaches to deep reinforcement learning do not adequately address partial observation while remaining sample-efficient. Our first contribution is a novel deep reinforcement learning algorithm, advantage-based regret minimization (ARM), which learns robust policies in visual navigation tasks in the presence of partial observability. Next, we are motivated by performance bottlenecks arising from large scale supervised learning for training visual perception systems. Previous distributed training approaches are affected by synchronization or communication bottlenecks which limit their scaling to multiple compute nodes. Our second contribution is a distributed training algorithm, gossiping SGD, which avoids both synchronization and centralized communication. Finally, we consider how to train deep convolutional neural networks when inputs and activation tensors have high spatial resolution and do not easily fit in GPU memory. Previous approaches to reducing memory usage of deep convnets involve trading off between computation and memory usage. Our third and final contribution is an implementation of spatially parallel convolutions, which partition activation tensors along the spatial axes between multiple GPUs, and achieve practically linear strong scaling.</p>
<p><strong>Advisor:</strong> Kurt Keutzer</p>2018-12-13T08:00:00ZZephyr: Simple, Ready-to-use Software-based Power Evaluation for Background Sensing Smartphone ApplicationsK. ShankariJonathan FürstYawen WangPhilippe BonnetDavid E. CullerRandy H. Katzhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-168.html2018-12-13T08:00:00Z2018-12-13T08:00:00Z<p>Zephyr: Simple, Ready-to-use Software-based Power Evaluation for Background Sensing Smartphone Applications</p>
<p>
K. Shankari, Jonathan Fürst, Yawen Wang, Philippe Bonnet, David E. Culler and Randy H. Katz</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-168<br>
December 13, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-168.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-168.pdf</a></p>
<p>Innovations in mobile hardware and software need corresponding advances in the accurate assessment of power consumption under realistic conditions. This is especially relevant for smartphone-based background sensing applications. Assessing the power consumption of such applications requires ease of use, deployment \emph{in situ} and well-understood error characteristics.
<p>Existing measurement methods, based on external power meters or power models, are increasingly unable to keep up with these requirements. External power meters require access to device batteries and do not capture context-sensitive power drain. Power models must be rebuilt for each specific device, adapted to each new OS version, and require administrator access to instrument fine-grained system-level APIs. These limitations impede the inclusion of accurate, universal evaluations in the research literature. </p>
<p>We propose a simple and portable alternative, Zephyr, which infers an application's power drain using the relative State of Charge change rate (SoCCR) via the phone's battery sensor. We validate our methodology through experiments that characterize SoCCR on Android and iOS devices and show that they are consistent with hardware readings, across identical phones, for the same phone over time and over both slowly and quickly varying workloads. </p>
<p>The Zephyr implementation is modular, open source, and available for Android and iOS today.</p></p>2018-12-13T08:00:00ZThe Sparse Manifold TransformYubei ChenDylan PaitonBruno Olshausenhttp://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-167.html2018-12-11T08:00:00Z2018-12-11T08:00:00Z<p>The Sparse Manifold Transform</p>
<p>
Yubei Chen, Dylan Paiton and Bruno Olshausen</p>
<p>
EECS Department<br>
University of California, Berkeley<br>
Technical Report No. UCB/EECS-2018-167<br>
December 11, 2018</p>
<p>
<a href="http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-167.pdf">http://www2.eecs.berkeley.edu/Pubs/TechRpts/2018/EECS-2018-167.pdf</a></p>
<p>We present a signal representation framework called the sparse manifold transform that combines key ideas from sparse coding, manifold learning, and slow feature analysis. It turns non-linear transformations in the primary sensory signal space into linear interpolations in a representational embedding space while maintaining approximate invertibility. The sparse manifold transform is an unsupervised and generative framework that explicitly and simultaneously models the sparse discreteness and low-dimensional manifold structure found in natural scenes. When stacked, it also models hierarchical composition. We provide a theoretical description of the transform and demonstrate properties of the learned representation on both synthetic data and natural videos.</p>
<p><strong>Advisor:</strong> Bruno Olshausen and Pieter Abbeel</p>2018-12-11T08:00:00Z