In-Depth Research Area Sessions
*Scroll to bottom for Speaker Details
|10:30am - 11:25am||
Program and Languages (CSE 1202)
Faculty: Nadia Polikarpova
|10:30am - 11:25am||Architecture (CSE 4140)
Faculty: Dean Tullsen
|10:30am - 11:25am||Artificial Intelligence (CSE 1242)
Faculty: Taylor Berg-kirkpatrick
*Scroll to bottom for Speaker Details
|11:30am - 12:30pm||Sysnet (CSE 1242)
Faculty: Aaron Schulman
|11:30am - 12:30pm||Embedded Systems (CSE 1202)
Faculty: Sicun Gao
*Scroll to bottom for Speaker Details
|1:30pm - 2:25pm||Robotics (CSE 1242)
Faculty: Henrik Christensen
|1:30pm - 2:25pm||Vision and Graphics (CSE 4140)
Faculty: Manmohan Chandraker
10:30am - 11:25am
Program and Languages (CSE 1202)
Speaker #1: Nico Lehmann
Title:: STORM: Refinement Types for Secure Web Applications
In this talk, I will present STORM, a web framework that allows developers to build MVC applications with compile-time enforcement of centrally specified data-dependent security policies. STORM ensures security using a Security Typed ORM that refines the (type) abstractions of each layer of the MVC API with logical assertions that describe the data produced and consumed by the underlying operation and the users allowed access to that data. I will also describe how STORM has been successfully used to implement end-to-end applications that show how it lets developers specify diverse policies while centralizing the trusted code to under 1% of the application, and statically enforces security with modest type annotation overhead, and no run-time cost.
Speaker #2: Shraddha Barke
Title:: Constraint-based learning of phonological processes
Phonological processes are context-dependent sound changes in natural languages. We present an unsupervised approach to learning human-readable descriptions of phonological processes from collections of related utterances. Our approach builds upon a technique from the programming languages community called *constraint-based program synthesis*. We contribute a novel encoding of the learning problem into Boolean Satisfiability constraints, which enables both data efficiency and fast inference. We evaluate our system on textbook phonology problems and datasets from the literature, and show that it achieves high accuracy at interactive speeds.
Speaker #3: John Renner
Title:: Scooter & Sidecar: A Domain-Specific Approach to Writing Secure Database Migrations
Web applications often handle large amounts of sensitive user data. Modern secure web frameworks protect this data by (1) using declarative languages to specify security policies alongside database schemas and (2) automatically enforcing these policies at runtime. Unfortunately, these frameworks do not handle the very common situation in which the schemas or the policies need to evolve over time—and updates to schemas and policies need to be performed in a carefully coordinated way. Mistakes during schema or policy migrations can unintentionally leak sensitive data or introduce privilege escalation bugs. In this work, we present a domain-specific language (Scooter) for expressing schema and policy migrations, and an associated SMT-based verifier (Sidecar) which ensures that migrations are secure as the application evolves. We describe the design of Scooter and Sidecar and show that our framework can be used to express realistic schemas, policies, and migrations, without giving up on runtime or verification performance.
Architecture (CSE 4140)
Speaker #1: Zixuan Wang
Title:: Enabling Efficient Large-Scale Deep Learning Training with Cache Coherent Disaggregated Memory Systems
Modern deep learning (DL) training is memory consuming, constrained by the memory capacity of each computation component and cross-device communication bandwidth. In response to such constraints, current approaches include increasing parallelism in distributed training and optimizing inter-device communication. However, model parameter communication is becoming a key performance bottleneck in distributed DL training. To improve parameter communication performance, we propose COARSE, a disaggregated memory extension for distributed DL training. COARSE is built on modern cache-coherent interconnect (CCI) protocols and MPI-like collective communication for synchronization, to allow low-latency and parallel access to training data and model parameters shared among worker GPUs. Our evaluation shows that COARSE achieves up to 48.3% faster DL training compared to the state-of-the-art MPI AllReduce communication.
Zixuan Wang ia a 4th year PhD candidate working with Prof. Jishen Zhao and Prof. Steven Swanson. Hia research is mainly on architecture and system support for memory systems.
Speaker #2: Jennifer Switzer
Title:: Junkyard datacenters: Carbon-efficient computing systems from old phones
1.5 billion smartphones are sold each year, and most are decommissioned less than two years later. The majority of these unwanted smartphones are neither discarded nor recycled, but languish in junk drawers and storage units. This computational stockpile represents a huge wasted potential: modern smartphones have an increasingly powerful and power-efficient processor, extensive networking capabilities, and a reliable built-in power supply.
This talk presents efforts to reuse smartphones as “junkyard computers”. Junkyard computers grow global compute capacity by extending device lifetimes, which saves on the carbon cost of manufacturing new devices. I show that the capabilities of even decades-old smartphones are within those demanded by modern cloud microservices, and discuss how to combine smartphones to perform increasingly complex tasks. I describe how current metrics are insufficient for measuring the carbon costs of compute, and propose new carbon-conscious efficiency metrics.
Jennifer is a 2nd year PhD student at UC San Diego. She received an MEng and BSE in Computer Science from MIT. Her research interests lie at the intersection of computer systems and sustainability. She is particularly interested in researching ways to extend the lifetimes of electronic devices.
Speaker #3: Zhiyuan Guo
Title:: Clio: A Hardware-Software Co-Designed Disaggregated Memory System
Memory disaggregation has attracted great attention recently because of its benefits in efficient memory utilization and ease of management. So far, memory disaggregation research has all taken one of two approaches: building/emulating memory nodes using regular servers or building them using raw memory devices with no processing power. The former incurs higher monetary cost and faces tail latency and scalability limitations, while the latter introduces performance, security, and management problems.
Server-based memory nodes and memory nodes with no processing power are two extreme approaches. We seek a sweet spot in the middle by proposing a hardware-based memory disaggregation solution that has the right amount of processing power at memory nodes. Furthermore, we take a clean-slate approach by starting from the requirements of memory disaggregation and designing a memory-disaggregation-native system. We built Clio, a disaggregated memory system that virtualizes, protects, and manages disaggregated memory at hardware-based memory nodes. The Clio hardware includes a new virtual memory system, a customized network system, and a framework for computation offloading. In building Clio, we not only co-design OS functionalities, hardware architecture, and the network system, but also co-design compute nodes and memory nodes. Our FPGA prototype of Clio demonstrates that each memory node can achieve 100 Gbps throughput and an end-to-end latency of 2.5 us at median and 3.2us at the 99th percentile. Clio also scales much better and has orders of magnitude lower tail latency than RDMA. It has 1.1x to 3.4x energy saving compared to CPU-based and SmartNIC-based disaggregated memory systems and is 2.7x faster than software-based SmartNIC solutions.
Zhiyuan Guo is a third year Ph.D. student in the Department of Computer Science and Engineering at University of California San Diego, where he works on enabling better resource utilization, flexibility and heterogeneity in datacenter. His research interests span operating systems, computer architecture and programming languages, with a focus on building efficient disaggregated and heterogeneous systems through co-designing system software and hardware. He received his BS in computer science from Beihang University in 2019.
Artificial Intelligence (CSE 1242)
Speaker #1: Jiarui Xu
Title:: "GroupViT: Semantic Segmentation Emerges from Text Supervision"
Grouping and recognition are important components of visual scene understanding, e.g., for object detection and semantic segmentation. With end-to-end deep learning systems, grouping of image regions usually happens implicitly via top-down supervision from pixel-level recognition labels. Instead, in this paper, we propose to bring back the grouping mechanism into deep networks, which allows semantic segments to emerge automatically with only text supervision. We propose a hierarchical Grouping Vision Transformer (GroupViT), which goes beyond the regular grid structure representation and learns to group image regions into progressively larger arbitrary-shaped segments. We train GroupViT jointly with a text encoder on a large-scale image-text dataset via contrastive losses. With only text supervision and without any pixel-level annotations, GroupViT learns to group together semantic regions and successfully transfers to the task of semantic segmentation in a zero-shot manner, i.e., without any further fine-tuning. It achieves a zero-shot accuracy of 51.2% mIoU on the PASCAL VOC 2012 and 22.3% mIoU on PASCAL Context datasets, and performs competitively to state-of-the-art transfer-learning methods requiring greater levels of supervision
Speaker #2: Mayank Sharan
Title:: Taming the Long Tail of Deep Probabilistic Forecasting
Deep probabilistic forecasting is gaining attention in numerous applications ranging from weather prognosis, through electricity consumption esti- mation, to autonomous vehicle trajectory predic- tion. However, existing approaches focus on im- provements on the most common scenarios with- out addressing the performance on rare and dif- ficult cases. In this work, we identify a long tail behavior in the performance of state-of-the-art deep learning methods on probabilistic forecasting. We present two moment-based tailedness measurement concepts to improve performance on the difficult tail examples: Pareto Loss and Kurtosis Loss. Kurtosis loss is a symmetric mea- surement as the fourth moment about the mean of the loss distribution. Pareto loss is asymmetric measuring right tailedness, modeling the loss using a generalized Pareto distribution (GPD). We demonstrate the performance of our approach on several real-world datasets including time series and spatiotemporal trajectories, achieving significant improvements on the tail examples.
Speaker #3: Sophia Sun
Title:: Probabilistic Symmetry for Improved Trajectory Forecasting
Trajectory prediction is a core AI problem with broad applications in robotics and autonomous driving. While most existing works focus on deterministic prediction, producing probabilistic forecasts to quantify prediction uncertainty is critical for downstream decision-making tasks such as risk assessment, motion planning, and safety guarantees. We introduce a new metric, mean regional score (MRS), to evaluate the quality of probabilis- tic trajectory forecasts. We propose a novel probabilistic trajectory prediction model, Probabilistic Equivariant Continuous COnvolution (PECCO) and show that leveraging symmetry, specifically rotation equivariance, can improve the predic- tions’ accuracy as well as coverage. On both vehicle and pedestrian datasets, PECCO shows state-of-the-art prediction performance and improved calibration compared to baselines.
Speaker #3: Yao-Yuan Yang
Title:: What You See is What You Get: Distributional Generalization for Algorithm Design in Deep Learning
We investigate and leverage a connection between Differential Privacy (DP) and the recently proposed notion of Distributional Generalization (DG). Applying this connection, we introduce new conceptual tools for designing deep-learning methods that bypass ``pathologies'' of standard stochastic gradient descent (SGD). First, we prove that differentially private methods satisfy a ``What You See Is What You Get (WYSIWYG)'' generalization guarantee: whatever a model does on its train data is almost exactly what it will do at test time. This guarantee is formally captured by distributional generalization. WYSIWYG enables principled algorithm design in deep learning by reducing generalization concerns to optimization ones: in order to mitigate unwanted behavior at test time, it is provably sufficient to mitigate this behavior on the train data. This is notably false for standard (non-DP) methods, hence this observation has applications even when privacy is not required. For example, importance sampling is known to fail for standard ERM, but we show that it has exactly the intended effect for DP-trained models. Thus, with DP-SGD, unlike with \erm, we can influence test-time behavior by making principled train-time interventions. We use these insights to construct simple algorithms which match or outperform SOTA in several distributional robustness applications, and to significantly improve the privacy vs. disparate impact tradeoff of DP-SGD. Finally, we also improve on known theoretical bounds relating differential privacy, stability, and distributional generalization.
11:30am - 12:30pm
Sysnet (CSE 1242)
Speaker #1: Lixiang Ao
Title:: FaaSnap: FaaS Made Fast Using Snapshot-based VMs
Speaker #2: Alisha Ukani
Title:: Locked-In during Lock-Down: Undergraduate Life on the Internet in a Pandemic
Speaker #3: Enze Alex Liu
Title:: Forward Pass: On the Security Implications of Email Forwarding Mechanism and Policy
Embedded Systems (CSE 1202)
Speaker #1: Xiyuan Zhang
Title:: ESC-GAN: Extending Spatial Coverage of Physical Sensors
Scientific discoveries and studies about our physical world have long benefited from large-scale sensing, from weather forecasting to wildfire monitoring. However, sensors in the real world are sparsely deployed, impeding timely and precise monitoring of the environment. Therefore, we seek to extend the spatial coverage of analysis based on existing sensory data, that is, to “generate” data for locations where no historical data exists. We observe that there are local patterns in nearby locations, as well as trends in a global manner (e.g., temperature drops as altitude increases regardless of the location). We propose ESC-GAN by modeling local, global, and multi-scale structure in an adversarial training framework, and achieve state-of-the-art performance for extending spatial coverage of sensory data.
Xiyuan Zhang is a second-year Ph.D. student at Computer Science and Engineering (CSE), UC San Diego, advised by Professor Rajesh Gupta and Professor Jingbo Shang. Her research interest focuses on developing sample-efficient algorithms in sequential data including data generation, augmentation, imputation, with past publications in data mining and AI venues (WSDM, AAAI).
Speaker #2: Gabriel Marcano
Title:: Soil Power? Can Microbial Fuel Cells Power Non-Trivial Sensors?
We explore the power delivery potential of soil-based microbial fuel cells. We build a prototype energy harvesting setup for a soil microbial fuel cell, measure the amount of power that we can harvest, and use that energy to drive an e-ink display as a representative example of a periodic energy-intensive load. We find that there is ample power over time to power our system several times a day. There remains, however, significant future work to make these systems reliable and maximally performant.
Gabriel Marcano is a second year PhD student in the Department of Computer Science and Engineering of UC San Diego. Before returning to academia, Gabriel worked for the MITRE Corporation for five years, primarily focusing on low power, embedded research and development. His current research continues this focus on low power embedded systems, now focusing on harvesting usable energy and powering embedded sensors using soil microbial fuel cells.
Speaker #3: Ranak Roy Chowdhury
Title:: Task-Aware Reconstruction for Time-Series Transformer
Time-series data contains temporal order information that can guide representation learning for predictive end tasks (e.g., classification and regression). Recently, there are some attempts to leverage such order information to first pre-train time-series models, followed by an end-task fine-tuning on the same dataset, demonstrating improved end-task performance. However, this learning paradigm decouples data reconstruction from the end task. Representations learned in this way are not informed by the end task and may therefore be sub-optimal for the end-task performance. In this work, we propose TARNet, a new model to learn task-aware data reconstruction that augments end-task performance. Specifically, we design a data-driven masking strategy that uses the self-attention score distribution to mask out data from important timestamps and reconstruct them, thereby making the reconstruction task-aware. This reconstruction task is trained alternately with the end task, sharing parameters in a single model, allowing the learned representation to improve end-task performance. Extensive experiments on tens of classification and regression datasets show that TARNet significantly outperforms state-of-the-art baseline models across all evaluation metrics.
Ranak is a third-year Ph.D. student working at the intersection of sensor fusion, IoT, and machine learning. His research is focused on enhancing the robustness of data mining methods to missing, irregular, asynchronous time-series data, and also developing self-supervised approaches for the same.
Speaker #3: Olivia Weng
Title:: ResNet Reshaper: Reshaping Residual Networks for Resource-Efficient Inference on FPGAs
Residual networks (ResNets) employ skip connections in their networks---reusing activations from previous layers---to improve training convergence, but these skip connections create challenges for hardware implementations of ResNets. In particular, they are problematic for inference accelerators on resource-limited platforms because they require extra buffers to store these activations, forcing higher utilization of on-chip/off-chip memories, requiring larger memory bandwidth and additional control logic. ResNet Reshaper is a fine-tuning process that modifies a ResNet's skip connections to make them more hardware-friendly with minimal to no loss in network accuracy.
Olivia Weng is a second-year PhD student working with Professor Ryan Kastner at UC San Diego. Her research focuses on applying hardware/software co-design to create efficient and fault tolerant architectures for deep learning.
1:30pm - 2:25pm
Robotics (CSE 1242)
Speaker #1: Sachiko Matsumoto
Speaker #2: Zhiao Huang
Title:: Differentiable physics and soft body manipulation
We introduce a new differentiable physics benchmark called PasticineLab, which includes various soft body manipulation tasks. We first compare reinforcement learning (RL) and gradient-based methods on this benchmark. Experimental results suggest that RL-based approaches struggle to solve most of the tasks efficiently. By optimizing open-loop control sequences with the built-in differentiable physics engine, gradient-based methods can rapidly find a solution within tens of iterations but still fall short on multi-stage tasks that require long-term planning. To enable long-horizon planning with differentiable physics, we develop a contact point discovery approach (CPDeform) that integrates optimal transport-based contact points discovery into the differentiable physics solver to overcome the local minima from initial contact points or contact switching.
Speaker #3: David Paz-Ruiz
Title:: TridentNet: A Conditional Generative Model for Dynamic Trajectory Generation
In recent years, various state of the art autonomous vehicle systems and architectures have been introduced. These methods include planners that depend on high-definition (HD) maps and models that learn an autonomous agent's controls in an end-to-end fashion. While end-to-end models are geared towards solving the scalability constraints from HD maps, they do not generalize for different vehicles and sensor configurations. To address these shortcomings, we introduce an approach that leverages lightweight map representations, explicitly enforcing geometric constraints, and learns feasible trajectories using a conditional generative model. Additional contributions include a new dataset that is used to verify our proposed models quantitatively. The results indicate low relative errors that can potentially translate to traversable trajectories. The dataset created as part of this work has been made available online.
Vision and Graphics (CSE 4140)
Speaker #1: Albert Chern
Title:: Geometry in Simulations
Computationally simulating physical phenomena is central to computer animations and computational sciences. However, discretizing the continuous space often results in equations with a large number of variables to solve. We approach the computational challenge by seeking better representations to these physical systems. Geometry, which focuses on the symmetries and invariants of the problems, provides a guidance for such changes of variables. In the talk, I will show two examples including synthesizing physics-governing surfaces and infinite domain problems.
Albert Chern is an Assistant Professor in Computer Science and Engineering at UC San Diego since 2020. He received his PhD in Applied and Computational Mathematics at Caltech in 2017, and worked as a postdoctoral researcher in mathematics at TU Berlin prior to UCSD. His research interests lie in the interplay among differential geometry, numerical partial differential equations, and their applications in geometry processing and physics simulations in computer graphics.
Speaker #2: Yu-Ying Yeh
Title:: Toward photorealistic 3D content creation for Augmented Reality
The development of augmented reality drives the need of photorealistic 3D content creation. We made efforts in this direction to create photorealistic 3D content for both single objects and indoor scenes. Our deep learning-based method for transparent shape reconstruction allows us to generate 3D shapes for transparent objects from a few mobile phone photos. We also collect a synthetic large-scale indoor scene dataset, OpenRooms, for learning inverse rendering and scene understanding tasks. Lastly, we build a framework, PhotoScene, to transfer material and lighting from photos to 3D scene geometries. This allows us to create digital twins from photos without laborious efforts from artists.
Yu-Ying Yeh is a PhD student working with Prof. Manmohan Chandraker. Her research interest mainly focuses on 3D content creation for photorealistic rendering. She is a recipient of the 2022 Google PhD Fellowship.
Speaker #3: Alexandr Kuznetsov
Title:: NeuMIP: Multi-Resolution Neural Materials
We propose NeuMIP, a neural method for representing a variety of materials at different scales. In this work, we generalize traditional mipmap pyramids to pyramids of neural textures, combined with a fully connected network. We also introduce neural offsets which enable rendering materials with parallax effects without any tessellation.
Alexandr Kuznetsov is a final year PhD student in Computer Graphics at University of California, San Diego, advised by Prof. Ravi Ramamoorthi. His research interests lie at the intersection of physically-based rendering and deep learning. He is working on developing novel neural representations for complex material appearances. Another direction is denoising for Monte-Carlo rendering.