This work addresses improvements in branch prediction mechanism to support high perfor- mance processors. The state of the art aims to balance the prediction latency and prediction accuracy using multi level correcting predictors . Prior published work focusses on scalar designs and prediction accuracy improvement for hard to predict branches employing tailor made, non generic […]
Autonomous Cyber Operations (ACO) agents promise effective security automation with minimal human intervention, yet their deployment raises three interconnected challenges: agents must be realistic (reproducing diverse attacker sophistication), secure (preventing autonomy from becoming an attack surface), and feasible (safely replicating human behavior at full autonomy). We argue that these three properties are requirements for ACO […]
Multimodal large language models (MLLMs) are rapidly becoming general-purpose AI systems, yet their capabilities are advancing faster than our ability to evaluate, improve, and validate their reliability in realistic use. Standard benchmarks mainly measure in-distribution final-answer accuracy, leaving critical gaps in safety, robustness, fine-grained reasoning evaluation, and reliability in real-world agentic settings. My research proposes […]
Presenter: Laura Waller, UC Berkeley Abstract: Computational imaging jointly designs hardware and algorithms to push beyond the classical limits of imaging, enabling measurement of new quantities (e.g. 3D, phase, and super-resolution) with simple, inexpensive hardware. These approaches have already transformed consumer photography; our goal is to achieve a similar transformation in scientific microscopy. In this […]
Modern system software increasingly needs runtime extensibility: userspace applications need safe ways to expose domain-specific extension points, GPU resource management needs workload-specific memory and scheduling policies, and kernel eBPF JIT compilers need different runtime optimizations as workloads and hardware vary. However, built-in policies are safe but difficult to specialize across rapidly changing workloads and hardware […]
Modern datacenter CPUs lose 25–66% of execution cycles to instruction-delivery stalls. This bottleneck persists, despite the recent trend towards accelerators and GPUs, as there is continuing demand by applications that only execute on CPUs. Two workload classes dominate today’s datacenter execution cycles: hyperscale server software (databases, build systems, and content stores), whose large instruction footprints […]
Autonomous systems, such as Autonomous Vehicles (AVs) and drones, are increasingly deployed across a wider array of contexts for both civilian and military use. As these systems become more common, they may be targeted by malicious actors seeking to exploit and abuse them, compromising safety-critical operations. Among the ways to protect these systems simulation based […]
A central goal on the path toward general AI is to build systems capable of deliberative reasoning before action. Such systems should inspect what they know, identify what they need, seek or construct useful information, and revise their reasoning through intermediate cognitive states. This dissertation studies this goal through the lens of Inner Monologue (IM), […]
This dissertation investigates how to make modern AI systems safer as they grow more capable. It addresses two central sources of risk: malicious misuse, in which adversarial users coerce models into harmful behavior, and internal misalignment, in which models themselves pursue goals that diverge from human intent through deception, sandbagging, or other covert behaviors. The […]
Recently, a class of non-invasive hardware side-channel attacks has been discovered in field-programmable gate arrays (FPGAs). These attacks extract remnants of prior users’ activity that persist as transistor defect states within reconfigurable routing resources. These remnants are known as FPGA Pentimenti. Resistive random-access memory (ReRAM) is a compelling candidate for pentimenti-like attacks beyond FPGAs. However, […]
Presenter: Sai Teja Peddinti, Google Abstract: As the digital landscape expands, traditional models of threat mitigation and user support are failing to keep pace with the unprecedented security, privacy, and safety challenges. Fortunately, the rise of large language models (LLMs) offers a powerful new paradigm for defense. This talk explores how LLMs are being leveraged […]
As quantum computing transitions from theory to practice, understanding which algorithms suit near-term devices becomes critical. Current quantum computers are severely constrained by limited qubit counts, short coherence times, and high error rates that quickly degrade computation into noise. This thesis addresses two interconnected questions: what non-trivial computational tasks can near-term devices execute and how […]
The growing complexity of high-performance computing (HPC) systems poses a fundamental challenge for domain scientists, whose primary objective is to obtain scientifically valid results rather than to optimize resource utilization. Modern leadership-class facilities combine heterogeneous CPUs, GPUs, and specialized accelerators across systems that simultaneously support traditional scientific simulations and AI-driven workloads. This creates a vast, […]
Modern high-performance GPU kernels increasingly rely on subgroup-level execution, including subgroup-level communication, subgroup operations, and matrix operations. These features are essential for workloads such as matrix multiplication and FlashAttention, but their language-level guarantees remain difficult to reason about. Existing programming models often leave unclear which threads participate in subgroup operations, when subgroup threads are required […]
Modern software increasingly relies on distributed systems to provide accessible, scalable, and reliable services. Choreographic programming brings a global perspective to distributed system development: programmers write a single program that describes the behavior of a whole system, and a compiler projects that global description into local programs run by each node. By making distributed control […]
Edge computing and distributed AI systems increasingly operate under heterogeneous resources, dynamic workloads, and frequent failures, requiring both adaptivity and fault tolerance for efficient execution. In heterogeneous edge clusters, nodes differ significantly in CPU throughput, memory capacity, and network bandwidth, while modern distributed GPU clusters supporting agentic LLM inference must recover large amounts of runtime […]
AI-assisted clinical decisions in medicine, and particularly in dermatology, demand fine-grained understanding across diverse skin tones, body sites, and disease types, yet expert-annotated datasets are scarce, demographically imbalanced, and almost devoid of rare presentations. This dissertation develops four deep learning systems for this low-label, low-coverage regime. We introduce HealNet, which learns wound healing stages from […]
Modern AI systems are rapidly moving beyond static text generation toward capable models and agents that reason, use tools, store memories, and update persistent state, yet safety methods still often assume a fixed model whose behavior can be controlled by output-level refusal. This leaves critical gaps in understanding why aligned models fail under adversarial pressure, […]
Due to their increasing complexity, autonomous decision-making agents rely on increasingly advanced algorithms, from classical control theory to reinforcement learning (RL) and, more recently, large vision-language models. While these algorithms help automate the decision-making in complex systems, they bring newer attack vulnerabilities that an adversary can exploit. In this dissertation, we study the security of […]
A community gathering of people interested in PyTorch and the projects that use it – not an official PyTorch organization. Sponsored by Red Hat and University of California Santa Cruz Location: Engineering 2, Room 180 Food, Socializing, and Excellent talks from the PyTorch Ecosystem 5:30 – 6:30 Food and Socializing 6:30 – 7:00 Talk 1 […]