A5061781904- Topic Modeling
- VLSI and FPGA Design Techniques
- Natural Language Processing Techniques
- Parallel Computing and Optimization Techniques
- VLSI and Analog Circuit Testing
- Privacy-Preserving Technologies in Data
- Interconnection Networks and Systems
- Advancements in Photolithography Techniques
- Machine Learning and Data Classification
- Photonic and Optical Devices
- Advanced Memory and Neural Computing
- Adversarial Robustness in Machine Learning
- Low-power high-performance VLSI design
- Multimodal Machine Learning Applications
- Advancements in Semiconductor Devices and Circuit Design
- Explainable Artificial Intelligence (XAI)
- Imbalanced Data Classification Techniques
- Neural Networks and Reservoir Computing
- Semiconductor materials and devices
- Advanced Surface Polishing Techniques
- Sparse and Compressive Sensing Techniques
- 3D IC and TSV technologies
- Stochastic Gradient Optimization Techniques
- Speech and dialogue systems
- Advanced Data Storage Technologies
University of California, San Diego
2016-2025
University of California, Berkeley
1992-2024
Allen Institute
2023-2024
Umbo Computer Vision (United Kingdom)
2023
University of Southern California
2022-2023
Southern California University for Professional Studies
2022-2023
University of Hong Kong
2007-2023
Cornell University
2023
Google (United States)
2023
Stanford University
2023
Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these are as yet poorly characterized. In order to inform future research, prepare for disruptive model capabilities, ameliorate socially harmful effects, it is vital that we understand the present near-future limitations of language models. To address this challenge, introduce Beyond Imitation Game benchmark (BIG-bench). BIG-bench...
Bill Yuchen Lin, Chaoyang He, Zihang Ze, Hulin Wang, Yufen Hua, Christophe Dupuy, Rahul Gupta, Mahdi Soltanolkotabi, Xiang Ren, Salman Avestimehr. Findings of the Association for Computational Linguistics: NAACL 2022.
We present LLM-Blender, an ensembling framework designed to attain consistently superior performance by leveraging the diverse strengths of multiple open-source large language models (LLMs). Our consists two modules: PairRanker and GenFuser, addressing observation that optimal LLMs for different examples can significantly vary. employs a specialized pairwise comparison method distinguish subtle differences between candidate outputs. It jointly encodes input text pair candidates, using...
Conventional federated learning (FL) approaches are ineffective in scenarios where clients have significant differences the distributions of their local data. The Non-IID data distribution client causes a drift model updates from global optima, which significantly impacts performance trained models. In this paper, we present new algorithm called FLIS that aims to address problem by grouping into clusters jointly trainable distributions. This is achieved comparing <italic...
Clustered federated learning (FL) has been shown to produce promising results by grouping clients into clusters. This is especially effective in scenarios where separate groups of have significant differences the distributions their local data. Existing clustered FL algorithms are essentially trying group together with similar so that same cluster can leverage each other's data better perform learning. However, prior attempt learn these distribution similarities indirectly during training,...
The traditional approach in FL tries to learn a single global model collaboratively with the help of many clients under orchestration central server. However, learning might not work well for all participating data heterogeneity. Therefore, personalization becomes crucial handling challenges that arise statistical heterogeneity and non-IID distribution data. Unlike prior works, this we propose new obtaining personalized from client-level objective. This further motivates participate...
Though successful, federated learning (FL) presents new challenges for machine learning, especially when the issue of data heterogeneity, also known as Non-IID data, arises.To cope with statistical previous works incorporated a proximal term in local optimization or modified model aggregation scheme at server side advocated clustered approaches where central groups agent population into clusters jointly trainable distributions to take advantage certain level personalization.While effective,...
With the relentless scaling of technology nodes, design co-optimization (DTCO) for conventional (Conv.) cell structure is starting to reach its limitations due limited routing resources, lateral p-n separations, and performance requirements. As a result, system (STCO) has been proposed exploit benefits 3-D architectures. Complementary-FET (CFET) technology, which stacks p-FET on n-FET or vice versa, can release restriction separation reduce in-cell congestion by enabling direct connections....
This paper proposes a new paradigm for learning set of independent logical rules in disjunctive normal form as an interpretable model classification. We consider the problem decision rule training neural network specific, yet very simple two-layer architecture. Each neuron first layer directly maps to if-then after training, and output second disjunction set. Our representation neurons this enables us encode both positive negative association features rule. State-of-the-art net approaches...
Humans can perform unseen tasks by recalling relevant skills acquired previously and then generalizing them to the target tasks, even if there is no supervision at all. In this paper, we aim improve kind of cross-task generalization ability massive multi-task language models, such as T0 FLAN, in an unsupervised setting. We propose a retrieval-augmentation method named ReCross that takes few unlabelled examples queries retrieve small subset upstream data uses update model for better...
In this article, we propose an automated standard cell synthesis framework, SP&R, which simultaneously solves P&R without deploying any sequential/separate operations, by a novel dynamic pin allocation scheme. The proposed SP&R utilizes the multiobjective optimization feature of satisfiability modulo theories (SMT) to obtain optimal layouts. To achieve practical scalability develop various search-space reduction techniques, including breaking symmetry, conditional assignment/localization,...
Language models (LMs) have demonstrated their capability in possessing commonsense knowledge of the physical world, a crucial aspect performing tasks everyday life. However, it remains unclear whether they capacity to generate grounded, executable plans for embodied tasks. This is challenging task as LMs lack ability perceive environment through vision and feedback from environment. In this paper, we address important research question present first investigation into topic. Our novel...
Proprietary LMs such as GPT-4 are often employed to assess the quality of responses from various LMs. However, concerns including transparency, controllability, and affordability strongly motivate development open-source specialized in evaluations. On other hand, existing open evaluator exhibit critical shortcomings: 1) they issue scores that significantly diverge those assigned by humans, 2) lack flexibility perform both direct assessment pairwise ranking, two most prevalent forms...
Current asynchronous tools are focussed mainly on the design of a single interface module. In many applications, one must interacting modules that potentially communicate in complex and intricate ways. When designing communicating modules, several difficult problems arise. First, even if each individual module can be synthesized correctly, according to environmental assumptions specified for module, composition may not work properly. Thus, needs have way model how interact with other, verify...
Crossbar architectures are widely used to implement high-performance network switches and routers. A crossbar switch can transfer cells between multiple ports simultaneously by closing cross points. This configuration must be determined an intelligent centralized scheduler that ensure fairness high utilization. In this paper, we describe the design implementation of two symmetric scheduling algorithms for configuring crossbars in input-queued support virtual output queueing. Our target is a...
Syntax directed translation based compilation from high-level concurrent programs has matured significantly over the past few years. They have been applied to significant designs in domains of digital signal processing and microprocessor designs. For data-path dominated designs, like those found applications, syntax approaches shown generate efficient asynchronous implementations. However for control-dominated where data parts play a relatively minor role, we believe solutions produced by...
Pin accessibility encounters nontrivial challenges due to the smaller number of routing tracks, higher pin density, and more complex design rules. Consequently, securing rule-correct routability has become a critical bottleneck for sub-10-nm IC designs (particularly in detailed stage) costing days runtime. To reduce turnaround time, designers demand new methodologies analyze feasibility given layout architecture (e.g., conditional rules, assignment patterns, etc). There are several...