A5034500430- Topic Modeling
- Natural Language Processing Techniques
- Privacy-Preserving Technologies in Data
- Speech and dialogue systems
- Machine Learning in Healthcare
- Artificial Intelligence in Healthcare and Education
- Recommender Systems and Techniques
- Multimodal Machine Learning Applications
- Brain Tumor Detection and Classification
- Stochastic Gradient Optimization Techniques
- Algorithms and Data Compression
- AI in cancer detection
- Network Packet Processing and Optimization
- Healthcare Systems and Public Health
- Human Mobility and Location-Based Analysis
- Electromagnetic Scattering and Analysis
- Semantic Web and Ontologies
- Magnetic Field Sensors Techniques
- Innovative Teaching and Learning Methods
- Online Learning and Analytics
- Engineering Applied Research
- Data Mining Algorithms and Applications
- Data Quality and Management
- Caching and Content Delivery
- Risk Perception and Management
Google (United States)
2021-2025
Stanford University
2007
Recent artificial intelligence (AI) systems have reached milestones in "grand challenges" ranging from Go to protein-folding. The capability retrieve medical knowledge, reason over it, and answer questions comparably physicians has long been viewed as one such grand challenge. Large language models (LLMs) catalyzed significant progress question answering; Med-PaLM was the first model exceed a "passing" score US Medical Licensing Examination (USMLE) style with of 67.2% on MedQA dataset....
BackgroundMedicine is inherently multimodal, requiring the simultaneous interpretation and integration of insights between many data modalities spanning text, imaging, genomics, more. Generalist biomedical artificial intelligence systems that flexibly encode, integrate, interpret these might better enable impactful applications ranging from scientific discovery to care delivery.MethodsTo catalyze development models, we curated MultiMedBench, a new multimodal benchmark. MultiMedBench...
Large language models (LLMs) have shown promise in medical question answering, with Med-PaLM being the first to exceed a 'passing' score United States Medical Licensing Examination style questions. However, challenges remain long-form answering and handling real-world workflows. Here, we present 2, which bridges these gaps combination of base LLM improvements, domain fine-tuning new strategies for improving reasoning grounding through ensemble refinement chain retrieval. 2 scores up 86.5% on...
An accurate differential diagnosis (DDx) is a cornerstone of medical care, often reached through an iterative process interpretation that combines clinical history, physical examination, investigations and procedures. Interactive interfaces powered by Large Language Models (LLMs) present new opportunities to both assist automate aspects this process. In study, we introduce LLM optimized for diagnostic reasoning, evaluate its ability generate DDx alone or as aid clinicians. 20 clinicians...
Personalization methods in federated learning aim to balance the benefits of and local training for data availability, communication cost, robustness client heterogeneity. Approaches that require clients communicate all model parameters can be undesirable due privacy constraints. Other approaches always-available or stateful clients, impractical large-scale cross-device settings. We introduce Federated Reconstruction, first model-agnostic framework partially suitable inference at scale....
Medicine is inherently multimodal, with rich data modalities spanning text, imaging, genomics, and more. Generalist biomedical artificial intelligence (AI) systems that flexibly encode, integrate, interpret this at scale can potentially enable impactful applications ranging from scientific discovery to care delivery. To the development of these models, we first curate MultiMedBench, a new multimodal benchmark. MultiMedBench encompasses 14 diverse tasks such as medical question answering,...
Large language models (LLMs) have revolutionized natural processing. However, effectively incorporating complex and potentially noisy user interaction data remains a challenge. To address this, we propose User-LLM, novel framework that leverages embeddings to contextualize LLMs. These embeddings, distilled from diverse interactions using self-supervised pretraining, capture latent preferences their evolution over time. We integrate these with LLMs through cross-attention soft-prompting,...
We study the relative effects of data augmentations, pretraining algorithms, and model architectures in Self-Supervised Learning (SSL). While recent literature this space leaves impression that algorithm is critical importance to performance, understanding its effect complicated by difficulty making objective direct comparisons between methods. propose a new framework which unifies many seemingly disparate SSL methods into single shared template. Using framework, we identify aspects differ...
Self-consistency with chain-of-thought prompting (CoT) has demonstrated remarkable performance gains on various challenging tasks, by utilizing multiple reasoning paths sampled from large language models (LLMs). However, self-consistency relies the answer extraction process to aggregate solutions, which is not applicable free-form answers. In this work, we propose Universal Self-Consistency (USC), leverages LLMs themselves select most consistent among candidates. We evaluate USC a variety of...
Large language models (LLMs) hold immense promise to serve complex health information needs but also have the potential introduce harm and exacerbate disparities. Reliably evaluating equity-related model failures is a critical step toward developing systems that promote equity. In this work, we present resources methodologies for surfacing biases with precipitate harms in long-form, LLM-generated answers medical questions then conduct an empirical case study Med-PaLM 2, resulting largest...
LLM-powered personalization agent systems employ Large Language Models (LLMs) to predict users' behavior from their past activities. However, effectiveness often hinges on the ability effectively leverage extensive, long user historical data due its inherent noise and length of such data. Existing pretrained LLMs may generate summaries that are concise but lack necessary context for downstream tasks, hindering utility in systems. To address these challenges, we introduce Reinforcement...
Answering complex natural language questions often necessitates multi-step reasoning and integrating external information. Several systems have combined knowledge retrieval with a large model (LLM) to answer such questions. These systems, however, suffer from various failure cases, we cannot directly train them end-to-end fix failures, as interaction is non-differentiable. To address these deficiencies, define ReAct-style LLM agent the ability reason act upon knowledge. We further refine...
Deep retrieval models are widely used for learning entity representations and recommendations. Federated provides a privacy-preserving way to train these without requiring centralization of user data. However, federated deep usually perform much worse than their centralized counterparts due non-IID (independent identically distributed) training data on clients, an intrinsic property that limits negatives available training. We demonstrate this issue is distinct from the commonly studied...
This work shows how to improve and interpret the commonly used dual encoder model for response suggestion in dialogue. We present an attentive that includes attention mechanism on top of extracted word-level features from two encoders, one context label respectively. To interpretability models, we design a novel regularization loss minimize mutual information between unimportant words desired labels, addition original method, so important are emphasized while de-emphasized. can help not only...
In human-human conversations, Context Tracking deals with identifying important entities and keeping track of their properties relationships. This is a challenging problem that encompasses several subtasks such as slot tagging, coreference resolution, resolving plural mentions entity linking. We approach this an end-to-end modeling task where the conversational context represented by repository containing references mentioned so far, relationships between them. The updated turn-by-turn, thus...