Fellowships for PHD Students

AWS Security:
Low-cost computer vision techniques for object interactions for physical security.
Near-real-time anomaly detection at scale (e.g., streaming petabytes per hour) to identify malicious activity in Linux cloud-based environments.
Running distributed big data, AI/ML, and/or streaming models with minimal latency and cost using host, network, and audit telemetry sources. Such data engineering can be used for selecting interesting subsets of data including complex, multi-event sequences, improving matching and micro-summarization efficiencies for continuous log processing, and improving event summarization processing efficiency while maintaining accuracy.
Use of AI/ML (e.g., GenAI) to automate and improve security response workflows (for example, from written reports or threat detections), specifically through incident summarization, response planning, and communication management.
Artificial General Intelligence & Search:
Large Language Models (LLMs):
Retrieval augmented generation (RAG), fine-tuning and alignment (SFT, RLHF), and efficient inference: ensuring accuracy and reducing hallucinations; maintaining privacy and trust; reasoning over long contexts
Long form context methods
Improving data efficiency; effectively distilling models for real-time inference, data quality checks
Multi-lingual LLMs and challenges for cross-language defects (e.g. cross-language hallucinations)
Synthetic data generation for LLM learning
Adapting LLMs for dynamic content (e.g., feeds, web content) in knowledge-augmented scenarios
Tool and Code Empowered LLM
External Knowledge and Domain Knowledge Enhanced LLM and Knowledge Updating
Vision-Language:
Multimodal learning and video understanding: retrieval with multimodal inputs (e.g., video, image, text, speech)
Adversarial ML with multimodal inputs
Comprehensive video understanding with diverse content (open-vocabulary)
Shared multimodal representation spaces, aligned codecs
LLM and VLM based Intelligent Agents
Search and Retrieval:
Personalization in Search, semantic retrieval, conversational search: understanding descriptive and natural language queries for product search; retrieving information using LLMs’ output
Search page optimization (ranking) using heterogeneous content such as related keywords, shoppable images, videos, and ads
Tool Learning for Proactive Information Seeking
Efficient Generative AI:
Novel model architectures for improved performance (accuracy & efficiency)
Training large neural network models with efficiency: High performance distributed training and inference algorithms for Generative AI systems, quality metrics and evaluations
Responsible Generative AI:
This may include, but is not limited to measurement and mitigation, guardrail models, privacy concerns, detecting and mitigating adversarial use cases, and machine unlearning and model disgorgement
Responsible AI for audio, image and video generation
Privacy preserving continual learning/self-learning
Fact Checking and Factual Error Correction for Truthful LLMs

Application Instructions