Fellowships for PHD Students – AI2AI – JHU-Amazon Initiative for Interactive AI

Computer Vision:
Visual representations for a wide variety of video and image understanding tasks, such as activity/action recognition, scene understanding, body pose estimation, etc., especially those capable of supporting few-shot learning.
Speech and Language:
Paralinguistic information from speech; speech recognition in highly challenging environments; dialectal, code-mixed, and multilingual language understanding; robust language understanding for evolving vocabularies, entities, topics, and domains; knowledge extraction, representation and injection into speech recognition and language understanding systems; reasoning on knowledge graphs; machine translation and massively multilingual processing; etc.
Green AI:
Efficient learning via innovation in training and inference algorithms, in models (e.g., sparse models, retrieval augmented architectures), via knowledge distillation and hardware acceleration, and learning with limited resources (e.g. with less data or on edge devices).
Multimodal AI:
Interactive, multi-turn, multi-modal retrieval with text and image inputs, learning from cross-modal context in multimodal settings.
^NEW Security:
Automatically identify malicious vs. benign activity in Linux cloud-based environments (using IP/domain reputation, attack chain alert fusion, command line executions, obfuscation detection, etc.) Computer Vision and Video Indexing techniques for physical security and identification of malicious activity.
^NEW Responsible AI Approaches for Data:
Automatically assessing accuracy of human-labeled data; redacting sensitive, private or objectionable content from data; generating realistic synthetic data for model training; quantifying and mitigating bias in training data to attain equitable model performance; and estimating the complexity of a human-labeling task for allocating quality-control resources.