Problems I am thinking about these days:
Theory:
[1] Learning Accuracy Preserving Low Resource DNN Models
[2] Effects of LR Schedule on GD based Non-convex optimization methods
[3] Understanding the Topology and geometry of DNNs
[4] Manipulating the loss surface of DNNs
[5] Learning Disentangled Respresentations
NLP:
[1] Paraphrase Detection and Generation
[2] Text Classification
[3] Named Entity Recognition
[4] Learning universal Ontology
[5] char-word-sentence Embedding
Dialog Agents
[1] Dialog Simulation
[2] Unsupervised Dialog Evaluation
[3] User Behavior Model
[4] Automated Crowdsourcing to collect Dialog Data
[5] Annotator Model (Automated Annotation Model)
[6] User Agent Self Play
[7] Information theoretic Coherence