Visions of AI Lighting talk

Computer vision for intelligent systems by Gim Hee Lee, NUS

Formal methods in AI by Kuldeep Meel, NUS

• provably correct systems
  • provably correct probabilistic reasoning
• explainability
  • interpretable learning
• verification
  • verfication of probabilistic AI systems

Machine reasoning and deep spiking networks by Shaowei Lin, SUTD

• have neural and symbolic modules the work together
• homotopy type theory as the logical language
• robust learning of sequences (paths)

Transfer learning by Sinno Pan, NTU

• No knowledge accumulation, need to extract the knowledge learned
• vector of weights is too abstract to represent knowledge
• graphical models as the bridge between NN weights and first order logic

When AI meets Game Theory by Bo An, NTU

• GT for security, security resource scheduling
• GT for urban intelligence, optimal policy making, dynamic ERP
• adversarial machine learning
  • combating fraud on Alibaba using deep RL

Preferences and recommendations from Data and AI by Hady Lauw, SMU

• multi-modal preference signals
  • reviews
  • photos
  • rating
• visual sentiment analysis
• comparison of products in review/tweets
• preference from social links

Towards collaborating human-machine intelligent systems by Harold Soh, NUS

• machines should account for human psychology
• joint decision-making
• rich way interactions, then learn from it

State of the art in Language Error Correction by Ng Hwee Tou, NUS

• errors in writing, grammatical, word choice
• modeled as a machine translation from “bad English” to “good English”

Human centric AI by Jiewen Wu, AStar

multi agent planning in urban environment by Akshat Kumar, SMU SIS

• reason with aggregate data using collective graphical models

Clinical data analysis, Rajan, NUS IS

• multiview
• clustering

Data privacy for machine learning, NUS

• 33 bits of data to identify an individual
• black box access to machine learning models can leak data
  • reconstruction attack, tracing attack
• model that generalize should also be privacy preserving
  • does not depend on individual training records
• privacy needs to done during training