Neuro-symbolic AI algorithms
ApplyProject Description
Symbolic, logic-based languages are inherently interpretable by humans. Symbols are entities standing for other entities and can be combined to form more complex expressions. Symbol systems are therefore well suited to explain and answer questions of “how” and “why” an intelligent agent (human or artificial) arrived at a decision. Knowledge-based systems based on logic have traditionally been used successfully in question answering (formulated as computing entailments, i.e., statements that must be true if all the axioms are assumed to be true) and can generate novel and “surprising” answers through deductive inference. However, they are not well suited to dealing with incomplete or noisy information or identifying patterns from unstructured data. Machine learning methods, in particular neural networks, can deal with noisy and
incomplete data substantially better than symbolic, logic-based methods. However, they operate mainly as black boxes which do not make the logic underlying a decision making process available. Neuro-symbolic methods in Artificial Intelligence aim to combine logic-based AI methods and neural methods to overcome the limitations of both.
The aim of the project is the identify, implement, evaluate, and improve neuro-symbolic methods. The baselines and experiments will focus on one of two possible areas of application: biomedical data where a large number of knowledge bases has been developed, or common sense knowledge.
Program -
Computer Science
Division -
Computer, Electrical and Mathematical Sciences and Engineering
Faculty Lab Link -
https://cemse.kaust.edu.sa/borg
Center Affiliation -
Computational Bioscience Research Center
Field of Study -
Artificial Intelligence
About the
Researcher
Robert Hoehndorf
Desired Project Deliverables
Month 1: identification of algorithm, technical presentation
Month 2: implementation, baseline experiments
Month 3: algorithm evaluation
Month 4: analysis, improvement and tuning
Month 5: experimental results, theoretical results
Month 6: write-up