A Declarative Learning Based Programming Framework for Integration of Domain Knowledge and Statistical Learning
Abstract
We propose a framework and a novel underlying formalism for integrating domain knowledge and statistical learning. Our ambitious objective is to find a breakthrough and novel abstractions for designing AI complex systems based on data and knowledge representation rather than based on underlying computational units. While our approach is general, we envision its application in the context of automated decision support, where it is crucial to synthesize data and context in a meaningful way and to minimize retraining and reconfiguration efforts required by a rapidly evolving environment.The targeted abstraction layer will be high-level and intuitive for the domain experts to interact with, allowing them to design intelligent models rapidly without worrying about the underlying computations. The framework allows to seamlessly integrate knowledge about the world and goals expressed via declarative KR models and procedural knowledge about tasks and missions. Our proposed technique directly exploits domain knowledge to guide the training of the statistical models. Moreover, our models make consistent and informed predictions based on the domain knowledge at the decision-making phase.Our highly expressive formalism can handle uncertainty in various forms including soft constraints, preferences or use partial information while addressing scalability of computations. Knowledge is automatically used to identify data of interest, to generate learning examples and configurations,extract features, and to guide the training by enforcing the consistency of the outputs with the available knowledge or perform reasoning to generate the target outputs. Exploiting domain knowledge will help to learn novel concepts when there is little to no supervision. Our approach allows for working on heterogeneous data and we focus in particular on text and visual data, given its relevance to decision-support. Our novel architecture is built around adeclarative learning-based programming (DeLBP) backbone which organizes the integration of different machine learning techniques (e.g. deep learning), KR models, (e.g. ILP, constraint-based reasoning) and broad-coverage deep language understanding (TRIPS ontology and parsing model). While our model addresses the integration of domain knowledge in general, we focus on theencoding and the integration of TRIPS which is an extensive ontology and parsing knowledge into our framework. This will help in processing natural language data as well as in providing information about the world knowledge and relationships between various concepts when processing any arbitrary data.Our integrated approach will facilitate closing the loop from raw data to knowledge and from knowledge to training new concept learners in a lifelong learning process. Our goal is to pave the avenue for the next generation of languages and integrated computational models for building AISystems. To summarize, our key contributions are: 1) A unified representation model for various declarative and procedural uncertain knowledge that operates on learning building blocks, 2) Using various formalisms for inference based onknowledge representation and reasoning models, linear programming techniques, probabilistic inference, 3) Proposing a novel global heterogeneous optimization framework that enables scalable inference-based-training and knowledge-based learning based on declarative and procedural knowledge, 4) Integrating an existing ontology, TRIPS, along with a deep language semantic parser into our knowledge-based learning framework. This will address the lack of direct supervision. Moreover, the existing annotated data can be used to improve the parser. This will be a showcase for the loop of using knowledge to understand the data and using data to enrich the knowledge. The overall novelty of our proposal is providing a novel abstraction to help domain experts torapidly configure problem specifications and specify actionable knowledge to solve new proble
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 23, 2019
- Source ID
- N000141912308
Entities
People
- Parisa Kordjamshidi
Organizations
- Florida Institute for Human and Machine Cognition
- Office of Naval Research
- United States Navy