KB4DL: Building a Knowledge Base for Deep Learning

Deep Learning (DL) has received considerable attention from the AI community. However, we suffer from the lack of ability in interpretation and annotation of the outcomes from extensive and exhausting learning efforts. In this thesis, we propose an automatic metadata extraction both from DL models as well as their publications and populate ontologies in Deep Learning. The automatically populated ontology represents the state-of-the-art (SOTA) results in Deep Learning based on the metadata extracted from models (model name, accuracy, network models, output, parameters) as well as ones from publications (paper title, author, topics, images, tables, data source). The evaluation of the ontology has been conducted with the models and publications from the SOTA research publications. Studying and understanding deep learning models is the base step for any deep learning engineer and availability of many deep learning frameworks makes it difficult for a developer to understand the model implementation and to study it deeply. In addition, users have to deal with huge computational requirements of the Machine Learning world and have to invest a lot of time in analyzing the model based on every parameter. It is even more tough to find the right data to replicate or evaluate the model, as data pre-processing is a crucial step in getting the right model. To handle these situations, in this thesis, we have developed an approach to extract possible information from a research paper, link it to datasets, tag the pre- trained models and built a deep learning knowledge base out of it. We have further used Natural Language Processing techniques such as Lemmatization, POS tagging, Named Entity Recognition to process the filtered data from research articles/papers and obtain Triplet <Subject, Predicate, Object> format. These triplets are used to validate the extracted meta data and use to fill in any missing data. The application links a Model -> Source Code -> Research Paper, which has great potential in understanding the models based on their framework. We have created a web application where users can upload models/documentation, view existing models from the created knowledge base, categorize the models based on framework, language, the class name. The application also visualizes the model and its architecture in detail and also presents the user with pre-trained models (if existing). We have extracted images, tables, topics, categories, datasets and pre-trained models from various sources and linked these details to a publication. In this way, the application and its knowledge base present a system for users to host, find, compare, study and analyze the deep learning models, which help the developers in understanding the model's behavior. Using this we can also find model similarities based on architecture and this also gives scope for building a recommendation system for the future.

Name thesis:degree_name

M.S. (Master of Science)

Level thesis:degree_level

Masters

Discipline thesis:degree_discipline

Computer Science (UMKC)

Grantor dc:publisher

University of Missouri -- Kansas City

Year dc:date.issued

2019

Author dc:creator

• Punyamurtula, Ruthvic

Advisor dc:contributor.advisor

• Lee, Yugyung, 1960-

Language dc:language.iso

en_US

Handle dc:identifier.uri

https://hdl.handle.net/10355/69014

OAI identifier oai:identifier

oai:mospace.umsystem.edu:10355/69014