Gå til hovedinnhold
0
Jump to main content

Novel Tsetlin Machine Mechanisms for Logic-based Regression and Classification with Support for Continuous Input, Clause Weighting, Confidence Assess...

Darshana Abeyrathna Kuruge of the Faculty of Engineering and Science at the University of Agder has submitted his thesis entitled «Novel Tsetlin Machine Mechanisms for Logic-based Regression and Classification with Support for Continuous Input, Clause Weighting, Confidence Assessment, Deterministic Learning, and Convolution» and will defend the thesis for the PhD-degree Friday 22 April 2022. (Photo: Private)

In addition to overcoming the mentioned limitations, the Tsetlin Machine (TM) with the proposed modifications can perform comparably or better in comparison with state-of-the-art machine learning algorithms.

Kuruge Darshana Abeyrathna

PhD Candidate

Kuruge Darshana Abeyrathna of the Faculty of Engineering and Science at the University of Agder has submitted his thesis entitled «Novel Tsetlin Machine Mechanisms for Logic-based Regression and Classification with Support for Continuous Input, Clause Weighting, Confidence Assessment, Deterministic Learning, and Convolution» and will defend the thesis for the PhD-degree Friday 22 April 2022. 

He has followed the PhD-programme at the Faculty of Engineering and Science at the University of Agder with specialisation in ICT.

Summary of the thesis by Kuruge Darshana Abeyrathna:

Novel Tsetlin Machine Mechanisms for Logic-based Regression and Classification with Support for Continuous Input, Clause Weighting, Confidence Assessment, Deterministic Learning, and Convolution

The human (and animal) brain learns things by seeing, tasting, hearing, etc., through experience.

Machine Learning (ML) is the artificial twin of human learning. ML is a branch of artificial intelligence that focuses on learning from interaction with an environment where experience translates into numerical values.  

The Tsetlin Machine (TM) and continuous values

There exist many ML algorithms. Different algorithms provide different advantages and disadvantages.

Hence, the choice of ML algorithm appropriate to learn from a particular environment depends on the type of the environment and the user requirements.

The Tsetlin Machine (TM) is one such ML algorithm, which is relatively new in ML research.

The TM has unique properties and advantages. However, it cannot be applied when the environment represents experiences as continuous values.

Being inherently binary, the TM does not natively process continuous input and struggles with producing continuous output.

Further, its binary nature can make pattern representation and pattern learning uneconomical for applications that require interpretability, high inference speed, low energy consumption, and small memory footprint.

Additionally, when the user wants to estimate the confidence of the classification, this can't be measured with the TM.

Four changes overcomes the limitations of the Tsetlin Machine 

In this Ph.D. research, we address the above limitations of the TM by changing its architecture and learning procedure in four ways:

  • We propose approaches to deal with continuous values by (1) statically representing continuous input features as binary thresholds, (2) incorporating adaptive learning of continuous values directly in the TM architecture, and (3) recasting the TM architecture to support regression, introducing the Regression TM (RTM) which produces continuous output.
  • We have created a novel architecture with weights attached to the clauses, which learn the patterns and their importance in data. The weights make the learned model more compact, increasing interpretability and inference speed, while reducing memory consumption. 
  • The inclusion or exclusion of input features and their negations in clauses is decided by Tsetlin Automata (TAs). We reduce the energy consumption in the learning phase by introducing a new learning automaton called a multi-step variable-structure finite-state automaton. This automaton makes deterministic state jumps while the TAs in the original TM relies on energy costly random number generation.
  • Finally, we propose an approach for replacing the hard decision function used in the TM output generation with the logistic function. This approach enables measuring the classification confidence.

In addition to overcoming the mentioned limitations, the TM with the proposed modifications can perform comparably or better in comparison with state-of-the-art machine learning algorithms.

Disputation facts:

The trial lecture and the public defence will take place on campus in Auditorium C2 040, Campus Grimstad, and online via the Zoom conferencing app - registration link below.

Professor Frank Reichert, Faculty of Engineering and Science, University of Agder, will chair the disputation.

The trial lecture Friday 22 April at 10:15 hours

Public defence Friday 22 April at 12:15 hours

 

Given topic for trial lecture«A Tsetlin Machine SWOT analysis (strengths, weaknesses, opportunities, and threats) when compared to State-of-the-Art Machine Learning»

Thesis Title«Novel Tsetlin Machine Mechanisms for Logic-based Regression and Classification with Support for Continuous Input, Clause Weighting, Confidence Assessment, Deterministic Learning, and Convolution»

Search for the thesis in AURA - Agder University Research Archive, a digital archive of scientific papers, theses and dissertations from the academic staff and students at the University of Agder.

The thesis is available here:

https://uia.brage.unit.no/uia-xmlui/handle/11250/2990166

 

The CandidateKuruge Darshana Abeyrathna (1991, Kegalle, Sri Lanka) Bachelor degree from Asian Institute of Technology, Thailand and Masters degree Thammasat University, Thailand. Present position: Assistant Professor, UiA

Opponents:

First opponent: Associate Professor Marek Druzdzel, PhD, Bialystok University of Technology, Poland

Second opponent: Associate Professor Kristian Hovde Liland, PhD, NMBU Norway

Professor Andreas Prinz, University of Agder, is appointed as the administrator for the assessment committee.

Supervisor in the doctoral work were Professor Ole-Christoffer Granmo, University of Agder (main supervisor) and Professor Morten Goodwin (co-supervisor)

Opponent ex auditorio:

The chair invites members of the public to pose questions ex auditorio in the introduction to the public defense, with deadlines. It is a prerequisite that the opponent has read the thesis. Questions can be submitted to the chair Frank Reichert on e-mail frank.reichert@uia.no