Biography
Biography: John Kontos
Abstract
The poster is proposed to conceptualize different Machine Learning Methods, e.g. the ANN-based ones, as equivalent to different String Matching methods for retrieving from a table of examples. The aim of the proposal is to clarify what Machine Learning methods can be reduced to in order to avoid the anthropomorphic concept of Machine Learning. Such a reduction may also help in converting “learned” systems to “explainable” systems that enhance trust. The table below shows a correspondence between String Matching methods and Machine Learning methods.
String Matching method Machine Learning method
1. Sequential Serial
2. Non-sequential Serial (Decision Trees (DT))
3. Weighted Total Parallel (Single ANN)
4. Weighted Partial Parallel (Multiple ANNs)
5. Hierarchical Weighted Partial Parallel (“Deep” ANNs)
First we notice that Method No.2 differs from No.1 in the different order of symbol by symbol matching procedure. This change of order by computing a DT (e.g. ID3) is meant to reduce the computing effort necessary for matching between a set of examples and a string under test. Since methods Nos 3 to 5 according to recent literature may be reduced to DTs we may consider them as variants of serial string matching. Future work must analyze the “unseen” example cases.