Machine Learning (Details)
The following provides information about some of the composition techniques which help to communicate certain scientific themes and information related to the study. Further techniques are detailed in a forthcoming thesis.
Ricky first became aware of Professor Lidbury’s study from the ME Research UK periodical Breakthrough in Spring 2017. The article explains that machine automation (referred to machine learning in the study), allows for the rapid surveying of single samples of tissue. As human genomes vary greatly, hundreds or thousands of samples from those with and without a condition would normally be required. This can be very costly, particularly in the field of ME research where funding may be scarce and where it is difficult to access large numbers of “well-defined and clinically assessed patients” (ME Research UK, 2017, p.13)(1). Machine learning however, allows researchers the possibility to analyse accurate data derived from algorithmic predictions, where obtaining thousands of samples might otherwise be viably difficult.
The following provides some examples of composition techniques used to communicate the machine learning aspects of the study:
Computer Calculations -> Timbre
Timbre can be described as the perception of a particular sound quality (where quality refers to the make-up of the sound, rather than a level of ‘good’ or ‘bad’ sound). For instance, a piano and violin may play the same note, but we are able to distinguish the qualities that distinguish a piano sound from that of a violin. Different timbres can even be created on the same instrument, emanating from the way it is played (the articulation).
A Central Processing Unit (CPU) is a computer component which one could describe as being the computer's 'brain'. Its function is therefore, critically important for the processing of calculations involved within machine learning. The timbre of the pizzicato strings articulation from the start, aims to represent the sound of computer calculations (Figure 1). While it is not possible to truly hear the actual CPU inside a computer, the quick, sharp sounds are indicative of the technological function of a CPU or reflect the notion of a robotic-like process at work.
Figure 1: Pizzicato strings and syncopation to reflect a sense of computer calculations, important for machine learning. Chaggar, R., 2021. 2. Machine Learning from The Search for ME Biomarkers.
Decision Trees -> Melodic Motifs
For the investigation of comparing ME participants to healthy controls, the machine learning algorithm called random forest was applied to “identify predictor patterns from the pathology results” (Lidbury et al., 2019, p.2)(2). The 'forest' term refers to decision trees.
Decision trees are the result of an algorithm which categorise (or partition) different data. In Figure 2 on the right, a boundary (y = 0) has been implemented as the best position to sort data according to triangles and crosses. Boundary (y = 2) would not be suitable for a clear separation as it would leave some crosses in the triangle section. On the left, the decision has been taken to sort the data according to more precise boundaries to separate the triangles and crosses into their own quarter.
Figure 2: Illustrating the categorisation of data, whereby triangles and crosses represent two different classes of data. The broken lines are used to partition the spaces they occupy. X and Y are two variables. Yang, Z.R., 2010. Machine learning approaches to bioinformatics. Singapore; Hackensack, N.J.: World Scientific.
When multiple partitions are possible from the given data, a tree-like structure indicating the algorithm process can be illustrated:
Figure 3: A decision 'tree' example illustrating the data sorting. Yang, Z.R., 2010. Machine learning approaches to bioinformatics. Singapore; Hackensack, N.J.: World Scientific.
Melodic motifs and the use of pitch were techniques applied to represent both the decision process and and tree-like formations. The decision process aspect to the study is musically communicated through melodic patterns (motif) exchanges. The motifs represent the notion of data and how it is treated within the machine learning process, rather than the musical notes being mapped to any specific data from the study.
For instance, the descending 'branches' of data (or the development of a tree structure mentioned above) is represented musically in violin 1 and 2, at bar 34 in terms of the motif played by violin 1 followed by violin 2 at a lower pitch:
At bar 35 (Figure 5) a variation of the motif is presented. A response by violin 2, is given which concludes on the first beat of bar 36. Although ‘leading into’ bar 36, a firm restatement of the bar 34 motif represents a decision for that motif to be selected, rather than a continuing with the motif from 35. Disregarding the bar 35 motif in favour of the bar 34 motif, depicts the data-sorting aspect of machine learning. The ascent at bar 37 represents a resetting of the algorithm, much like a machine or robot might reset to a default or starting position upon executing a function.
Figure 5: New motif at b. 35 subsequently disregarded for the motif at b. 34, proceeding to a descension of pitch to represent the decision tree branch characteristic of the algorithm. Chaggar, R., 2021. Can’t Stand Standing from Search for ME Biomarkers.
(1) Abbot, N., 2017. Diving. Breakthrough, Spring 2017(25), pp.13–14.
(2) Lidbury, B.A., Kita, B., Richardson, A.M., Lewis, D.P., Privitera, E., Hayward, S., de Kretser, D. and Hedger, M., 2019. Rethinking ME/CFS Diagnostic Reference Intervals via Machine Learning, and the Utility of Activin B for Defining Symptom Severity. Diagnostics, 9(3), p.79. https://doi.org/10.3390/diagnostics9030079.