By Joe Greaney
Computers already power modern publicity, but what about creative production?
It’s no secret that artificial intelligence, or AI, has become the backbone of publicity and advertising in almost every respect of the word. From online publishing to printed paper, the decision of advert placement has largely left human hands altogether, as programs and algorithms determine where they might have greatest influence – for better or worse.
It should be unsurprising, then, that platforms such as Spotify, iTunes and Amazon Music should seek to do this also. However, the more novel aspect comes with regard to deploying these techniques as a selling point for their product, rather than contributing to the negative sentiments most consumers seem to share. With almost one third of the British online population expressing irritation from such techniques, how is streaming still so popular?
As the saying goes, it’s not what you know, it’s how you think. In Spotify’s case, the simulated decisions made by a virtual mind employ a variety of techniques to try and recommend its users the songs it ‘thinks’ they would like. The simplest of these is a process called collaborative filtering, a tactic used all over the internet to predict what people like and will want to buy. Simply put: the decision system takes the things you’ve expressed interest in, examines other user profiles that share those recorded interests, and recommends things to you that are on their profiles, but aren’t on yours, and vice versa. Due to its simplicity, however, this method is inherently flawed – new or unknown content on profiles are never recommended to anyone and popular content becomes inescapable. Whilst a myriad of factors on Spotify, such as data about artists, albums, playtimes and lyrics can inform this approach, modernity presents a more novel and useful solution – deep learning.
Although the deep learning approach is much more rigid than the popular press would make it seem, with the differences and similarities of content inferred from the collection provided and the rules on how to ‘read’ for these determined by a team of human beings, the degree of predictability is as infinite as the number of factors it is permitted to generate. In Spotify’s case, music is converted into a visual image representing the volume along different frequencies. Divided into separate ‘frames’ over and over again, each step of division is analysed and compared amongst several ‘neurons’ – each a collection of inputs from the group of neurons before it, and outputs to the group coming after. The results of this analysis are plotted on a graph, grouping similar results together along axes of factors the system deems key to their differences, via more than 250 automatically generated content filters.
This cross comparison is used in combination with the collaborative method before to not only determine what people like based on labels, but based also on what this algorithm ‘thinks’ it sounds like. The distance between artists along each of the separate axes informs Spotify’s systems on their audible differences, with this factor weighed against what it knows people already like. The results for likes and dislikes from the radio feature are also considered, although as they reportedly contradict predictions made by the AI more often than not, they do little to sway the system’s decision.
Knowing how these systems work, it can be daunting to stand out among the crowd, as the pressure to sound like everyone else and be liked by the masses becomes more important to business than ever. If you know what to do, however, these mechanisms can turn into your greatest weapon – if automation is informing advertisers, then perhaps it can inform content creators on what they should make.
On platforms like YouTube, this is already happening. This chase for popularity only inspires the proliferation of the same subject matters, as videos are deconstructed, reassembled and uploaded en masse. After seeing its success, a new question arises – what if it were replicated elsewhere?
For Spotify, the solution is clear. The algorithms are public, the processes openly published. So what if the more technologically informed songwriter were to analyse the most popular contemporary tracks of a given period, intercept the inferred patterns, and apply the results towards the creative process? Pop music is already perceived as formulaic to the extent of it being a common comedic trope, and with the advent of deconstructed instrumentals and simplistic lyrics pioneered by artists such as SOPHIE and labels such as PC Music, the capacity for noise in data inputs is dramatically mitigated, and so the outputs should prove less abstract than typical examples of similar production techniques.
Whilst such a process, then, seems not only novel and exciting but extremely profitable as well, it seems a sure eventuality that one day it should take the music world by storm. However, with machines on both ends of creative industry, the possibility of content painfully becoming equal parts bizarre and monotonous should, hopefully, keep music and song writing interesting and genuine, and very much from the heart. Otherwise, imagine a microphone left next to an amplifier for the next forty years.
©Joe Greaney 2019