Stacks Image 136

June 18th, after 25 years, Douglas Hofstadter is back in my home town: Turin Polytechnic hosts a "speech" held by him together with Andrea Padova, titled "Può una macchina estrarre l'anima di Chopin?" (i.e.: "Can a machine extract the soul of Chopin?").

Short answer: no, or not yet, at least.

Hofstadter starts with a little story about how he accidentally discovered, during one of his courses on what we could call, for a lack of better term, "Artificial Intelligence", the existence of "EMI", a program by David Cope whose main purpose was to compose "original" music in the style of well known artists.

Of course, the use of verbs like "compose", or "write", or anything that presumes will and desire is always a bit suspect when dealing with programs and Artificial Intelligence.
In fact EMI, at best, could be considered as sort of "remixer". Starting with a group of compositions provided as a sort of "training set", the system worked its way up by recognising and labelling short fragments (according to a sort of "genetic code" with 4 different "signatures") and then applying the same recognize/label algorithm to progressively longer sequences.

By this process, the system was able to extract recurring patterns that were somehow distinctive of the style of a specific artist. The next step was to remix these patterns horizontally (that is, creating a new temporal succession made of these) and vertically (i.e. shifting the same signature fragment up or down in scales).

According to Hofstadter, this approach, while completely different from his own ideas about Artificial Intelligence, was still impressively effective in "capturing" some non-formal aspects of an author's style. In particular, some of the "fake" Chopin's pieces were able to suggest something that Hofstadter can only describe as "Poland spirit".

Oddly enough, the same program, when "confronted" with Bach's oeuvre, failed spectacularly at capturing the "style". Why is it so, muses Hofstadter, considering that Bach has always been considered a sort of mathematical genius? Shouldn't a computer excel at recombining the abstract, pythagorean perfection of Bach, and struggle more with a more romantic, "by the heart" music like Chopin's?

To make us understand more clearly the successes (and failures) of EMI, Hofstadter mixed his speech with Padova's renditions of various pieces, some written by EMI, some written by Chopin (or Bach) and some others written by humans in imitation of these two author's work, and asked the public to try and guess who was who.

(Unfortunately my musical talent is a distant second to my abilities as a photographer, so even if I got most of them "right" it was definitely pure luck - or maybe my brain works a bit like EMI - your pick).

The general consensus (supported by many music experts in the room, and by Padova himself) is that Bach works, especially when it comes to contrappunto, follow plenty of formal and painstakingly defined rules, so the "remixing" approach used by EMI sounds, to the trained ear, like a horrible hodgepodge of well known ideas from Bach's genius.

So, in the end, despite the initial enthusiasm by Hofstader (he discovered EMI a couple of decades ago) and some surprising results from the original software, automatic composition of music - no matter if "original" or "in the style of" - still lacks any creative spark. On the other hand, someone from the public asked, in the final questions/answer segment, if we shouldn't build a "computer" that can actually listen to music, before trying to make one that writes it - and Professor Hofstadter hinted at some research he is doing with his students about it ... maybe next time (hopefully not in another 25 years) we can hear how this turned out.

Stacks Image 137
Hofstadter chatting with one of the organizers before the talk
Stacks Image 138
Hofstadter and Padova at the end of the talk