The Quietus - A new rock music and pop culture website

A Quietus Interview

Blossoming Intelligence: Emptyset Interviewed
The Quietus , October 16th, 2019 07:41

Ahead of their appearance at LEV Matadero and in the wake of their AI-derived LP, Blossoms, Emptyset talk to Kristen Gallerneaux and Bernie Brooks about machine learning and the potential ramifications of applied neural networks

Portrait by James Ginzburg

Emptyset's new LP, Blossoms, is haunted. Not by a ghost, but by an intelligence. Intelligence as a thing, not a trait. Still, it's not unlike an always-there spectral presence. People who claim to have been afflicted by hauntings often describe the uncanny sensation of being forever surrounded by something inhuman, whose actions, because of the thing's inherent inhuman-ness, are difficult to parse. In English folklore, too, creatures like fairies are often depicted as being terrifyingly whimsical, horribly arbitrary - by human standards, anyway. The implication being that their logic, while extant, was more or less unknowable.

These examples aren't perfect but go some distance toward describing the experience of listening to Paul Purgas and James Ginzburg's latest album, released earlier this month via Thrill Jockey. Blossoms, while clearly the work of Emptyset, is also the work of something else figuring things out in its own way - the ever-present intelligence in question: a neural-network-based artificial intelligence system. It's very high tech, very scientific, but the overall effect is that of an Emptyset record possessed.

"From the beginning of Emptyset, the project has been considering processes of emergence. We saw the first album as starting from a void, dimensionality appearing - a point and straight lines forming structures. Our second album, Demiurge, added feedback, creating the beginnings of complex dynamics and chaotic systems. Then, our third LP was thinking through ideas of recursion and unfolding intelligence," Emptyset explain. "Finally, in 2016, Borders looked at how, through gesture and tactility, we could participate in the schematic of this sonic universe that we had been iterating - from cosmogony to self-recognition. Within this lineage, Blossoms considers ideas of evolution and adaptation, possibilities of intelligence creating intelligence, as well as aspects of learning and mimicry."

In a nutshell, through a process called "seeding" Ginzburg and Purgas taught a neural network how to make an Emptyset record using their music as a data set. The end result contains everything one might expect of an album by the duo: clattering racket, booming, euphonic resonances, insectoid buzzing. But these elements almost sound remixed or perhaps repurposed. Again, my thoughts drift back to possession. More than anything, Blossoms sounds like some sort of alien intelligence attempting to speak through Ginzburg and Purgas, as if creating a trial-and-error vocabulary with the data they've so graciously given it.

What made you gravitate towards AI in the first place? What was the appeal?

Emptyset: Over the past ten years or so, we have been considering various aspects of the history of intelligence, represented through literature as well as the domains of cybernetics, biomimetics, and design. With Blossoms, we had a chance to think more broadly about emergence and evolution in a very direct way, shifting into a space where we could address intelligence not as metaphor but as a material process in itself.

During my first blind listen to the LP, I built up this whole narrative in my head about how the AI was responding to actual plant life. And given their botanical titles, it’s easy to imagine the tracks carrying a strong sense of plant-derived intelligence, but I guess there isn’t such an explicit connection? Is it more metaphorical, like growing new work from the mulch of a decade?

E: The term "seed" in machine learning represents the initial value that initialises randomisation. Chaos, in mathematical terms, can relate to the way in which complex patterns emerge within dynamic systems that are dependent on initial states. This mirroring between organic and virtual systems became apparent when listening to hours of results from the neural network - there was something akin to watching a patch of earth becoming populated with a small ecosystem of grass, flowers, pollinators - with every output being in a relationship to both the original data that the system had learnt from and then again with each iteration and pattern that emerged as the system produced new audio.

At the same time, it is also the contrasting of something biological that unfolds from the chemical logic of DNA - with warm and living connotations - that when set against the potentially cold complex automation of artificial intelligence becomes unsettling.

This idea of mulch is also interesting. It carries a positive connotation: the decay of the past becoming nourishment for new, evolved life forms. Whilst there is an inevitability that intelligent systems will eventually create intelligent systems, the process of making Blossoms definitely clarified our curiosity and concern for the form this future evolution may take.

The press release mentions a software model and machine learning system - is this something you developed yourselves? Any interesting challenges that came out of the development process?

E: Over two years, we researched and worked with various programmers operating on the cutting edge of neural-network-based AI systems that could synthesise audio from the ground up. There were a lot of dead ends, failed experiments, and technical limitations we ran into along the way. Likewise, there were very few developers outside of the major tech corporations that could work at the level we required, and we had decided from the very beginning that we wanted to create the project with as independent a spirit as possible.

Much of our research was focussed around reading white papers for various systems and keeping up with the incremental developments in the field. It was only in March of this year that we managed to find a machine learning system that would be capable of realising the kind of results and complexity we were looking for.

One of the biggest challenges was audio quality. As much of the existing system architecture was designed around the frequency range of the human voice, it was incredibly difficult to find methods of realising something close to a full frequency audio output - something closer to music. Luckily, a white paper emerged that offered a breakthrough and from there we were able to construct a prototype software setup quite quickly, and then begin to lay down the foundations of the record.

Typically, your releases have operated on some explicit system of rules. For Blossoms, what were the rules applied to the older material and how it interacted with the new material?

E: The rules that applied to Blossoms were set by the training data we seeded to the neural network. Over the course of Emptyset, we have produced studio-based material LPs and EPs as well as more expanded performances and installations. Training the system on both our studio records and audio from these expanded projects, as well as a new set of improvisations on elemental materials, allowed us to unify the spectrum of our work into one singular data set.

This created a situation where the software was having to find patterns within the divergent material that might not reflect what one might normally think of as sonic or musical correlation. In essence, it was trying to listen to all this audio of a very broad range and somehow make sense from it all, and unify it into one convergent cognitive output - a kind of stream of consciousness. So, this idea of interaction was forced by non-hierarchically placing all these sounds from different moments together into one collection of data and then using it as the seed for the system.

This avoided our early problem, which was an intelligent system resolving to an average of the material it was trained on - as we found out in our initial tests, in which the outputs were producing quite predictable results when seeded with a narrow range of sounds. As with a genetic process, it was adding diversity that ultimately proved fruitful and in turn created a much broader and richer set of outputs.

Can you talk about the structural unfolding of the record? It seems in the beginning like there is an audible process of learning or self-awareness catching up with itself, then a structure locks in, and then the sense of "uncontainable eerieness" - as per Mark Fisher - really starts to unfold with 'Bulb'.

E: Within the context of the record, there are pieces which were built from earlier stages of the neural network’s output. 'Bulb' does represent one of the later products of the system. What was interesting to us was that in the beginning the outputs were extremely simple and uninteresting, and then as it progressed this sweet spot began to appear, something similar to an uncanny valley, where the output sonics felt in a relationship to the source material. [There was] a sense of resemblance, but then, equally, there was something uncomfortably alien about the way in which it was perceiving and understanding musicality - in the Fisher sense something unfurling with an unsettling absence of the human.

Towards the end, the system got so "good" at mimicry that its products weren't necessarily interesting. They were too close to literal combinations and convolutions of the source material. It was too convincing, and likewise, felt too familiar to methods of human production. So, it was this interstitial moment within the process that we became fascinated by, finding ourselves drawn to this rougher, semi-congealed sense of cognition, rather than the more refined and fully formed final stage outputs of the system.

How did it feel to hear an AI so accurately simulate your work? Not too many people have had that experience just yet.

E: It wasn't so much a case of hearing a simulation of our work, it was more the experience of hearing the way in which a pattern recognition system interpreted our work. The initial shock was hearing it produce texturally, harmonically, and rhythmically complex audio that had some kind of "musicality" to it, that was obviously derived from the material it was trained on. In that sense, it wasn't a personal shock as in, "This machine has absorbed something essential about us." It was the shock of hearing that a system was capable of constructing musical and sonic coherency that was both informed by the source material and communicating something alien or mutagenic.

Was there any hardware of note used, or am I correct in thinking Blossoms is entirely software driven and sourced from samples of the processed interactions with wood, metal, drum skins?

E: In order to create a final sound, we placed the pieces within reverb that we generated from an impulse taken in the Trawsfynydd nuclear power station in Snowdonia. We worked there in late 2012 as part of a commission for Tate Britain, and it appeared as part of our release Material. That project had also been a poignant moment for us in terms of trying to relate to systems that humans had built that seemed to have a life of their own, to be so autonomous as to be almost entirely inhuman and alienating. With this in mind, it seemed appropriate to use this reverberation signal as a container to work across the record.

There was a degree of processing needed, compression, EQ-ing, etc. to control some of the dynamics of the material, and we used some additional hardware processes to do this. As the album is produced entirely from assembled neural network sounds, it required some sonic taming in places to enable the raw synthesis to be more manageable to the human ear.

How do the potential ramifications of this sort of tech concern you? For example, do you think a near future in which streaming services like Spotify have phased out artists altogether is plausible?

E: The ramifications are concerning. The results exceeded our expectations in the sense that we didn't expect to be able to create the record directly from the outputs of the neural network. We expected to adapt them or hybridise them with other non-AI derived sounds. Obviously, in a project that accommodates abstraction there is a tolerance for results that are extreme or challenging - those artefacts of the process would not transpose into more popular forms of music or music that depends on a humanness or real-world-ness. But we feel that within five years systems will overcome this.

By the end of our modelling, the results carried so many of the qualities of the source material that they weren't interesting for our purposes, but the ability to perform more accurate mimicry and iteration will lead to the ability to synthesise any sonic aesthetic based on a training data set. Spotify is potentially an enormous data set to train a neural network on, and given enough processing power, it does seem inevitable that it would be able to create new music based on selecting parameters to cross-pollinate, such as "Cocteau Twins" and "ECM jazz". After that first phase, you could have an infinite morphology of generations and generations of new music derived from derivations. Ultimately, it does seem certain that in this environment, curatorial practices will supersede what we think of as creative practices further than they already have - whether or not consumers relate to this process will be seen.

Emptyset play LEV Matadero, Madrid on October 19. Blossoms is out now on Thrill Jockey