Circuit bending is an iconic audio production and sound design technique invented in the 60s, whose main premise is to rewire the original purpose of electronic circuit design, capable of producing sound, in a way that leads to unexpected and creative output.

What I find appealing about it is its potential to question and ultimately subvert the given state of things, without the necessity of becoming an expert beforehand. The technique of circuit bending is founded on the concept of “immediate canvas” coined by Reed Ghazala, one of the pioneers of the technique. In the context of electronic design, it means that the modifications can be made intuitively, experimentally, and without a deep insight into the working of the hacked circuit.

Such an anti-design approach encourages free experimentation, unconstrained creativity and the embracement of uncertainty, which results in a production of glitch: an object from the fringe of the compositional thought, conventionally considered undesirable and redundant.

I make a nod towards this anti-design approach by experimenting with the concept of “network bending”, understood as a spiritual ancestor of the classical technique. As a matter of fact, the concept of network bending is not so new, and even the term itself has been already applied within a similar context. What differentiates my interpretation, is the focus on the foundational notion of immediate canvas. In my research in this field, I look for techniques that are highly accessible, both in terms of computational requirements and expertise required for experimentation. 

Image: Network bending interface designed by Błażej Kotowski for AMRO festival. 

Inference of a neural net can be understood as a result of transforming the input information by a very complex, non-linear mathematical function, driven by a very large set of numerical parameters called weights and biases. These numbers are learned by the network in the computationally meticulous process of training on prior data distribution (the dataset). Together with the architecture of the network, they constitute the final identity of any pre-trained model. The architecture of a network is a design that arranges these weights and biases into a circuit-like structure defining the flow of the information. The allegory of bending the network in the spirit of bending the electronic circuit, can be understood as intervening in the numerical and architectural identity of the pre-trained model.

Image: Workshop on network bending at “Art Meets Radical Openness”, in Linz, AT. Credit: Violetta Wakolbinger

Recently, for the AMRO festival I have prepared a workshop on network bending. I prepared a set of tools, including an intuitive interface, through which participants could experiment with various techniques of network bending, creatively, and often unexpectedly modifying the behavior of the pre-trained neural net model. The results of such a playful practice could be seen as leading to the failure in function of the neural network. Indeed, depending on the depth and nature of the modifications, the process often takes the produced sound far away from the original intention of the model, but on the way, a world of many strange and interesting sounds is revealed. The sounds that otherwise reside in the inaccessible realm constrained by the thought of the model's original author.

I myself find this world fascinating, and in my research I try to systematize and document it by coining new terminologies, workflows and reproducible approaches to network bending. So far, it has proven to be a very inspiring direction that I incorporate into my work as a sound artist. I have recently presented the results of this work during the AMRO festival. I will be presenting the work again on Tuesday, 11th of June, in the framework of +RAIN festival, under the project name of “Disturbance, De: Configuration”.

There will also be a chance to play with the technique, and the described tools during the Generative Music AI Workshop, where I will take the role of a teaching assistant, working for 5 days with a mixed group of 30 music technologists and musicians, exploring the concepts of ethical and creative generative AI.

Image: Performing with use of network bending techniques at AMRO, Linz, AT. Credit: Martin Bruner

About Blazej Kotowski

Blazej will take part in RAIN+ LIVE Festival at UPF, on June 11th 2024.