Information Technology Reference
In-Depth Information
object for Max/MSP, to establish an internal representation of the musical envi-
ronment, a composed generative music system, and a process of machine-learning
establishing a connection between the two. A generative music system designed
by Young acts as a flexible, parametrically controllable improvisation system with
composed elements. A feedforward neural network is then used to learn a set of
mappings from the musical environment to the parameters of the generative system
in real-time as the performance is taking place. Young's systems exhibit elements
of mirroring and shadowing in their generative rules, and come close to a notion of
negotiation, as the system continually updates an internal model of what is happen-
ing in the music (without predetermined rules governing what is learned or when),
which it can try to manipulate.
Similarly, Bown and Lexer (
2006
) have explored the use of recurrent neural net-
works that exhibit simple low-level dynamical behaviours such as repetition with
variation and coordinated activity with an input. These networks can be embedded
in a Live Algorithm by hand-coding connections between standard audio analysis
tools and the recurrent neural network at one end, and between the recurrent neural
network and a stochastic generative music system at the other end. In an extreme
case, the recurrent neural network updates at sample rate, receiving the input audio
signal directly, and generating the output audio signal directly from the activation
of a single output node.
6.6 Further Considerations
This concluding section offers some directions for future Live Algorithm research.
The list cannot be comprehensive since it is impossible to predict which route(s)
will further the ultimate aims of the Live Algorithm agenda, but it is expected that
these topics will play some part in the process.
6.6.1 Embodiment
Brooks (
2009
) and other researchers in embodied robotics have argued against the
symbolic, representational AI approach to cognition, favouring instead a physically
grounded framework in which robots are situated in the world (they deal with the
world by perception and immediate behaviour, rather than by abstract representa-
tions and symbolic manipulation), and are embodied in the world in the sense that
their actions have immediate feedback on their own sensations. The complexity of
the environment is a key issue; rather than building fragile approximate models of
the world, the embodied approach utilises the world itself in order to pursue a goal.
The complexity of the world is (hopefully) tamed by working in and with the world,
rather than by attempting to imagine and represent the world. A consequence of
this is that embodied and situated systems can themselves have complex, emergent
behaviour.
