Below are some insights from Professor Peter McOwan in response to some questions from the Wellcome trust about the project and thoughts from other team members:
On the intent of the work to ‘mimic’ biological systems
There are a number of interesting biological analogs in the system. The system learns the difference between types of music, and the robots dance preferentially to punk.
Auditory processing
The computer program first does a spectral analysis, similar to the first stage of processing in the human cochlea, and finds the pattern of tones in the music, and from this extracts the features that define the music.
The system then uses a neural network, a computer model of how the weights in brain neurons change as we learn, to discover the difference between patterns of different types of music. It does this by ‘listening’ to a range of music, the network is then told if the music is punk or not (so called supervised learning), and through these experienced examples and changing the weights in the network (learning and memory) its able to find the typical patterns that indicate punk music. Once its taught, the system can give a punkyness rating to any new music it hears
Learning, auditory receptive fields and ‘taste’
In effect the system has learned a set of acoustic receptive fields, in analogy to the auditory cortex, and these receptive fields will be preferentially excited by punk music. Different networks can be trained with different sets of music, so in effect the different musical experience will lead to potentially differing auditory receptive fields, individual experience will shape individual response and arguably develop individual ‘taste’ and ‘aesthetics’.
Motor neurons and dance
The robots have a series of defined dance movements; these would correspond to motor units in the brain, an analogue of the motor map of motor receptive fields in the brain which execute particular muscle movements when stimulated. Punk music excites these robot motor units (sensory motor integration). The actual pattern of dancing for a particular given level of punkiness can differ between robots, this could be considered a type of ‘personality’.
Mirror neurons in the audience
The robots will not have any sensory input other than the music, they cannot observe others around them for example, of hear the actions of others around them. Therefore the motor neurons concept, which requires feedback from the environment e.g. same neuron fires when doing an action as seeing an action (visual mirror neurons) or hearing an action (auditory mirror neurons) wont actually be built into the robots. When a robot acts believably with a particular move in response to a particular sound feature, human observers ‘joining this dance’ will in effect be exploiting their mirror neurons, the same neurons that see the movement will do the movement.
The role of mirror neurons in potentially developing empathy can also come into play here in discussions.
Mirror Neurons in the robots?
However it is possible to argue that if particular sounds, for example a loud guitar riff from the band performer was always associated with a particular pogo move (a high jump for example) then the robots in hearing that rift (which is the sound associated with the action of jumping), performs a high jump, then there is a model for a mirror neuron in effect.
