School of Informatics
Univ. of Edinburgh
James Clerk Maxwell Building
The King's Buildings
Mayfield Road
Edinburgh
EH9 3JZ
Email:
As part of my of my PhD I have been developing C++ software to simulate neural networks. This software is available for use and further developments by others. Go To CNS
I am a PhD student at the Institute of Perception, Action and Behaviour within the School of Informatics.
My research is within the general area of biorobotics - the methodology of building robot models to evaluate biological hypotheses (for a review see Webb 2002, Webb 2000). In biorobotics investigations the aim is "to gain an understanding of how the sensory and neural systems of animals, embedded in appropriate environments, successfully control behavior. " (Webb 2001b). Such understanding can be gained by attempting to replicate animal / environment interactions in robot / environment interactions; i.e. by asking what control problems must the robot solve to display qualitatively comparable behaviour to the animal? Understanding can also be gained by evaluating existing technological and biological hypotheses, sometimes side by side. The "essence of the biorobotic strategy," says Grasso (2001), is to "construct a robot that is competent to test a hypothesis or set of hypotheses that have been suggested by the biology and then allow the robot s behavior to inform you of the acceptability of that hypothesis."
The animal/environment interaction I am examining is context generalisation in classical conditioning. Simply put, this is the problem of understanding how an animal "identifies" the stimuli which predicts a later reward or punishment. Ito et al. (1997) present a study where this ability was disrupted in the fly Drosophila, and suggest a model for this behaviour. My aim is to evaluate an implementation of their model using a robot. More details will be added to this page as work progresses...
References
Institute of Perception, Action and Behaviour
School of Informatics
Univ. of Edinburgh
James Clerk Maxwell Building
The King's Buildings
Mayfield Road
Edinburgh
EH9 3JZ
Email:
In order for a robot to exhibit learning behaviour, it needs a control system incorporating some adaptive ability. For control systems based on neural networks, this in turn means that the neural network as a whole, or neural elements individually, need to be able to change adaptively. A constraint added for real-time control of robots is that these networks can't take too long to simulate.
Following the synaptic plasticity hypothesis proposed over 100 years ago by the eminent neuroanatomist Santiago Ramón y Cajal, I am implementing a neural network learning rule which alters synaptic weights. The mechanism is based on the cellular mechanisms underlying learning in the small sea-snail Aplysia.
The diagram below shows a simplified representation of the neural circuit underlying the withdrawal reflex of Aplysia (taken from Memory: From Mind to Molecules). Important contributions to learning are made at the sites where synapses (termed modulatory) make connection to other synapses (their targets).
These synapse-on-synapse connections have been investigated for many years, with the result that a great deal is known about how such connections contribute to a strengthening of the targeted synapse. The diagram below offers just a hint of the complexity of the myriad cellular mechanisms that combine modulatory input (via the serotonin neurotransmitter) to synapse activity from spiking (via the cAMP concentration) to enhance transmitter release.
Traditional connectionism, and neural networks for robot control, model synapses in a very limited way, typically describing them with a single scalar weight. Weight updates rules are generally of the Hebbian kinds. Uncontroversially, there is clearly a great difference between these simplified models and real neurons, and in terms of learning mechanisms, specifically with the above example from Aplysia, there are important structural differences.
Learning models inspired on the mechanisms observed in Aplysia have recently be proposed and tested for robot control (i.e. the work of Bob Damper et al.) My own work continues along this path, attempting to incorporate, into simplified models, significant features of the cellular mechanism underlying learning in real neurons.