Tuesday, February 06, 2007

Brain Based Artificial Agents

Brief notes on "Brain-Based Devices: Intelligent systems based on principles of the nervous system", Krichmar and Edelman (2003), Proceedings of the 2003 IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, p940-945.

Krichmar et al have, for a number of years now, been working on the creation of physical devices (i.e. mobile robots) which are controlled by simulated nervous systems - hence the name brain based devices. The Darwin series of robots are the embodiment of this work, and this paper describes Darwin VIII in particular (I believe that Darwin X is the latest incarnation). The intention of this summary, however, is to look more at the principles involved rather than look specifically at the Darwin VIII device.

Brain based devices are constrained by four basic design principles in accordance with the selectional principles by which natural systems develop and operate:

- The device must engage in a behavioural task
- The device's behaviour must be controlled by a simulated nervous system which reflects the brains structure and dynamics
- The devices behaviour is modified through a reward system which indicates the relevance of sensory information to the simulated nervous system
- The device must be situated in the real world

The fourth point explains the need for mobile robotics: they provide the only practical means of 'embodying' the artificial nervous system in the real world. Given yesterdays discussion, the third point is of particular interest. The need for a value signal of some sort is a central part of the behaviour that Darwin VIII learns. The desired behaviour for this device was to learn how to categorise perceptual stimuli without explicit teaching - its task was to 'taste' blocks with two different patterns on top of them by moving around an arena and gripping them. The two different patterned blocks had different electrical conductivities - high conductivity which 'tasted good', and a low conductivity which 'tasted bad' (i.e. the former elicited a positive response from the value system, and the latter a negative response). The task thus required not only recognition, but also the correct motor controls in order to perform the physical separation of the blocks.

The artificial nervous system has six major elements which make up the simulated brain: an auditory system, a visual system, a 'taste' system, sets of motor neurons, a visual tracking system, and a value system. Please refer to the paper for further details. There were a number of behaviours set a priori, however, the choice of which to use in a given situation is performed by the artificial nervous system. Using this architecture, associations were learned between the patterns on the blocks and their 'tastes'. In addition to this, second-order conditioning experiments could be carried out with different stimuli - associations could be learned between different sensory modalities aswell. One of the more interesting results of this work however was the 'emergence' of perceptual categories - i.e. the ability of the artificial nervous system to categorise the stimuli encountered in its environment.

Following on from emergence of perceptual categories, the discussion turned to the binding problem. The Darwin VIII architecture, in creating these categories, was able to solve this problem, by bringing together disparate pieces of perceptual information, to enable the distinction of objects from the background. It was found that the neural circuits responsible for each of the modalities fired synchronously when an object was detected. Different objects were distinguishable in the differing temporal characteristics of these firing patterns.

As a side issue, the artificial nervous system was composed of nearly 20,000 neural units in the six regions described previously. The initial synaptic weights were initially randomly assigned. This allowed the study of what were essentially different 'individuals' with the same architecture, but different initial conditions. These individuals never displayed the same neural activity, even though the architecture was the same, and the overall behaviours displayed were similar. This may have interesting implications when applied to studies of individual differences in humans.

In the final discussion of the paper, Occam's razor is looked at in terms of the development of cognitive architectures: often, the view is taken that the simplest explanation is the best one by modellers. However, in neuroscience this is rarely the case due to the vast complexity of the brain and the myriad of interconnections and cross-influences that need to be taken into account for a full 'picture' to be developed. A number of levels of organisation can be looked at (from the synaptic to the organism), the examination of each is with merit, however, for the full picture, these must be integrated. Thi is where the authors say that brain based devices come in useful. In addition to this role in studying the human brain, the techniques may of course also apply to the construction of devices and methods for industrial and commercial applications. All based on neurobiological rather than computational principles of construction.

No comments: