01 July 2010

A brain and a robot walk into a bar.. (MEA2010, day 3)

Not surprisingly there are a few posters here on two-way connections between robots and neurons cultured on multielectrode arrays (MEAs). One of them offer an open source software package to do it, called Cult2Robot. The authors use the software to let a culture of neurons on an MEA move a robot in four directions and avoid obstacles, all through a Bluetooth connection. Spike rates in the culture are monitored and whenever they cross a threshold at one of the four edges of the square MEA the robot goes in that direction. If the robot's sensors detect an object in any of the four directions, electrical stimulation is applied to neurons on the opposite site of the array, making them more prone to fire and bring the robot away from the obstacle.

It's a simple principle but it illustrates a point I've been trying to formulate for months:
  1. Activity in brains and neural networks can be understood as a fixed number of neurons with a spike rate 0-200 Hz (120.000 neurons in this particular culture) 
  2. Some patterns of activity result in defined actions (here supra-threshold activity along an MEA edge results in ipsilateral movement) 
  3. Some actions are adaptive, others maladaptive, depending on the circumstances 
  4. Given adaptive sensory- and/or reward-feedback, neurons change their activity to produce more adaptive output (here objects are avoided by stimulation of neurons with contralateral output)
  5. The number of adaptive activity patterns a network can reliably assume in the context of changing sensory- and/or reward-feedback is a measure of its operant control (in brains we call this creativity, intelligence, self-discipline etc; as an infant learns new words, its operant control increases)
  6. By using sensory- and/or reward-feedback protocols and good electrophysiological or brain imaging techniques we can map the dynamic range of activity states a brain or network can assume and explore/model the network properties that determine its degree of operant control
A pressing question is how adaptive various networks can be. Could we for example program the MEA-culture-robot above to move not just in four directions but in the 360 directions of a circle? Could the network learn that certain spatial and/or temporal patterns of network output drive power-moves in the robot (like jumping, climbing or crawling) that scale difficult obstacles, the presence of which might be indicated by specific sensory feedback patterns? The link between specific obstacles and appropriate outputs could be strengthened by application of dopamine... you get the idea. Moreover, to make use of the rich and variable activity of neural networks, they should be given the ability to control the robot's actions along continuums like amplitude, duration, and correlation with other actions. And remember, its not just academic curiosity driving these explorations: a constant theme of this conference has been neural prostheses, and the ability of human brains to generate and respond to many arbitrary patterns of activity along continuums is exactly what gives them the ability to control and respond to brain computer interfaces that restore function and improve lives daily, all over the world, but which are still very immature and problematic.

No comments: