Home Research BCCN I Project Details Homeostasis Learning
Personal tools

Homeostasis Learning

Computational Neuroscience

In general organisms will try to maintain homeostastis responding to a disturbance with a compensatory reaction (often a reflex). The pain on touching a hot surface is usually followed by a withdrawal reflex. Due to the fact that all more advanced creatures possess far-sensors, which react to distance signals (like heat-radiation), animals can learn the correlation between an earlier occurring far-sensor signal and the corresponding – later occurring – near-sensor event. In our example one would like to learn the conjunction that heat-radiation predicts pain, which is a causal coupled temporal sequence of sensor events. Hence such a mechanism serves to learn predicting the environment of the animal/agent in a better way (predictive learning) and as a result the animal/agent will have learned anticipatory disturbance compensation.

Computational Neuroscience Over the years we have developed several algorithms for correlation based temporal sequence learning (ISO, ICO, ISO3-learning), which represent a form of differential Hebbian learning related to spike timing-dependent plasticity (STDP). Note these algorithms – different from conventional reinforcement learning – converge to stable synaptic weights as the consequence of a pure input-condition (input control).

Arm Rotate
Deflection of a mechanical arm disturbed by  a second one. After learning the deviation vanishes (taken from the ICO-Paper)

Currently we are using these algorithms for two sub-projects

Arm Rotate 1) To learn load compensation at a multi-joint robot arm, where the movement of one joint will lead to a predictable (hence learnable) torque change at the other joints. Here it is of importance to learn the “right degree” of compensation, which corresponds to a learned forward model of the system. To this end we are employing recruitment mechanisms similar to those used at real muscles.
Runbot_F2.jpg 2) Furthermore we are using the same learning algorithms to improve the dynamic stability of our fast biped walking robot “RunBot”, which is currently the fastest biped walker existing, attaining a relative speed of 3.5 leg-length per second (human record walkers arrive at about 4-4.5 leg-length per second)

Main cooperation partners:

Belongs to Group(s):