Computational Neuroscience
Florentin Wörgötter
The BCCN-group on Computational Neuroscience is mainly interested in understanding neuronal systems embedded in their environment.
Over decades neurons and brains have mainly been investigated as stimulus-response systems, where the output of the system does not affect its own inputs. Animals (and humans), however, operate differently in quite a fundamental way: Whatever action we perform, it will almost always immediately affect our sensory inputs. In writing these lines I see my fingers moving over the keyboard, feel their touch, hear the clicking of the keys, notice letters appezring bksp bksp bksp bksp bksp appearing, where every typo (feedback error-signal) will enact a correction closing the perception-action loop.
Hence, such perception-action loops are the normal mode of operation of all animals. Within such a loop some signals that come back to the animal’s sensors are of direct relevance (like the seeing of a typo in writing), while others do not immediately drive the loop (like the clicking of the keys).
Fundamentally, it is only the animal/human/agent who can decide which of the arriving sensor signals are relevant for the momentarily existing task and which are not. The behaviour of any creature is therefore controlled by measuring its own inputs (input-control) and normally not by reinforcement from an external observer1. Improved fitness will arise if an animal can do this in a predictive way, hence if it can use predictive mechanisms to anticipate the outcome of its own actions and, to some degree, also the “behaviour of the world”.
Our research agenda: The goal of our studies is to understand how autonomous behaviour arises in animals and agents through the development of complex perception-action loops and the learning of adaptive, anticipatory behaviour through input-control with minimal external interference.
To this end we investigate (using neuro-physiological data as well as robots):
Input: Information processing in the visual system and its use in machine vision.
Adaptation: The biophysics of synaptic plasticity and correlation based learning mechanisms in animals and robots.
Reasoning: Decision making, planning and the discovery of the structure of the agent’s environment.
Output: The sequencing of actions towards goal-directed behaviour.
We believe that every external reinforcement will first by the
animal transformed into an internal signal, measured and compared to
other possible inputs of relevance, before it might or might not be
used to influence behaviour and/or learning.
Lectures:
- Computational Neuroscience for Informatics SS
- Computational Neuroscience for Biology SS
- Computational Neuroscience - Learning
- Seminar Computational Neuroscience SS 2008
Projects involved:
|
|
Homeostasis LearningContact person: Kolodziejski, ChristophAdaptive neuronal systems learn to avoid a disturbance and therefore keep homeostasis are constructed and investigated in this project.
|
|
|
Goal-directed Learning (learning to achieve a goal)Contact person: Kolodziejski, ChristophHow can non-goal directed signals from the environment be combined and be utilized to achieve goal directed behaviour? This is the main question this project is faced with.
|
|
|
Modelling spike-timing dependent plasticity.Contact person: Wörgötter, FlorentinSpike-timing dependent plasticity is related to the differential Hebbian learning rule, which we use also for controlling robots (ISO/ICO rule). This relation can be used to derive a flexible formalism that allows modelling STDP in an efficient way. Thus, we can now investigate STDP local at dendrites or the change of STDP as a consequence of pre- and post-synaptic parameter.
|
|
|
Dynamics of collective behaviour in large scale self-organizing systemsContact person: Not specifiedMost animals usually do not live as solitary creatures. Instead they organize into social groups of sometimes large numbers, then often called a herd or a swarm. In the process of doing so each individual (“EGO”) will experience all others (“ALTER”) as being part of its own environment. A general principle for improved survival is predictive learning, which is learning to predict the near-future development of EGO’s environment. Animals/agents that have such a learning mechanism available will necessarily also try to learn predicting the behaviour of all ALTERs, creating a system where the agents will not only mutually influence each others behaviour but also each others learning.
|
|
|
Action descriptions and action chainingContact person: Wolf, AlexanderRobots that are supposed to operate in a dynamic real world environment will have to be able to perform complex movements. The control of the robot cannot be preprogrammed, but has to be learned, for instance by imitation, and constantly updated. The actions are represented hierarchically: Primitive actions (e.g. moving the arm in one direction) form the basis for complex tasks (e.g. fixing a screw in the wall). Here we examine, how humans perform a sequence of subtasks using a manipulandum, and how the order in which they are perform or decision processes influence the details of the movements.
|
|
|
Decision making and PlanningContact person: Wörgötter, FlorentinThe ability for decision making and planning is a major trait of higher vertebrates. We are especially interested in the link between learning and planning. Fact is that your planning at a certain level will influence what you will experience next and, hence, what you can then learn. But this process also goes the other way round. To do this in embodied (robot) systems is the goal of this project.
|
|
|
Adaptive - Fast Dynamic Walking Robot “RUNBOT”Contact person: Manoonpong, PoramateThe goal of this roject is to design a dynamic, biomechanical system for bipedal walking under neural control and learning. This has led to RunBot, which achieves a speed of about 3.5 leg-length per second and is currently the fastest planar, dynamic, biped robot, which can learn to adapt to the terrain.
|
|
|
Physiology and Modelling of the Visual System (mostly old work)Contact person: Wörgötter, FlorentinThis project contains also a summary (Publications) of the older work of this group. Over years we had been concerned with the physiology of the visual cortex and the LGN being especially interested in the influence of cortico-thalamic feedback. In paralell we had developed models of the primary visual pathway where the most recent model concerns the retina.
|
|
|
Philosophy, The Arts and Public Awareness of ScienceContact person: Wörgötter, FlorentinThis project addresses aspects, which are usually not in the center of interest of a natural scientist. Still, every now and then, we found ourselves in a position that we felt the need to “think a bit beyond” the data and the quantifications treated and performed done in our daily business and to deal with issues of wider interest. Several publications have arisen from this, addressing issues of “mind and body”, which should be – at least – an interesting read hopefully stimulating discussions and some more “thoughts-beyond”.
|