Learning in silico

Researchers investigate learning processes by simulating a cubic millimeter of the brain

Freiburg, 24.04.2007

The ability of the brain to learn lies in the special properties of the nerve cells and in particular of  their connections, the synapses. All brain activity is mediated by information in the form of short electric impulses that are passed from one “firing” cell to the next. In so doing, the cells cultivate their capability to propagate signals. If cell A emits a pulse that evokes a response in cell B, this strengthens the contact between cells A and B. If there is no such causal relationship, or if cell B fires before cell A, the connection is weakened. As a result of this phenomenon, known as “spike-timing dependent plasticity” (STDP), frequent pairings cause strong neural pathways to develop. Conversely, connections which are infrequently used decline. This “plasticity” of the brain, its ability to adapt physiologically and structurally, is considered to be the foundation of learning. On the basis of a complex computer simulation of 100,000 neurons with 10,000 contacts each – corresponding to about one cubic millimeter of cortex - Abigail Morrison, Ad Aertsen and Markus Diesmann have discovered that STDP may be insufficient to explain the learning processes of nerve cells. The results of the scientists from the Bernstein Center for Computational Neuroscience, the University of Freiburg and the RIKEN Brain Science Institute in Tokyo will be published in the July issue of Neural Computation.

From earlier studies the researchers knew that their computer simulation reproduced many dynamical properties of living cortical tissue. The virtual neurons fire with about the same frequency as in the brain, and the activity neither escalates nor decays – the system exists in a “dynamic equilibrium”. Now they have extended their model to take the plasticity of neuronal connections into account.  To this end, Morrison developed a mathematical formulation of the STDP learning rule that fitted the available experimental data significantly better. This development allows the model to become considerably more realistic.

To investigate whether the computer model can simulate learning processes, the researchers repeatedly stimulated a specific group of neurons. Their initial observations were in agreement with the predictions of the learning model: as the stimulated neurons transmitted the stimuli to their downstream neurons, these contacts were strengthened. However, this occurred  at the expense of  the contacts from their upstream neurons in the network.  As the stimulated group responded to the external stimulus, their other inputs became redundant and decayed. After a while, the researchers determined that the entire stimulated group had decoupled itself from the rest of the network.

STDP is therefore not sufficient to explain learning in large neuronal networks, additional requirements must be satisfied to enable the system to learn. There are already strong indications as to what these requirements might be. With the simulation of large networks, Morrison and colleagues have a powerful tool to appraise a variety of different models and uncover the secret of neuronal learning.

Source: Morrison, A., Aertsen, A., & Diesmann, M. (2007). Spike-timing dependent plasticity in balanced random networks. Neural Computation, 19 (6) 1437-1467
http://www.mitpressjournals.org/doi/abs/10.1162/neco.2007.19.6.1437

Drucken