From Neuronal Models to Neuronal Dynamics and Image Processing

Neurons make contacts with each other by synapses: The presynaptic neuron sends its output to the synapse via an axon, and the postsynaptic neuron receives it via its dendritic tree (“dendrite”). Most neurons produce their output in form of virtually binary events (action potentials or spikes). Some classes of neurons (e.g., in the retina) nevertheless do not generate spikes but show continuous responses. Dendrites were classically considered as passive cables whose only function is to transmit input signals to the soma (= cell body of the neuron). In this picture, a neuron integrates all input signals and generates a response if their sum exceeds a threshold. Therefore, the neuron is the site where computations take place, and information is stored across the network in synaptic weights (“connection strengths”). This “connectionist” point of view on neuronal functioning inspired neuronal networks learning algorithms such as error backpropagation [1], and more recent deep-learning architectures [2]. Recent evidence, however, suggest that dendrites are excitable structures rather than passive cables, which can perform sophisticated computations [3, 4, 5, 6]. This suggest that even single neurons can carry out far more complex computations than previously thought. Naturally, a modeller has to choose a reasonable level of abstraction. A good model is not necessarily one which incorporates all possible details, because too many details can make it difficult to identify important mechanisms. The level of detail is related to the research question that we wish to answer, but also to our computational resources [7]. For example, if our interest was to understanding the neurophysiological details of a small circuit of neurons, then we