Practicals & Theses

List of practicals and theses at the institute (see also TI Research Presentations)

Details for Deciphering the C. Elegans “Brain”

Deep neural networks are nowadays one of the most active areas of research. Google, Apple, IBM and Microsoft are all on the hunt for people with expertise in this area. For example, Google bought in 2014 the startup DeepMind for $500 Million (see

Unfortunately, the learned deep neural networks, although very effective in classification tasks, for example, remain more or less a black box. One cannot say for sure, what every neuron in the learned network does. This is one of the main obstacles for the wide acceptance of such networks. Scientists and engineers want to understand the systems they design, such that they are able to also make predictions about their future behavior.


Fortunately, nature may come to rescue. One of simplest and most well studied organisms in the world is the C.Elegans worm (see Biologists and physicists have figured out all the neurons within the worm’s “brain”, their synapses, the strengths of the synapses, and their parameters. Moreover, they have even built mathematical models allowing to simulate this “brain”. Below is and excerpt of such a model, describing the withdrawal of the worm in case it is tapped either on the head or on the tail:

The red triangles are the sensorial neurons (PLM is the posterior neuron, and AVM is the anterior neuron), the blue triangles are the motor neurons, and the circles are control neurons. The arrows are chemical activation or inhibition synapses and the lines are electric synapses, with associated strengths. All these values were determined experimentally.

What is not known, however, is why the network is like that. The basic intuition is that when the head is touched, the worm should move backwards, and when it is touched on the tail, it should move forward. However, why does one need all these neurons, and why are they connected this way, it is still a mystery. If one considers that every neuron is a controller, and the links allow neurons to synchronize with each other, then the above circuit is a natural, distributed controller. If one understands the role of each neuron and synapse, then one will be able to design distributed controllers inspired by nature.

Fortunately, one has a model of the neural network, and one can simulate it, such that one can take out neurons and/or synapses, and figure out this way, what is the role they play in this network. The purpose of this master thesis is to perform such experiments, and come up with an explanation of why it was designed this way by nature. This should lead to general principles for distributed controller design.


Univ.Prof. Dipl.-Ing. Dr.rer.nat. Radu GROSU (main responsibility)