Parking with a Worm's Brain

The recent success of deep neural networks (DNN) in unsupervised and supervised learning, in particular the success of DNN autoencoders, convolutional DNN, and recurrent DNN, have lead to an unprecedented interest in DNN, and also to bold predictions of their role in taking over humanity and rendering us useless. However, DNN are considerably simplified, artificial models, that only capture the functional behaviour of the neurone's synapses, and completely ignore the dynamic dynamic behaviour of the neurones themselves. If DNN were the key to learning and to intelligence, why did nature create biological rather than artificial neurones? What are the advantages of biological neurones? In this talk, I show that biological neurones are very well suited to devise, or automatically learn sophisticated nonlinear controllers. We use a dynamic model that captures with decent precision, the behaviour of neurones and their synapses, in the C.elegans nematode. This model is powerful enough for designing or learning a parallel parking algorithm with just 39 neurones, where the role of each neurone is well understood. Moreover, the model turns out to be very robust, and easily extensible to tolerate faults. In fact, the neural model we use represents a Turing complete formalism, where parallel composition, and not sequential composition, is the norm.