Why the brain of hawk tobacco is more efficiently trained than computer neural networks

Hawk moth tobacco model organism to study the brain of animals

Despite all the advances in machine learning and neural networks, the principles of operation of computer systems are markedly different from those of biological systems. Scientists still do not understand the fundamental biomechanical mechanisms underlying reliable and fast learning of biological neural networks. Therefore, they continue to carefully examine.

One of the most suitable objects of such research is the olfactory system of insects. In particular, the olfactory system of butterflies (for example, tobacco hawk moth, Manduca sexta ) is a relatively simple biological neural network that demonstrates learning ability. Therefore, the butterfly is an ideal model organism for understanding the mechanics of learning a neural network.

In previous years, scientists recorded neurosignals in individual components of the hawthorn brain. These components have a structure and mechanisms that are standard for biological systems:

• the use of neuromodulators (octopamine and dopamine) for brain training - by the way, in humans they are also important for emotions, mood regulation and other mental functions;
• cascading network structure;
• significant changes in dimension (for example, in the number of neurons) between networks;
• sparse data coding in multidimensional networks (sparse coding);
• random connections;
• presence of a noisy signal .

The very presence in the biological system of some incomprehensible and seemingly unnecessary elements (like the notorious "junk" DNA) actually plays an important role in the functioning of the entire system. We just do not yet understand what exactly this or that element is needed for. Of particular interest is the release of octopamine / dopamine in the learning process. It is not yet clear how such stimulation contributes to the creation of new sparse codes in the mushroom body of the brain. The mushroom body contains approximately 4,000 Kenyon cells that encode smells for recording in the butterfly's long-term memory.

In order to better understand the work of biological neural networks, scientists from the University of Washington in Seattle built a computational model for the tobacco hackweed olfactory system that most closely matches the records of the neuroactivity of his brain and all known biophysical processes that occur there, including octopamine stimulation.

Computational neural network models of the mushroom body (MB) of the insect brain were created earlier, but now researchers have focused on studying the role of octopamine in the process of associative learning, as well as the connection of the mushroom body with the antenna lobe (AL) department. For this, the architecture and neural dynamics of the entire system were modeled, including octopamine stimulation, the growth of synapses, the perception of odors in AL, and the reading of neurons in the downstream information flow.


The scheme of work of the AL and MB departments in the brain of a model organism. 30,000 chemical sensors in the olfactory butterfly system (RN) excite a noisy preamplifier network (AL), which transmits a signal to a layer of plastic sparse memory of a fungoid body (MB), which excites neurons of action (EN), interpreting signals from a fungal body as concrete actions, for example, "fly up." Green lines correspond to excitatory connections, and red lines to inhibitory compounds.

The authors managed to build a computer model of the neural network, which demonstrated the ability of reliable training, at the same time possessing significant similarities with the real biological system. The computer model has revealed critical functions in the moth brain - and their influence on the learning process.

This scientific work will allow in the future to mathematically describe these functions and apply in the development of a framework for more efficient computer neural networks with reliable and fast learning. Biomedical machine learning systems.

Scientists are sure that the identified mechanisms of cascade networks, sparseness and hebbian plasticity will well complement the original principles of neural networks, first formulated in 1962 in the Nobel work of Hubel and Wiesel , who described the hierarchical structure of processing visual signals in the brain of cats. On the basis of this work, in 1980, the world's first mathematical model of the Neocognitron neural network was created, which in many ways became a model for modern deep neural networks. However, this work did not take into account many functional systems in the work of the living brain. The study of tobacco hackweed made it possible to identify and clearly mathematically simulate these elements - the missing elements of AI computer systems.

The scientific article was published on February 8, 2018 on the site of preprints arXiv.org.