Your source for the latest research news

Friday, 31 January 2020

A new type of artificial neural network inspired by the human brain



The artificial intelligence was an outstanding technological development in recent years. The development of ever more optimized neural networks allows AI to solve complex tasks and learn new solving methods on its own. However, this adaptability shows its limits: when contextual conditions change, AI often has difficulty adapting directly to these variations. In humans, this adaptation is due to neuromodulation. This is why a team of researchers has tried to reproduce this cognitive capacity to adapt it to a new type of neural network, and the results have proved very satisfactory.



Despite the immense progress made in the field of AI in recent years, we are still very far from human intelligence. Indeed, if current AI techniques make it possible to train IT agents to perform certain tasks better than humans when they are specifically trained, the performance of these same agents is often very disappointing when put in conditions (even slightly) different from those presented during the training.

Human beings are able to adapt very effectively to new situations using the skills they have acquired throughout their lives. For example, a child who has learned to walk in a living room will also quickly learn to walk in a garden. In such a context, learning to walk is associated with synaptic plasticity, which changes the connections between neurons, while the rapid adaptation of walking skills learned in the living room to those necessary for walking in the garden is associated to neuromodulation.


Reproducing human neuromodulation in the context of artificial intelligence

Neuromodulation changes the input-output properties of the neurons themselves via chemical neuromodulators. Synaptic plasticity is the basis of all the latest advances in AI. However, no scientific work has so far proposed a means of introducing a mechanism of neuromodulation in the networks of artificial neurons. This result, described in the journal PLOS ONE , is the result of a collaboration between neuroscientists and researchers in artificial intelligence from the University of Liège.

To implant artificial neuromodulation, the researchers created a neural network made up of two sub-networks: the first collects and analyzes contextual information, and the second processes this information to decide what actions to take. The first thus acts as a neuromodulator on the second, so that the whole network adapts quickly to contextual changes. Credits: Nicolas Vecoven et al. 2020


These ULiège researchers developed a completely original artificial neural network architecture, introducing an interaction between two subnets. The first takes into account all the contextual information concerning the task to be solved and, from this information, neuromodulates the second sub-network in the manner of the chemical neuromodulators of the brain.

Artificial neuromodulation: it allows an effective adaptation to changes

Thanks to neuromodulation, this second sub-network, which determines the actions to be performed by the intelligent agent, can therefore adapt very quickly to the task at hand. This allows the agent to efficiently resolve new tasks.

This innovative architecture has been successfully tested on classes of navigation problems for which adaptation is necessary. In particular, the agents trained to move towards a target, while avoiding obstacles, were able to adapt to situations in which their movement was disturbed by extremely variable wind directions.



Teacher. Damien Ernst: “The novelty of this research is that, for the first time, the cognitive mechanisms identified in neuroscience find algorithmic applications in a multitasking context. This research opens perspectives in the AI ​​exploitation of neuromodulation, a key mechanism in the functioning of the human brain”.


Bibliography:

RESEARCH ARTICLE
Introducing neuromodulation in deep neural networks to learn adaptive behaviours

Nicolas Vecoven, Damien Ernst, Antoine Wehenkel, Guillaume Drion

PLoS ONE 15(1): e0227922.

doi:10.1371/journal.pone.0227922

No comments:

Post a comment