Informatics student presents at NeuroMONSTER 2021 conference
Master of Informatics student, Jerry Zhao, gave a presentation on ‘Self-organising deep recurrent neural networks for behaviour control’, a project collaboration with Edinburgh Centre for Robotics student, Billy Lyons, and supervisor Dr Michael Herrmann at NeuroMONSTER 2021 conference. The project suggests control of behaviour brain function is incomplete unless accompanied by observation and evaluation from motor actions.
The NeuroMONSTER conference 'Mathematics Of Neuro-Science, Technology and Engineering, Rhodes' was held on 21st and 22nd September and served as a platform for short talks and posters presenting general mathematical models of brain function.
In Jerry’s exploratory symposium, he explored the notion that the main function of the brain is control of behaviour is incomplete unless accompanied also by ideas about observation and evaluation of feedback from the motor actions as well as the initial generation of behaviour.
In the NeuroMONSTER conference, I shared our group’s ideas on incorporating general principles from brain science in understanding and designing a plausible model of the brain. I also get inspiration from my fellow speakers. I am very interested in these workshops about neuroscience or artificial life, and I believe exploring interdisciplinary field is very useful for my research.
‘Self-organising deep recurrent neural networks for behaviour control’
Jerry Zhao, Billy Lyons, Michael Herrmann
In this report, researchers study the interaction of these components within a paradigm of self-organising sensorimotor loops in a modular counter-propagating neural network. Only the first module of the network is directly connected to sensors and actuators, while each module interacts with other modules by the same rules as the first one with the environment, realising the principle of surface relativity. An additional locality principle implies that the weights in each module adapt based on the local information in the flow of the sensory feedback from the environment and inputs from other modules. The modules aim at maximising an estimate of predictive information. Op- timisation of this objective function leads the network system to a critical point between stability and instability, which entails a sensitive engagement with other modules and avoids global synchronisa- tion. Finally, a duality principle is in place to express that sensing and acting are computationally complementary in a similar way as in classical control theory. Researchers demonstrate the validity of approach using an architecture proposed in which is now more comprehensively analysed. Tests confirm experimentally that the brain-style principles apply. For various simulated and real robotic platforms, researchers show how controllable behaviour of various levels of complexity emerges from the self-organisation of neural feedback, and how it can be further organised into purposeful behaviour based on higher level reinforcement learning.
About Jerry Zhao
Jerry Zhao's research interests are reinforcement learning, self-organization and robot control. In his Honours project, he studied and analysed the homeokinesis learning algorithms and did experiments in real robots. Behaviour generation was archived for different kinds of real robots in efficient and systematic ways. The theory in practice was tested in order to assess hybrid learning algorithms based on deep RL and to prepare also applications in other domains such as search and rescue, agriculture, mining, or mobility on demand.
Particularly, Jerry is interested in constructing a multi-layer control architecture that can explain some functionalities of the brain in terms of long-time planning, self-coherence, exploration and automatic development.
Jerry's work as an intern at IPAB in the summer was also important for developing my research interest. Jerry did some work on a multi-layer DEP model with interesting simulation videos and statistical analysis. Supervisor, Dr Michael Herrmann helped Jerry with support and guidance.