Ramin M. Hasani will be giving a talk on Learning with a Worm’s Brain at the Sharif University, Department of Computer Engineering.You may find further information below.
A 1mm-long worm with only 302 neurons in its nervous system, exhibits remarkable behavioral plasticity which are not yet realizable by modern AI systems. Functional and structural properties of the C. elegans’ brain inspired us to think about how to design behavior the way nature does it.
We report: In order to create “behavior” with neural circuits, it is essential to take two inseparable principles into account:
1. The near-optimal evolution of network structures
2. Deployment of many forms of synaptic plasticity (Learning) mechanisms.
We looked into the fundamental architectural and physiological design principles underlying signalling amongst neurons, and designed artificial neural circuits which automatically learn to perform tasks like parking a rover robot, manipulating an arm robot and in general controlling movement. Our brain-inspired networks are substantially smaller in size compared to their deep learning setting counterpart, they are significantly robust to noise, they perform parallel sensory processing and most importantly, they are fully explainable.
An example of our networks automatically learned to perform parallel parking of a rover, comprises a circuit composed of 39 biological model-neurons wired through 80 synapses. The network is initially designed based on the relational properties that neurons form within the nervous system of the worm, and then learned with an evolutionary strategy to perform the task. https://youtu.be/sVrahfZ-sdE
Our work introduces a new paradigm in designing explainable neural processing units which enables us to create reliable behavior for robots as well as distributed multi-agent systems for parallel data processing in arbitrarily-chosen settings.
Ramin is a PhD research assistant in Computer Science at Vienna University of Technology. His main research focus is on developing bio machine learning algorithms for "Modeling and Learning the Analog Behavior”. As of Nov 2017, he will be a visiting scholar at MIT, CSAIL group, investigating novel machine learning algorithms for the control of distributed multi-agent systems.
He is actively collaborating with several Industrial and academic institutions such as: Infineon Technologies, Austria, OpenWorm Foundations, USA, Imperial College London (VAS Group), UK, and Institute of Molecular Pathology, Austria (Zimmer’s Lab). Ramin has completed an M.Sc. in Electronic Engineering at Politecnico di Milano (2015) and has got his B.Sc. in Electrical Engineering – Electronics at Ferdowsi University of Mashhad (2012)