Embedded Files
Topic leaders
Topic leaders
Gert Cauwenberghs (UC San Diego)
Rodolphe Sepulchre (Cambridge)
Expert panel
Expert panel
Eve Marder (Brandeis)
Todd Coleman (Stanford)
Terry Sejnowski (Salk)
Guillaume Drion (U Liege)
John Doyle (Caltech)
Timothy O’Leary (moderator)
Invited speakers
Invited speakers
Eve Marder (Brandeis)
Todd Coleman (Stanford)
Terry Sejnowski (Salk)
Team
Team
Luka Ribar (Cambridge)
Thiago Burghi (Cambridge)
Jun Wang (Harvard)
Soumil Jain (UC San Diego)
Frederic Broccard (UC San Diego)
Alessio Franci (UNAM)
Guillaume Drion (ULiege)
Goal
Goal
Understanding the principles by which neural networks are able to achieve robust and reliable activity in spite of a great degree of uncertainty in their basic components is one of the central questions of both neuroscience and neuromorphic engineering. The reliability of neuronal behavior coexists with an amazing degree of adaptability due to constant learning and changing external conditions, providing one of the key goals for bioinspired hardware to achieve. Neuromorphic control draws inspiration from the feedback architecture of biophysical neural networks in order to address the challenge of learning reliable and adaptive rhythms while being robust to uncertainty in the components.
This topic area aims to demonstrate and explore those neuromodulatory neuromorphic principles in silico using NeuroDyn, a digitally programmable, continuous-time analog neuromorphic system that models the membrane dynamics and channel kinetics of generalized Hodgkin-Huxley (HH) neurons with conductance-based synapses. The generalized form of HH dynamics, supported by the functional flexibility of NeuroDyn, permits complex neurodynamics that elicits transitions between tonic spiking and phasic bursting through adjustment of on-chip programmable parameters. The versatile architecture of NeuroDyn offers participants a real neuromorphic platform to test neuromodulation algorithms designed in software. The topic area aims to utilize the existing NeuroDyn hardware in order to study the design and control of complex neuronal patterns that are at the heart of central pattern generating networks controlling essential rhythmic behaviors such as walking and breathing. The main emphasis will be on robust control and robust learning.
Projects
Projects
Participants will demonstrate in software and in hardware that a small neuronal network can be made reliable in spite of variable components.
The approach will be highly modular: the coexistence of variability and reliability will be demonstrated in a single spiking neuron, then in a single bursting neuron, then in subcircuits of bursting neurons interconnected by inhibitory synapses, and finally in circuits made by interconnecting several subcircuits.
At each step, participants will be given minimal instructions, then design their circuits in software, and finally validate their design on the NeuroDyn platform.
Robustness of the designs in the face of real-world perturbations and hardware variability will be one of the key challenges of the projects to overcome.
Time permitting, participants will investigate simple tuning and adaptation rules to compensate for external perturbations such as supply and temperature variation.
Equipment
Equipment
Participants will be given remote access to a system with four NeuroDyn chips featuring up to 16 neural compartments, 96 gating variables, 48 chemical synapses, and 24 electrical junctions. All components and their 1,560 parameters in the model will be fully adjustable over a programmable user interface on the worldwide web, and the participants will be able to configure voltage clamp or current clamp interfaces to individual neurons to measure and apply voltage and current, just like an intracellular neurobiological test rig except with simultaneous control and measurement over 16 channels, as 16 virtual “pipette electrodes”.
Participant preparation
Participant preparation
We have prepared some material to get you introduced to the challenge. The video tutorial below is a seminar that provides an overview of what we mean by neuromodulation control and of the control objectives that we will want to explore during the workshop.
The main goal for Week 1 will be to understand the basic principles of single cell neuromodulation, and to develop in Python the control algorithm of a "rebound bursting cell". This single-cell control will be the basis for creating and controlling small rhythmic circuits in the following week.
We have assembled different pdfs and notebooks to guide you through those steps.
Reading materials
Reading materials
Please visit the website of the monograph "Neuromorphic Control Principles".
A good starting point is the pdf of Chapter 2, entitled "Single neurons as feedback control systems". There you will learn how to think of a neuronal model as a control system and how to design the controller of basic spiking or bursting neurons. Notebooks are written in Julia, and the webpage contains all the information to install what you need to install in order to run them. The notebooks 2.2 and 2.3 should be of direct interest to the worshop.
See also the following arXiv preprint:
"Neuromorphic Control," Luka Ribar and Rodolphe Sepulchre, arXiv:2011.04441.
An introduction to the NeuroDyn hardware and its modeling in software is in the following papers:
"Analog VLSI Biophysical Neurons and Synapses with Programmable Membrane Channel Kinetics," T. Yu and G. Cauwenberghs, IEEE Trans. Biomedical Circuits and Systems, vol. 4 (3), pp. 139-148, 2010.
“Biophysical Neural Spiking, Bursting, and Excitability Dynamics in Reconfigurable Analog VLSI,” T. Yu, T.J. Sejnowski, and G. Cauwenberghs, IEEE Transactions on Biomedical Circuits and Systems, vol. 5 (5), pp. 420-429, 2011.
“Assimilation of Biophysical Neuronal Dynamics in Neuromorphic VLSI,” J. Wang, D. Breen, A. Akinin, F. Broccard, H.D.I. Abarbanel, and G. Cauwenberghs, IEEE Trans. Biomedical Circuits and Systems, vol. 11 (6), pp. 1258-1270, 2017.
“Neuromorphic Dynamical Synapses with Reconfigurable Voltage-Gated Kinetics,” J. Wang, G. Cauwenberghs, and F. Broccard, IEEE Transactions on Biomedical Engineering, vol. 67 (7), pp. 1831-1840, 2020.
Notebooks
Notebooks
Please visite the website "Telluride" for more specific instructions to the Telluride challenge.
This website contains three notebooks directly aimed at the Telluride NMC challenge.
Notebook 1 is a design methodology to design spiking or bursting neurons from elementary circuit design principles.
While the models in the monograph are conductance-based, the models in that notebook are circuit-based. Depending on your background, you might want to start from conductance-based models or from circuit-based models. Ideally, you might want to experiment both as they are closely related.
Eventually, you will want to design a bursting neuron from the interconnection of Hodgkin-Huxley conductance-based models, as those are the neurons programmed in NeuroDyn.
Notebook 2 will guide you through converting a Hodgkin-Huxley model into a model that can be mapped to the NeuroDyn hardware. You should not start Notebook 2 until you have understood how to design a bursting model from the interconnection of two Hodgkin-Huxley models.
Notebook 3 will guide you to the specifics of creating the bursting model in NeuroDyn. We hope that each participant can use that notebook as a starting point to Weeks 2 and 3.
Please use the slack channel nmc to ask your questions and interact on this particular challenge.
Video tutorial – INC Chalk Talk
Video tutorial – INC Chalk Talk
Page updated
Google Sites
Report abuse