Real-time multicompartment Hodgkin-Huxley neuron emulation on SoC FPGA

Good morning everyone and welcome to our new seminar of the Frontiers in Neuroscience Webinar Series. My name is Serena Jacobuy and I am excited to introduce our speaker for today, Professor Timothée Levi from the University of Bordeaux. Professor Levi is an esteemed editorial board member of Frontiers in Neuroscience and will be presenting the latest article published in the Neuromorphic Engineering section.

I would like to remind you to visit our website to learn more about Frontiers and to contact us if you are interested in collaborating. Don't forget to subscribe to our channel on Cassini to receive notifications about upcoming events and to watch previous webinars. Now, I am pleased to turn the floor over to Professor Levi. Please feel free to begin.

Thank you, Serena. I will now introduce our recent paper titled "Real-time Multi-compartment Hodgkin-Huxley Neuron Emulation on SoC FPGA." This work was a collaboration between the University of Bordeaux and the University of Tokyo in Japan.

I will start by discussing the context of my talk. We design biomimetics for neurological disorders. I will explain what biomimetic means to us, show some applications, and then conclude. Millions of people are affected by disorders like Parkinson's and Alzheimer's.

Unfortunately, all of these people are drug resistant, making treatment difficult. However, there have been significant advancements in biophysics interface, particularly in neurons. Various technologies now allow for recording and stimulation at different levels, such as neurons, single cell networks, and even the brain as a whole.

A new field called electrocytic therapy emerged around 10 years ago. It is similar to pharmaceutical therapy, but instead of using drugs, it involves micro electrical stimulation. This therapy was highlighted in a video by Nature journal.

In our lab, we focus on two main areas of research. The first involves designing a biometric neural network. This includes using a small FPGA on the left and observing the output of our Neuron model on the Scope. Additionally, we work on bio resistance to enable bidirectional communication between Artificial neurons and living neurons. All of this research is conducted in real time.

Our objective is to modify, replace, and control biological neural activity. We have two main applications for this work. The first is developing therapeutic protocols, including electrolytics and designing new neuroprostheses. The second application is in artificial intelligence, using neural networks for signal processing tasks. Both applications use the same system, requiring real-time communication for efficient operation. Closed-loop communication, allowing bi-directional communication, is essential for our bio-hybrid platform, which involves communicating with living cells using a biomimetic neural network.

Maintaining real-time operations poses significant challenges, which is why we focus on embedded systems. Setting up closed-loop experiments in this multidisciplinary field can be complex. To clarify, let's discuss the meaning of bio breed and the schematics for its setup. On the left side, you will find the biological preparation, where you can work at various levels such as in vitro or in vivo. In the in vitro setting, different types of neurons, like neuron cultures, can be studied.

You can use slides of the brain or spinal cord, as well as IP cells that can differentiate into various types of neurons. On the right is a different neuromorphic computing platform, typically consisting of a conventional PC followed by different hardware such as FPGA or ASIC. These platforms can work in real time, but have limited resources compared to PCs. Interface components such as amplifiers, filters, noise removal, and signal preprocessing are necessary for both acquisition and stimulation on both sides.

I will now introduce our unique network and explain why I chose a biomimetic approach. Neurons can be divided into three main parts: dendrites, soma, and axon. The main function of neurons is to produce action potentials through ion exchanges facilitated by ion pumps. Neurons are connected through synapses, which can be either electrical or chemical.

Depending on the ion exchange, the voltage will change, leading to the creation of an action potential based on sodium or potassium ionic current. Biomimetics aims to closely mimic biology, rather than just being bio-inspired like many AI systems. The ultimate goal is to replace biological components. To achieve this, it is essential to recreate the action potential in terms of both frequency and shape. The shape of the action potential is important for reasons that will be explained later.

In order to closely mimic biology, it is important to consider the various classes of neurons that exist. For example, cortical neurons have many different classes and types that we aim to replicate. By focusing on biophysical details and utilizing multi-compartment models, we can accurately represent the topology of neurons. We work at the biological time scale to facilitate communication between different components. When creating a neural network, it is essential to incorporate synapses, axonal delay, internal plasticity, and noise. The choice of neuron model is crucial and should be based on the specific application requirements.

A comparison of neuron models can be seen in a table by Izhikevich, with biological plausibility on the vertical axis and implementation costs on the horizontal axis. The simplest model is located in the lower left, while the most complex model is in the upper right. Our team primarily implements the Hodgkin-Huxley model due to its close resemblance to biology. We have also worked with the Izhikevich, Tubul, and Kohno models in collaboration with the University of Tokyo.

Overall, selecting the appropriate neuron model is essential for accurately representing biological processes in neural networks.

The classic simulation step is not ideal for temporal stimulation in this model. It is important to understand the different models and what they can reproduce. In the first column, there are not many models that are biophysically meaningful. The Moorlier model is one example.

In our team, we are working on modeling the sinusoid. The membrane potential of neurons is governed by a differential equation. This electrical schematic model includes different ionic conductances and capacitance representing the neuron's membrane. The differential equation is nonlinear, with ionic conductances dependent on voltage and the probability of opening or closing ionic channels.

Each probability of activation or deactivation is represented by a differential equation. When implementing this, the actual potential is reproduced. Different types of neurons exist, leading to the desire to produce all types of neurons. For example, there are regular spiking neurons, which are excitatory, and fast spiking neurons, which exhibit symmetry. In the brain, there is a prevalence of regular spiking and fast spiking neurons. The main difference between the two is that fast spiking neurons act more like oscillators, maintaining a fixed frequency during stimulation, while regular spiking neurons exhibit frequency adaptation over time.

Unfortunately, the single compartmental model has limitations, especially when dealing with neurological disorders. This is because it lacks spatial dimensions and often does not accurately represent morphology or delays. Additionally, it may not account for different types of spikes, such as te spike or don't you spike. Neurostructures can be affected by diseases like ALS. To address these limitations, both single compartment model neural networks and multi-compartmental models are implemented.

The multi-compartmental model involves compartmentalizing neurons using different compartments based on an approximation of the cable equation. This approach allows for a more accurate representation of the spatial topology of neurons by incorporating a theory of compartments.

In fact, there is a third dimension, denoted by X, which represents all spatial dimensions. Multicompartmental neurons have different connected sections that are divided into segments, also known as compartments. Both terms are used interchangeably.

The conduction of the action potential in each compartment changes its properties as it evolves. Implementing compartmental models, especially in hardware, can be difficult due to the need to use different methodologies to solve differential equations. For single compartments, the forward Euler methodology is commonly used, but it is not stable for multicompartmental models. In these cases, the Crank-Nicholson methodology is necessary, requiring more resources to solve the differential equations.

We published a paper in EMBC three years ago to explain the difference between single and multi, as well as stable and unstable. If you want more explanation, you can refer to the paper.

The state of the art of multicompartamental models is challenging to implement in real time and hardware. Current research shows limited implementations on FPGA using the LIF/IZ model instead of the Hodgkin-Huxley model. Other implementations on GPU and CPU are not in real time, with long simulation times like 2 hours for 1 second of simulation. Our goal is to achieve real-time simulation on FPGA. We use a System-on-Module named CRR from AMD, which combines FPGA with a Processing System (PS) and Programmable Logic (PL). This platform addresses the main drawbacks of FPGA, such as communication with external devices and floating-point operations. The board we use includes external and internal ports for connecting to devices like the Maxwell system for biological cell stimulation.

The K26 board features a Zynq UltraScale FPGA, a processor, and the ability to run Linux. It is affordable, priced at around €400, making it accessible for labs. The board is flexible and useful for various applications.

The implementation on the board includes a single compartment for 1,024 neurons and 2^20 synapses, as published in Nature Communications. Over 1000 neurons and 4 million synapses were successfully implemented in real-time with flexible parameters and fully configurable synapses.

Additionally, a multi-compartmental implementation was achieved with 16 neurons, each capable of having up to 64 segments in real-time. The flexibility of the system allows for the emulation of neurological disorders.

I will focus on the compartmental and use the same hardware topology for both. The implementation may be different, but we use the same board for real-time computation. We have very low latency online monitoring and connect to a computer via Ethernet to access monitoring data, such as spike and raster load, and send stimulation.

And it's quite very important.

In the single compartment model, we use both floating point and fixed point implementations to accommodate the lower precision required for forward Euler methodology. This is why we have both float and fixed point options available. However, for the multi-compartmental model in BioemuS, we must use full floating-point coding. All coding is done in floating point and we utilize an ODE solver for this purpose. This implementation involves using memory and communication through AXI Lite within the SoC, with all computations for unichannel synapses being performed in the FPGA.

First, all parameters are stored in the RAM before computation on the hardware platform. For monitoring, there are multiple output options such as Ethernet or a Wi-Fi module. This is important for creating robots where data can be sent outside the robot. Digital conductors inside the board can be used for monitoring with PMODE. Monitoring can be done through DMA, Scope, or internet, with spikes and membrane potential saved for later analysis. Results will be shown later.

Resources in an FPGA are primarily used for synapses, as there are over a million of them. This allows for ample space for memory. However, the logic part only utilizes around 15% of the FPGA.

Compared to other FPGA boards, the VCU 118 is more expensive but offers more resources. For example, only around 8% of resources are used for this type of implementation. This board is also used for larger scale neural network implementations.

We began working on a multiparameter course application focusing on ALS studies. Specifically, we studied motor neurons using a rat model that was specifically prepared for ALS research. The modeling of neurons was conducted using patch clamp techniques to create a detailed model of the neurons, including recording and imaging for spatial topology. This work was done offline. Pascal Brancheau's paper provides a detailed model and topology of the neuron, which served as a starting point for our own modeling efforts in our platform.

We use the neuron software for modeling and comparing our results. This software, developed by Yale, is well-known for analyzing single or multiple neurons. We use it as a reference point for our comparisons. Starting with the same model, we implement it in both neurons and our platform. Neuron software requires 3 seconds to simulate 100 milliseconds, causing latency. In contrast, our platform operates in real time.

On the right, you can see the results comparing the simulation of software to the hardware implementation in Bioeum. It is evident that the behavior is exactly the same. We emulated for 150 milliseconds and monitored the results. The data can be accessed on another PC or saved to an SD card if desired. Different representations were also made for various dimensions depending on the segment, showing the propagation of the action potential throughout the network. All of this was compared to the software neuron, yielding identical results.

In conclusion, this system operates in real time and is the first multi-compartment HH implementation on FPGA in the literature, which is a significant achievement. Our goal was not to use more neurons compared to larger implementations on GPU, but to maintain real-time performance. This is crucial for conducting long-term simulations to study the behavior and dynamics of neurological disorders over time, which is not feasible with GPU implementations that take hours to simulate just one second.

Being able to simulate in real time allows us to communicate with different types of neurons - healthy, metal, and disorder ones - and even apply stimulation to modify their dynamics. This capability is the next step in our research. It should be noted that while we initially stated there were no synapses, there are actually synapses implemented in the system, although not all connectivity has been fully established yet.

In terms of activity monitoring, the system can provide spiking activity reports directly on the onboard file for monitoring. Additionally, membrane potentials can be obtained from the output, allowing for analysis of shape and frequency.

You can have all those features on board so that if you want to do something else later, you can. I am showing you this because we had some noise, which is very important. We use a synthetic noise model from Alland text and implement it in hardware as biometric noise. This is part of a process. The external delay is also crucial for determining the topology of the network and the connectivity between different neural networks.

We can see the raster plot displaying the membrane potential for each neuron in real time. Our collaboration with Universaly, experts in the organoid model, involved the development of three types of organoids from human IP cells. These include single organoids, assembloids where two organoids are placed close together to form connections, and connectoids where two organoids are placed at a distance and connected through microfluidic channels. Our model of these organoids was compared to biological recordings, resulting in a publication in Nature Communications. The single organoid model showed more connectivity at the surface, while the connectoid model showed more synaptic activity and connectivity at the external part due to the formation of axon bundles through microfluidic channels.

The comparisons between the logical model and the logical recordings, as well as our model, are as follows. We use 512 neurons per organoid, including excitatory and inhibitory neurons, and we conduct an emulation for 5 minutes. During this 5-minute emulation, the hardware operates in real time, allowing for monitoring of all spikes on board. This real-time operation results in the same outcomes as observed in logical recordings.

When a single organoid is present, two different dynamics are observed. As organoids are assembled or connected, they begin to form connections and create a network. Additionally, we introduce chemical inhibition or excitation, such as choline or CQNX, to the emulation process. By doing so, we are able to achieve results that are consistent with biological responses, enabling us to inhibit or stimulate our network effectively.

In conclusion, our research in the frontis paper focuses on the multicompartmental aspect. I demonstrate the emulation and provide guidance on implementation. While the implementation can be challenging, we were able to successfully execute it, representing a significant advancement in the state of the art of multicompartmental model implementation. Additionally, I discuss the single compartment implementation as it is related to the multicompartmental model. The choice between a single or multicompartmental model depends on the specific application and the desired emulation for experimental setups.

We have developed real-time biomimetic systems to study neurological disorders, which will aid in the development of new neuroprostheses. Neuroprostheses involve using artificial hardware to stimulate biological parts for learning purposes, such as in cases of diseases like stroke. We presented the hardware architecture, highlighting the need for precision in the single compartment model and the use of more complex differential equation solvers. Additionally, we showcased some applications of this technology.

The specific multicom model has shown promising results and is a novel system in literature. It will be used to emulate more diseases. A publication was found in Frontiers this year, published in December 2024. For the single compartment model, a publication was also found last year in June in Nature Computation. This new tool provides online access to all data and systems, and can be used by purchasing the same FPGA board.

In the future, we aim to focus on characterizing the biological aspects of our work in order to improve emulation. This includes selecting the number and type of synapses, as well as the overall topology. We plan to conduct more experiments with our neuroscientist collaborators, particularly in motor neurons for LSDs. I would like to express my gratitude to all the PhD students involved in this research, as well as our collaborators at the University of Kyoto, European and French projects, and all members of the lab. Thank you.

Thank you for your attention. If you have any questions, please feel free to post them on the platform. Thank you.

Thank you for your attention. If you have any questions, please feel free to post them on the platform. Thank you.