Martin N P Nilsson
Update 2021-10-05: The article “Opålitliga nervceller?” by me and Henrik Jörntell was published in the Swedish clinical magazine Neurologi i Sverige no. 3-2021; Update 2021-03-11: Added a Radio Sweden interview on absent-mindedness; Update 2021-02-21: The neuron model I have worked on for a long time has now been published in Physical Review E! I have written a popular summary below under the heading “What are all those neurons doing, anyway?”. |
This is my private web page, describing my research in a popular way, followed by some other topics (see here for my professional page). If I’m not the Martin you are looking for, I keep a table of Martin Nilsson researcher namesakes. You can mail me at from.web@d…….m, where d…….m is the domain address of this web page.
Professional activity
I’m a mathematical physicist and associate professor (“docent” in Swedish) at RISE. I work in multiple disciplines, e.g., biophysics, signal processing, mechatronics, robotics, and computer science, but my research focuses on machine learning and AI. I’m especially interested in neurobiophysics and the brain’s tipping us off on how knowledge can be efficiently represented and processed in computers.
The brain includes a computer
Fig. The brain is a computer of the von Neumann variety (with some extras, because it is embodied).
Having spent several years at a computer architecture lab in Japan, I don’t hesitate to say that the brain is a computer of the von Neumann type (Fig. The brain). A computer architect might describe the human brain in the standard computer architectural lingo as an extreme-VLIW (Very Long Instruction Word) symmetric multiprocessor with nightly write-back cache (the hippocampal formation) and a high degree of instruction-level parallelism. If you accept the view of the brain as a computer, it becomes obvious that a significant piece of software is running on this computer. And it is also quite evident that knowing the entire brain “hardware” (a.k.a. the connectome) is by far insufficient for understanding the brain as a whole. There is no quick and easy machine learning algorithm that will suddenly solve this problem.
The key, I believe, is that we aren’t born with all that software, but acquire it during childhood. Finding out how the brain bootstraps, or how it learns to learn–call it metalearning if you like–is perhaps the most promising approach towards understanding the brain’s higher levels of abstraction.
If you are interested in this topic, feel free to watch the following video from a RISE seminar where I attempt to present it in a popular way: Beyond Deep Learning - What can biology teach us?
By the way, a great book discussing the brain by layers of abstraction is Ballard: Brain computation as hierarchical abstraction (MIT Press, 2015).
What are all those neurons doing, anyway?
In any case, to understand what the brain is doing, we must first understand what the nerve cells (neurons)—its building blocks—are doing. This is a non-trivial problem which requires fairly advanced mathematical tools for a full analysis, but I will stay away from diving deep into formulas here. By “doing”, I’m referring to the electrical signaling between neurons that carries information, responsible for most of the neuron’s energy consumption.
A neuron consists of dendrites, which receive signals from other neurons; the axon, which conveys signals on to other neurons; and the cell body (soma), which connects the dendrites with the axon. A somewhat different division serves better for understanding the electrical signalling. Here, soma and neighbouring (proximal) portions of the dendrites together constitute one compartment; the more faraway (distal) dendrite sections form another compartment, and the axon initial segment (AIS) forms a third one (see fig. Purkinje neuron).
Fig. Purkinje neuron and its division into three compartments (CC BY 4.0)
The electrical signals are sequences of voltage spikes (action potentials). However, it is not the shape of the spikes that convey information, but the distance between the spikes. What has puzzled researchers is that these distances vary substantially, even when input is pharmacologically blocked. This effect cannot be explained by the classical Hodgkin-Huxley model, but in a paper I published with the neurophysiologist Henrik Jörntell in Physical Review E in February 2021, we show that it is possible by our three-compartment model. The main cause of the variations appears to be thermal noise affecting passages (ion channels) in the wall (membrane) of the proximal compartment.
In connection with the publication of the Physical Review E paper, there was an interview intended for laypersons in Radio Sweden.
In the three-compartment model, the input is summed (integrated) by the distal compartment, whereas the proximal compartment creates a voltage ramp and adds noise, followed by the third compartment probing the voltage level and generating a spike upon passage of a certain threshold voltage.
Fig. Model agreement with experimental data. The red solid trace marks the model’s theoretical probability density, and the blue bars is the histogram of of interspike intervals, whereas black dashed trace marks the kernel density estimator (like a histogram but better). (CC BY 4.0)
The agreement between the model and the experimental measurements is spectacular, despite the model being mechanistic and having only the minimal number (three) of free parameters (Fig. Model agreement). “Mechanistic” means that the model is built upon underlying biophysical machinery, and not on “curve fitting”. This means that we can view the model as an explanation of how the signalling actually works.
The breakthrough which enabled us to find the model was the solution of the first-passage problem I published in Journal of Physics A: Mathematical and Theoretical in October 2020. This article describes a method for computing the probability for a stochastic (random) process to pass a time-variable boundary, and can be directly applied to the neuron, where the stochastic process represents the neuron’s internal voltage (membrane potential). I held a short presentation on this method at KTH (Royal Institute of Technology) in February 2021, and a video including subtitles is available on YouTube.
Well, what exactly does the neuron do, then?
I will sketch a bit more in detail how the neuron operates. This description assumes that you are somewhat familiar with electrical circuit diagrams. I start with the famous Hodgkin-Huxley (HH) model from 1952, for which Hodgkin and Huxley received a Novel prize in 1963. There are lots of material on it available elsewhere, so suffice to say that it is a simple equivalent circuit (Fig. HH model) that was designed to describe the propagation of action potentials in axons. It models sodium (Na+) and potassium (K+) ion channels explicitly as variable resistors, and predicts the shapes of spikes and axonal propagation very well. However, it is insufficient for describing the generation of spikes in the neighborhood of the soma, because the HH model is completely determinsitic.
Fig. HH model. V is the internal (membrane) potential.
I prefer drawing the HH circuit in a somewhat different way, using (non-linear) current sources instead of variable resistors or conductances (Fig. HH model with current sources). This is particularly convenient when we are interested only in the neuron’s operation in the vicinity of the resting potential (i.e., in the potential range where the neuron spends its time when it isn’t spiking).
Fig. HH model with current sources. Here, I have included synaptic input as an additional current source.
The HH model assumes that K channels and Na channels are uniformly distributed. Actually, the K channels are more prevalent in the proximal compartment, whereas Na channels are more prevalent in the AIS compartment. To transform the HH model into our three-compartment model, taking this heterogeneous distribution into account, we first stretch out the HH model as in Fig. Stretched-out. Note that this circuit is still electrically equivalent to the circuit in Fig. HH model with current sources.
Fig. Stretched-out HH model.
As the final step, we insert two axial resistors to create the three compartments (Fig. Three-compartment circuit). Although there is no big difference in structure (you can imagine the evolution having come up with this idea!), the behaviour of the circuit changes dramatically.
Fig. Three-compartment circuit (CC BY 4.0).
The circuit operates approximately as follows: Input (Isyn) to the neuron is lowpass filtered by the distal compartment. After a spike, the proximal compartment potential Vp starts up at a relatively low value. This potential increases slowly through a current via Rdp from the distal compartment, and the more input, the faster the charging of Cp. On top of this, there is a small but stochastically varying current (IK) exported by the K channels. If the total potential in the proximal compartment reaches a certain threshold voltage, this will be sensed by the Na channels in the AIS compartment (akin to an emitter follower!), and a spike is generated. The threshold has an interesting history, so I have written a digression on this topic, the Story of the Spiking Threshold.
The stochasticisty turns out to be essential, but it also makes the mathematics complicated. Luckily, we found that this circuit is still mathematically tractable, and behaves exactly in agreement with experimental data. For a detailed explanation of the nitty-gritty details, please see the paper in Phys Rev E.
Implications for current machine learning and AI techniques
Fig. Abstraction hierarchy of biological and biologically inspired learning.
The results in Physical Review E say concretely that the behaviour of neurons is considerably simpler than previously thought, and it enables us to find compact but accurate mathematical abstractions of neurons. However, there is a lot more to do. The model can potentially help us find higher-level abstractions for aggregates of neurons (Fig. Abstraction hierarchy), which may lead to new, biologically inspired methods for knowledge representation, machine learning, and AI. A problem with today’s machine learning methods is that they require enormous amounts of data, whereas the human brain achieves impressing results with considerably less data by instead using its structure.
Other activities
The following is a collection of random additional topics that I’m happy to share and that could perhaps save you some time and effort.
How To Make Collapsible Ski Poles
When you travel with roller skis, ski poles are rather awkward to transport. While the roller skis are short and fit into a handy bag, the ski poles are quite long (~1.5 m) and do not fit into the same bag. It would be nice to have collapsible ski poles that fit into the same bag as the roller skis!
Here is one way you can make collapsible ski poles.
Minimalistic Fitness Training
By minimalistic training I mean simple and concise training based on simple movements, low risk for injuries, and minimal investment in time and equipment. I think such minimalism is critical for the motivation to start and keep up training. Minimalistic training at home is especially useful when one’s ordinary gym, sports club, or running track is unavaliable, for example due to vacation, quarantine, or bad weather. While running fast is an efficient way to improve fitness, it can be hard to achieve at home. However, standing up cycling on a simple, foldable exercise bike can do wonders to the fitness.
Minimalistic Web Site Design
Designing web pages is fun, but maintaining them is a pain. This becomes particularly evident if one has to move the web site to a different ISP or change design tool with a proprietary format. Maintenance overhead is a good reason for keeping one’s web pages as simple as possible within reasonable limits. My suggestion is to use just simple Markdown with the Github style sheet. You can find more details here.
Using A Smartphone As A Desktop Computer
Smartphones are getting so powerful that they can be used as desktop computers. At the time of this writing (February 2021), there are not so many smartphones that satisfy the full requirements, but the trend is that this number will grow quickly. The main obstacle seems to be software support, but although it is developing quickly, there are some difficult software issues that cannot be solved easily.
StripCAD - A Circuit Editor For Stripboards
PC interface designed with StripCAD. The interface converts the RS232 status signals to two bidirectional digital I/O lines.
StripCAD is a free program for stripboard (a.k.a. “Veroboard”) circuit CAD running under Windows and Wine/Linux. StripCAD is intended to accelerate prototyping electronic circuits. The fundamental approach is (1) unify circuit schematics and circuit layout, and (2) to minimize soldering by mounting components vertically and using horizontal strips for interconnections. You can download it from here.