Romain Cazé

Neurons Perform Better Logical Calculations Than Expected

rcaze — Wed, 03 Dec 2025 09:05:51 +0000

Do artificial neurons underestimate the power of their biological counterparts? Ten years after the first theoretical predictions, researchers have shown that a single isolated neuron can carry out a calculation once thought impossible. Published in Scientific Reports, these findings suggest that employing more complex artificial neurons could improve the efficiency of neural networks.

If neural networks solve all kinds of computations, what can a solitary neuron do? Models that have been in place for more than half a century estimate that its capabilities remain limited, and current artificial intelligences are built on that assumption. Yet those systems are especially energy‑hungry and comparatively inefficient relative to what nature achieves, leading some scientists to suspect that a neuron actually has more tricks up its sleeve. Researchers from the Institute of Electronics, Microelectronics and Nanotechnology (IEMN, CNRS/University of Lille/Polytechnic University of Hauts‑de‑France), the Laboratory of Cognitive and Computational Neuroscience (LNC2, INSERM/ENS–PSL), and the Genes, Synapses and Cognition laboratory (GSC, CNRS/Institut Pasteur) have demonstrated that an isolated neuron is capable of far more logical calculations than commonly believed.

These belong to the class of linearly non‑separable operations, which classic models of isolated artificial neurons cannot solve. This limitation may be one reason for the poor energy efficiency of today’s networks. In the study, biological neurons solved on their own a problem that requires responding only when excitatory stimuli encoding shapes are paired with others encoding colors. To make the neurons compute, they were stimulated with an excitatory neurotransmitter, glutamate. The glutamate was trapped in “cages” and released by laser beams, allowing precise control over neuronal activation. The distinct locations where glutamate acted enabled the neuron to perform the desired operation. An electrode attached to the neuron confirmed that it fired precisely when the correct conditions were met, which it did reliably.

The IEMN team now aims to determine whether artificial neurons endowed with this capability would give rise to more performant artificial neural networks.Cette réponse me convient.

Microscopy image (two‑photon excitation) of a neuron used in the study.
The coloured dots on branches 1 and 2 indicate the sites that were excited by glutamate; the dots on branch 3 were activated by electrical stimulation. The scale bar represents 20 µm. The large white element corresponds to the pipette that records the neuron’s activity.

Références
Demonstration that sublinear dendrites enable linearly non‑separable computations.
Romain D. Cazé, Alexandra Tran‑Van‑Minh, Boris S. Gutkin & David A. DiGregorio.
Scientific Reports volume 14, Article number: 18226 (2024).
https://doi.org/10.1038/s41598-024-65866-9

In French: https://www.insis.cnrs.fr/fr/cnrsinfo/les-neurones-effectuent-de-meilleurs-calculs-logiques-que-prevu

Book chapter

rcaze — Mon, 11 Nov 2019 15:28:51 +0000

An post to point toward the book chapter I wrote at the end of my PhD.Caze2013Chap

Even passive dendrites extend neuron’s computation capacity

rcaze — Sat, 13 Oct 2018 16:14:28 +0000

A neuron receives its input through passive or active dendrites. Active dendrites can sum excitatory inputs both below and above their sum. This feature turns a neuron in a two-layer network capable of universal computation while linearly separable computations defined what a single neuron could do. Some neuron, however, lacks active dendrites. We thus wonder here what happens when dendrites are passive and process two excitatory inputs below their sum. We enumerate parameter sets and focus on excitatory inputs to determine how many computations can be implemented by a neuron model with either active or passive dendrites. We then generalize these numerical results to an arbitrary number of dendrites. First, we show that a single dendrite either passive or active suffices to compute linearly non-separable computations. Second, we prove that a sufficient number of passive dendrites enable a neuron to be universal for positive computations. Third, we show how a neuron can implement these computations using two distinct strategies: 1) Where a single dendrite suffices to trigger a somatic spike; 2) where somatic spiking requires the cooperation of multiple dendrites. Only a neuron with active dendrites can use the strategy (1) while neuron with either passive or active dendrites can use strategy (2). Finally, we employ strategy (2) to implement a linearly non-separable function in a biophysical model with passive dendrites inspired by stellate cerebellar interneurons. We show here that even passive dendrites enable a neuron to extend its computation capacity well beyond what we previously thought.

Even a passive dendrite extends neuron’s computation capacity

rcaze — Fri, 12 Oct 2018 17:44:42 +0000

A neuron possesses receptive organs called dendrites that are active or passive. Active dendrites can sum excitatory inputs both below and above their arithmetic sum. This feature turns a neuron in a two-layer neural network capable of universal computation. Linearly separable computations previously define what a single neuron could do. Some neuron, however, lacks active dendrites. We thus wonder here what happens when dendrites are passive and can only sum two excitatory inputs below their arithmetic sum. We enumerate parameter sets and focus on excitatory inputs to determine how many computations can be implemented by a neuron model with either an active or passive dendrite. We then analytically generalize these numerical results to an arbitrary number of dendrites. First, we show that a single dendrite either passive or active suffices to compute linearly non-separable computations. Second, we analytically prove that a sufficient number of passive dendrites enable a neuron to be universal for positive computations. Third, we show how a neuron can implement these computations using two distinct strategies: 1) Where a single dendrite suffices to trigger a somatic spike; 2) where somatic spiking requires the cooperation of multiple dendrites. Only a neuron with active dendrites can use the strategy (1) while neuron with either passive or active dendrites can use strategy (2). Finally, we employ strategy (2) to implement a linearly non-separable function in a biophysical model with passive dendrites inspired by stellate cerebellar interneurons. We show here that even passive dendrites enable a neuron to extend its computation capacity well beyond what we previously thought.

Rewriting abstracts

rcaze — Wed, 10 Oct 2018 09:31:09 +0000

Scientific articles follow a logic in their structure. They start by an abstract: a short text often below 300 words describing the article content. Next comes the introduction then result and methods. It ends with a discussion/conclusion. I would like to focus here on abstracts.

Abstracts constitute the beating heart of a paper. With the title, it will make or not people read the article. Yesterday I re-read the abstract of the paper I’m most proud of and I realize how poorly it was written.

I believe in second chances, even third or fourth. Writing is a painful and long exercise and there is no such thing as the perfect text from the first draft. In particular for abstracts. Every line, every word deserve to be carefully examined. I am going to rewrite the abstract first for some of my old papers but also of papers I enjoy reading. The latter being the most difficult thing to do because so of them are written by a native English speaker who writes much better than me. Yet I find the exercise useful not only for me but also to promote the paper.

A rewritten abstract offers a new perspective on a paper. I don’t speak about a lay abstract or a simplification like an author’s summary. I speak about a proper abstract with all the scientific rigour in it.

Thank you for reading! I hope to please you with my rewritten abstracts.

Computing using space in dendrites

rcaze — Fri, 16 Jun 2017 14:10:32 +0000

After two years of work my last paper is finally out 17_05NECO. What I keep realizing is that science can take an awful long time. New idea need time to mature and even more time to be accepted. This paper was an uphill battle against editorial and peer reviewers. But it is finally out and a good excuse to write a new post.

This paper demonstrate an important thing. Dendrites can play a crucial role even for the most simple computation. The computation studied here is stimulus selectivity in other words a simple passing of information to signal the presence or absence of a particular input. Even in this case dendrites enables to make this computation more resilient to synaptic and dendritic failure. The implementation proposed here may seem too complex for a simple computation: to make the preferred inputs the most dispersed rather than the strongest. But it shows that with dendrites you can do more with less. Be more resilient with less synapses.

Another coming work will show that for the same computation an implementation employing dendrites use less synapses than a classic one. This becomes particularly interesting when one see that saturating dendrites can be implement using a resistor (that will saturate in all cases).

A mountain becoming flat underwater.

rcaze — Tue, 06 Oct 2015 14:45:49 +0000

Some neurons respond preferentially to certain stimuli, like a sound or a picture reminding Jennifer Anniston. Hubel and Wiesel obtained the Nobel price, some 50 years ago, for the discovery of stimulus selective neurons in cats’ visual cortex and for the model associated with this discovery. This model shines by its simplicity. Imagine a mountain, each coordinates corresponds to a (visual) stimulus, the altitude associated to them corresponds to how strong a neuron responds to this stimulus. In more technical terms, the height of a given point equals the depolarization created by this stimulus. Now imagine that this mountain sits in the middle of a sea. The tip of the mountain outside of the water is the supra-threshold response, i.e. an activity level sufficient to trigger a neural activity noticeable by the rest of the brain. This metaphor seems a good way to understand neuron stimulus selectivity, but like all models it cannot explain everything.

Electrophysiology recently made astounding progresses and it is now possible to hyperpolarize a neuron in vivo. Using our metaphor it becomes feasible to increase the sea level. Scientists used this technique and made an unexpected observation. When they hyperpolarize a neuron -increase the sea level-, the mountain, as it goes underwater, becomes flat. Meaning that the neuron loses its selectivity and responds equally to all stimuli. Why? This is the topic of one of my current project. But I will talk more about it in another post, where I will try to explain why a neuron might start to respond as strong to Jennifer Anniston as to any other Hollywood inhabitants.

Indifference (in Science) can sometimes be frustrating.

rcaze — Fri, 25 Sep 2015 14:14:18 +0000

Motivated by my last post, I decided to more regularly update my blog.

In this post I am not asking a question but write about my life in Science and
my experience as a young researcher.

I enjoy a lot my life in Science, so I will try not to complain too much
(pledge), but life in science can sometimes be frustrating. I often
hear that the life of a young researcher can be difficult because of the
constant struggle. That he or she always has to fight against the existing
dogmas or ideas held by researchers higher up in the hierarchy. I tend to
disagree. Fighting against someone is motivating. You are never as
courageous as when you have a mighty opponent. Even a fierce and overwhelming
opponent that has much more means than you have. Worthwhile struggle is much
less frustrating than passive indifference. Science is just one area touched by
indifference and in our society overflowed by sounds, images and information
indifference sometimes is the only defense. So, I understand why people could
be indifferent. Still, I realize more and more that my frustration most often
comes from this indifference. Is there a solution? Well, it might be a clunky
solution but it is the only one I found. Indifference. I just say to myself
that if my work is worth something then someday, somewhere, somebody will use
it. Today I have a decent place to live, someone I love to live with, food in
my plate and people let me do what I love to do -Science. So no complain really.

A NEURON + Python tutorial

rcaze — Tue, 30 Jun 2015 07:51:40 +0000

I had the chance to give during the first week at OCNC (Okinawa Computational Neuroscience Course) a tutorial on NEURON neuron_tuto. This tutorial contains some self-advertisement ;), it demonstrates that a neuron with two passive dendrites can compute a linearly non-separable function, i.e. the feature binding problem. I hope it will be useful for your … Continue reading →

Neuron, human, brain and society.

rcaze — Thu, 26 Mar 2015 16:45:08 +0000

Yesterday I had dinner with some friends, and to explain my work I used an analogy that they liked. I used this analogy many times to justify my point of view. I am convinced that neurons are smarter than we thought, way smarter, and that it should change our views on the brain. I spent my PhD trying to define “smart” and to demonstrate this intelligence (see review below). I believe that our brain is capable of amazing things, e.g. being goal directed, because neurons are capable of these amazing things. In other words, that a neuron is as complex as a brain. It may seem a weird opinion, but I found an analogy/question making this proposition less strange. Do you think that a society is more complex than a human? I do not think so. I think that a human/neuron is as complex as a society/brain (even if there are complex surely in different ways).

Our discovery -presented in this review wrote with some experimental collaborators Tran-van-Minh et al._2015– demonstrates that sublinear summation in dendrites makes a neuron more intelligent than we thought. And I am convinced that it is one step toward my proposition about neurons and humans.