iiinova — where your digital self keeps evolving

Ido Aizenbud @idoai.bsky.social · 11/06/2026

TL;DR: A single cortical pyramidal neuron is not a point; it's a powerful, noise-robust, general-purpose computational unit. Dendrites are not just wires to help neurons connect; they are the substrate for complex nonlinear computation. (15/15) preprint: www.biorxiv.org/content/10.6...

020

Ido Aizenbud @idoai.bsky.social · 11/06/2026

If one biological neuron is this powerful, what happens when we build artificial networks from dendritic units instead of point neurons? TwinProp opens the door to bio-inspired architectures that do more with fewer units and less energy. More soon with @davidbeniaguev.bsky.social. (14/15)

110

Ido Aizenbud @idoai.bsky.social · 11/06/2026

If one neuron can solve tasks that require networks of point units, then point-neuron brain models miss a key source of computational power. TwinProp enables digital twins with full dendritic complexity; with connectomics, this opens a path to multi-scale brain twins. (13/15)

110

Ido Aizenbud @idoai.bsky.social · 11/06/2026

We reverse-engineered the XOR solution. The neuron splits the problem across 2 dendritic subunits: the apical tuft computes ¬P₀ ∧ P₁ and the basal tree computes P₀ ∧ ¬P₁. Their outputs converge at the soma that computes (¬P₀ ∧ P₁) ∨ (P₀ ∧ ¬P₁) = P₀ ⊕ P₁. (12/15)

100

Ido Aizenbud @idoai.bsky.social · 11/06/2026

To pinpoint the mechanisms, we performed causal ablations on 4-bit parity: Intact L5PC: 99.4% No active conductances: 78.1% Soma-only: 76.9% No NMDA: 73.8% LIF: 68.8% The computational power depends on dendritic structure, active conductances, and NMDA nonlinearities. (11/15)

110

Ido Aizenbud @idoai.bsky.social · 11/06/2026

NMDA receptors are key. As task complexity increases, NMDA currents grow monotonically, and spatial recruitment spreads across the dendritic arbor, especially into the apical tuft. (10/15)

100

Ido Aizenbud @idoai.bsky.social · 11/06/2026

What's the mechanism? Dendrites. PCA of dendritic voltages shows that harder tasks recruit higher-dimensional dynamics, more semi-independent dendritic branches doing local computations. It's a distributed, high-rank representation across the tree. (9/15)

131

Ido Aizenbud @idoai.bsky.social · 11/06/2026

A perceptron can't solve XOR, but our single neuron solves 10-dimensional parity and arbitrary random Boolean functions (~99% on 4-bit tasks), demonstrating genuine high-order feature binding, capabilities often attributed to multilayer networks. (8/15)

100

Ido Aizenbud @idoai.bsky.social · 11/06/2026

How far can it go? We tested parity: a classic hard benchmark for nonlinear computation, where one bit flip changes the answer. The L5PC solved XOR (d=2) perfectly, reached 99.4% on 4-bit parity, and 91.2% on 10-bit parity: 1,024 patterns requiring nonlinear integration. (7/15)

110

Ido Aizenbud @idoai.bsky.social · 11/06/2026

Can a single L5 pyramidal cell recognize spoken words? We encoded sounds into spike trains via a biologically plausible cochlea→VCN model, and: LIF: ~55% L5PC: ~71% 7-layer DNN upper bound: ~73% One neuron solves natural vision and speech tasks at near-DNN accuracy. (6/15)

110

Ido Aizenbud @idoai.bsky.social · 11/06/2026

Can a single L5 pyramidal cell classify images of cats and dogs? We encoded images into spike trains via a biologically plausible retina→LGN→V1→V2 model, then trained the neuron using TwinProp. Accuracy: LIF: ~55% L5PC: ~80% 7-layer DNN upper bound: ~83% (5/15)

110

Ido Aizenbud @idoai.bsky.social · 11/06/2026

Our solution: TwinProp, a digital-twin-based backprop algorithm that propagates gradients through a twin to optimize input weights. 1: Train a DNN on the neuron's I/O 2: Optimize synaptic weights via frozen DNN 3: Map optimized weights back to the detailed model & verify (4/15)

110

Ido Aizenbud @idoai.bsky.social · 11/06/2026

Until now, the field lacked a way to systematically ask what can a detailed neuron model compute. These models can't be solved analytically, and with thousands of synaptic weights to tune, manual search is intractable. How do we overcome these optimization challenges? (3/15)

100

Ido Aizenbud @idoai.bsky.social · 11/06/2026

For decades, neurons were modeled as point units in AI/neuroscience. But biological neurons have dendrites and nonlinear conductances. What capabilities are lost? See below! With @davidbeniaguev.bsky.social ,Noam Pnueli, Idan Segev & @mikilon.bsky.social: www.biorxiv.org/content/10.6.... (2/15)

120

Ido Aizenbud @idoai.bsky.social · 11/06/2026

What can a neuron compute? Real biological neurons are complex, but how capable are they? Using a new method, we found that a single cortical neuron can classify cats vs dogs, recognize spoken words, and solve 10-bit parity, all tasks thought to require entire networks. (1/15)

362

Reposted by Ido Aizenbud

Dan Goodman @neural-reckoning.org · 04/09/2025

Is anarchist science possible? As an experiment, we got together a large group of computational neuroscientists from around the world to work on a single project without top down direction. Read on to find out what happened. 🤖🧠🧪

Diagram of how the "collaborative modelling of the brain" (COMOB) project started. Starting material lead to group research or solo research, coming together in online workshops (monthly) in an iterative cycle, finishing with writing up together. The diagram is illustrated with colourful cartoon blob characters.

27528

Ido Aizenbud @idoai.bsky.social · 01/08/2025

Looking ahead, we reckon the next leap will come from combining high-res EM of entire human neurons with vol-imaging tools. Hybrid biophysical/AI models promise to clarify how single-cell properties scale up to network dynamics—and ultimately to circuits underlying language, creativity, and memory.

120

Ido Aizenbud @idoai.bsky.social · 01/08/2025

Human neurons are more functionally complex: their richer morphology and stronger synaptic nonlinearities give them extra computational power. Fitting deep neural nets to match human input-output dynamics consistently required greater network depth than for rodent models.

131

Ido Aizenbud @idoai.bsky.social · 01/08/2025

Single human neurons are wired to perform nontrivial logical computations! Due to extensive dendritic branching and specialized vol-gated currents, HL2/3 PNs support ~25 independent NMDA-spike compartments—almost twice than rat neurons—and can implement XOR-like operations via dendritic Ca²⁺ spikes.

163

Ido Aizenbud @idoai.bsky.social · 01/08/2025

Dendritic load imposed on human dendrites accelerates EPSP propagation down the dendrites, while dendritic high‐density of h-channels enables faithful transfer of theta-band signals (that are associated with various learning and memory processes) from dendrites to soma.

100

Ido Aizenbud @idoai.bsky.social · 01/08/2025

Load is all you need. The extensive membrane surface area of the dendrites “loads” the soma with additional capacitance and conductance. This load imposed on the AIS makes action potentials at the soma remarkably “kinky” – with a steeper rise of voltage, yet sensitive to rapid input fluctuations.

111

Ido Aizenbud @idoai.bsky.social · 01/08/2025

We bring together decades of human tissue recordings, detailed biophysical models, and machine-learning techniques to try and answer these questions. Here are some key insights:

100

Ido Aizenbud @idoai.bsky.social · 01/08/2025

What makes human pyramidal neurons uniquely suited for complex information processing? How can human neurons’ distinct properties contribute to our advanced cognitive abilities?

2379

Ido Aizenbud @idoai.bsky.social · 16/04/2025

Thanks! What kind of complexity measurements do you refer to?

000

Ido Aizenbud @idoai.bsky.social · 29/03/2025

Just got back from the GRC Dendrites meeting in Ventura, California! I presented my research on how single neurons can implement complex nonlinear functions — amazing discussions and brilliant minds all around. #Neuroscience #GRCdendrites #ComputationalNeuroscience

140

Reposted by Ido Aizenbud

Toviah Moldwin @tmoldwin.bsky.social · 19/02/2025

Now out in PLOS CB! We propose a simple, perceptron-like neuron model, the calcitron, that has four sources of [Ca2+]...We demonstrate that by modulating the plasticity thresholds and calcium influx from each calcium source, we can reproduce a wide range of learning and plasticity protocols.

journals.plos.org

The calcitron: A simple neuron model that implements many learning rules via the calcium control hypothesis

Author summary Researchers have developed various learning rules for artificial neural networks, but it is unclear how these rules relate to the brain’s natural processes. This study focuses on the ca...

1124

Ido Aizenbud @idoai.bsky.social · 08/01/2025

Great stuff! Can you add me as well?

110

Ido Aizenbud @idoai.bsky.social · 31/12/2024

Indeed, the human synapse has a strong effect on complexity, so even a rat morphology with human synapse will be much more complex than the same morphology with a rat synapse, as you can see in panels L and M, the effect is more pronounced within rat morphologies (meaning, smaller morphologies).

110

Ido Aizenbud @idoai.bsky.social · 31/12/2024

In this case, estimating the complexity of the neurons using the entropy of the weights is intersting, it is similar to checking whether there's a simpler DNN that will give the same perfect fit.

010

Ido Aizenbud @idoai.bsky.social · 31/12/2024

In principal, if you take a DNN that is expressive enough, it should perfectly fit the function of all neurons, so in this case, the FCI would be 0 for all of the neurons.

110

Ido Aizenbud @idoai.bsky.social · 29/12/2024

2. We have not tried to use Encoder-Decoder systems, but it is certainly possible and may be relevant.

100

Ido Aizenbud @idoai.bsky.social · 29/12/2024

1. We followed the architecture introduced in Beniaguev et al., 2021, and chose a depth of 3 that was enough to capture the variability in our model dataset. We also repeated some of the experiments with alternative depths of 2 and 7, and the results stayed similar (see discussion for more details).

110

Ido Aizenbud @idoai.bsky.social · 27/12/2024

Conceptually the same approach could be used to measure the functional complexity of biological neural networks, and actually of any model that we can feed with random inputs and simulate to get the outputs.

120

Ido Aizenbud @idoai.bsky.social · 27/12/2024

We argue this is a scalable approach to efficiently approximate the mutual information between the inputs and the outputs of a function (in this case, the function of a single neuron), therefore serving as a useful measure of functional complexity.

130

Ido Aizenbud @idoai.bsky.social · 27/12/2024

I'm not sure I understand your question, can you repeat it?

000

Ido Aizenbud @idoai.bsky.social · 27/12/2024

Size is indeed one of the important factors, but it is not the only one.

000

Ido Aizenbud @idoai.bsky.social · 26/12/2024

Check out the full story in our preprint. We welcome your feedback and hope this sparks new directions in exploring how single-neuron complexity scales up to network-level functions. (11/11) bioRxiv: www.biorxiv.org/content/10.1...

biorxiv.org

What makes human cortical pyramidal neurons functionally complex

Humans exhibit unique cognitive abilities within the animal kingdom, but the neural mechanisms driving these advanced capabilities remain poorly understood. Human cortical neurons differ from those of...

1161

Ido Aizenbud @idoai.bsky.social · 26/12/2024

Our findings introduce a quantitative framework linking single-neuron structure and biophysics to computational “depth.” As we refine our understanding of neuronal complexity, we move closer to explaining how human brains support advanced cognitive abilities. (10/11)

1110

Ido Aizenbud @idoai.bsky.social · 26/12/2024

Interestingly, the pattern also depends on the cortical layer. In rats, L5 pyramidal cells are more complex than L2/3. In humans, it’s reversed: L2/3 cells show greater complexity, hinting at a human-specific shift in cortical computation. (9/11)

2151

Ido Aizenbud @idoai.bsky.social · 26/12/2024

By disentangling morphological and synaptic contributions, we find that human neurons owe their enhanced complexity to distinct dendritic architecture and unique synaptic properties. These combined factors set human neurons fundamentally apart from rat neurons in terms of I/O complexity. (8/11)

1110

Ido Aizenbud @idoai.bsky.social · 26/12/2024

It’s not just size and shape. Synaptic properties, especially those governing NMDA receptor dynamics - play a critical role. Human neurons exhibit steeper NMDA nonlinearities, elevating their computational complexity. (7/11)

3192

Ido Aizenbud @idoai.bsky.social · 26/12/2024

Structural features matter. larger dendritic trees, more branches, and complex morphologies correlate strongly with higher complexity. Human neurons aren’t just big - they’re wired to integrate inputs in more sophisticated ways. (6/11)

1234

Ido Aizenbud @idoai.bsky.social · 26/12/2024

Applying the FCI, we find that human cortical pyramidal neurons consistently rank higher in I/O complexity than rat neurons. This suggests that human neurons can perform more intricate computations at the single-cell level. (5/11)

1185

Ido Aizenbud @idoai.bsky.social · 26/12/2024

Our new metric, the FCI (Functional Complexity Index), measures a neuron’s I/O complexity based on how well (or poorly) a fixed deep network can replicate its output. A good fit = simpler I/O, while a poor fit = more complex I/O. (4/11)

1121

Ido Aizenbud @idoai.bsky.social · 26/12/2024

In our new preprint, we examined 24 reconstructed human and rat neurons from all cortical layers, with experimentally matched electrophysiology. After simulating responses to diverse inputs, we fit a DNN to each neuron’s I/O - quantifying how “hard” it is to predict their output. (3/11)

1131