A Bayes-Markov Neuromorphic Model of Cortical Orientation Selectivity: A Computational Re-implementation and Quantitative Simulation Study

arXiv:2608.12388 · q-bio.NC, cs.LG, cs.NE · Submitted 2026-08-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "A Bayes-Markov Neuromorphic Model of Cortical Orientation Selectivity: A Computational Re-implementation and Quantitative Simulation Study".

Jane: The paper was written by Abolfazl Moslemi, Milad Sarabadani, Fatemeh Sefidian and Hossein Peyvandi from Sharif University of Technology and University of Tehran and Islamic Azad University and Pasargad Institute for Advanced Innovative Solutions.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title and Authors: Tom: Welcome back to the show, everyone. Today we're looking at a paper that's been making the rounds on arXiv, and it's a mouthful: "A Bayes–Markov Neuromorphic Model of Cortical Orientation Selectivity: A Computational Re-implementation and Quantitative Simulation Study." Jane, I have to say, just reading that title makes me feel like I need a neuroscience degree.

Jane: Tom, it sounds intimidating, but the core question is actually something we can all relate to. How does your brain know that a line is horizontal versus vertical? That's orientation selectivity, and this paper is trying to explain how that happens in the visual cortex.

Tom: Right, so we're talking about the very first stages of vision processing. The authors—Moslemi, Sarabadani, Sefidian, and Peyvandi—they're revisiting an older model from two thousand four by a researcher named Shirazi. They're not inventing a new theory from scratch; they're taking that old idea and giving it a proper, modern computational test.

Jane: And that's actually a really important thing to do. The original model was mostly theoretical. It proposed that orientation selectivity could emerge from a process of Bayesian inference, where the brain is essentially guessing the most likely pattern of cortical activity given the raw input it gets from the eyes.

Tom: So instead of the brain having hardwired detectors for every possible angle, it's running a probability calculation on the fly?

Jane: Exactly. And the original idea was clever, but it was hard to verify. This new paper takes that framework and actually builds it, runs simulations, and measures how well it performs. They're asking, does this theory actually hold up when you put it on a computer?

Tom: And does it? I mean, we have the paper here, so spoiler alert, it seems to. But what's the big deal about re-implementing an old model? Why not just come up with a new one?

Jane: Because if the old idea works, that's a huge validation. It means the original insight was correct. And this team went further. They added spiking neuron models, which are much more biologically realistic than the simple on-off switches used in the original. So they're testing the theory with more realistic hardware, so to speak.

Tom: I love that. It's like taking a classic car design and putting a modern engine in it to see if it can actually win the race. And it sounds like it did win. But I'm curious about the practical side of this. Who cares if a computer model can detect lines?

Jane: Well, Tom, that's exactly what we're going to get into. Because if we can figure out how the brain does this so efficiently, we might be able to build computer vision systems that work the same way. But that's a conversation for a bit later. For now, let's just say this paper is a big deal for anyone who cares about how brains work, and that includes people building artificial ones.

Tom: Alright, I'm hooked. So we've got a classic theory, a modern test, and a potential payoff for AI. Let's keep going and see what the actual results were.

Paper Summary: Jane: So, Tom, we've established that this paper, "A Bayes–Markov Neuromorphic Model of Cortical Orientation Selectivity," is a modern test of a two thousand four theory. But let's get into the actual meat of it. What did they find?

Tom: They found that the model works. It produces sharp orientation selectivity, meaning it can clearly distinguish a horizontal bar from a vertical one. They even got a perfect score of one point zero on their orientation selectivity index, which is their measure of how well the model can tell orientations apart.

Jane: And that's not just a lucky guess. They tested it with a bar at different angles, from zero to one hundred eighty degrees, and the model's response peaked sharply at the preferred orientation and was completely suppressed at the orthogonal one. It's a clean, textbook tuning curve.

Tom: But it's not just about the perfect case. They also threw noise at it. They added random perturbations to the input, simulating a noisy visual signal, and the model held up. It was robust to moderate noise, which is a big deal because the real world is messy.

Jane: And that's the key point. The model isn't just a toy that works in a perfect simulation. It's robust. They also tested contrast invariance, which is a fancy way of saying the model's selectivity doesn't change when you make the stimulus brighter or dimmer. That's a property that real visual cortex neurons have, and it's been a challenge for many models to replicate.

Tom: So the model is sharp, robust, and contrast-invariant. That sounds like a slam dunk. But what about the spiking part? I know you mentioned that earlier. The paper's title mentions "neuromorphic," which sounds very cool, but what does it actually mean here?

Jane: It means they took the abstract, rate-based model and implemented it with spiking neurons. Instead of just saying a neuron is "on" or "off," they simulated actual action potentials, the little electrical spikes that neurons use to communicate. They used two models: a simpler leaky integrate-and-fire model and the more complex, biophysically detailed Hodgkin-Huxley model.

Tom: And did the spiking version still work? Because I imagine that's a much harder thing to get right.

Jane: It did. The spiking neurons preserved the orientation selectivity. The Hodgkin-Huxley model, in particular, showed a really interesting property. It was all-or-none. When it was active, it fired at a very consistent rate, and when it wasn't, it was silent. There was no in-between.

Tom: That's fascinating. It's like the model is making a binary decision, but it's doing it through the natural dynamics of the neuron, not through some artificial rule. That's a really elegant result.

Jane: It is. And it suggests that the brain might be using this kind of inference mechanism, and the spiking dynamics are just how the computation gets implemented in real tissue.

Tom: So we have a theory, we have a simulation, and we have a biologically plausible implementation. What's the catch? What's the next step? There's got to be a catch.

Jane: The catch is that this is still a model. It's a proof of concept. The next step is to see if this can be scaled up and applied to more complex visual scenes, not just simple bars. But that's a conversation for our next segment, where we talk about the improvements this paper suggests.

Improvements and Implications: Tom: Welcome back. We're deep in the weeds of "A Bayes–Markov Neuromorphic Model of Cortical Orientation Selectivity," and Jane just teased that we're going to talk about what this means for the future. So, what's the big improvement here?

Jane: The big improvement is that this paper takes a theory that was mostly conceptual and makes it testable and practical. The original model had some operations that were hard to imagine a real biological circuit performing. This re-implementation cleans that up, making it more biologically plausible.

Tom: And they didn't just clean it up; they made it run fast. They vectorized the code, which is a technical way of saying they made the simulation run much faster on a computer. That means you can do systematic parameter sweeps, testing lots of different settings, which is crucial for understanding the model's behavior.

Jane: Right. And that speed is what allowed them to do those quantitative tests we talked about, like the noise robustness and the contrast invariance. You can't do that kind of systematic analysis with a slow, pixel-by-pixel simulation.

Tom: So, for the researchers, it's a better tool. But what about the bigger picture? I know we have Lu and Meng on the line. Lu, you're the AI visionary. What does this mean for artificial intelligence?

Lu: Tom, this is a fantastic question. For me, the most exciting part is the neuromorphic angle. The fact that they could implement this Bayesian inference with Hodgkin-Huxley neurons is a blueprint for building more efficient AI hardware. Current AI, like deep learning, is incredibly power-hungry. The brain does this kind of computation with a fraction of the energy.

Tom: So you're saying this could lead to brain-inspired chips that are way more efficient than our current GPUs?

Lu: Potentially, yes. If we can figure out how to build circuits that mimic this kind of local, probabilistic inference, we could see a new generation of hardware that's much better at processing visual information in real-time, like for robotics or autonomous vehicles.

Meng: But Lu, from an engineering standpoint, that's a long way off. We can't just take this simulation and turn it into a chip tomorrow. What's the more immediate practical impact?

Jane: That's a fair point, Meng. The immediate impact is in our understanding of the brain. This paper provides a concrete, testable model of a fundamental brain function. It gives neuroscientists a framework for designing new experiments to look for these specific patterns of activity in the visual cortex.

Meng: So it's more of a scientific tool than an engineering one right now. That makes sense. But it does validate the approach. It shows that Bayesian inference is a computationally viable way to explain how the brain works, which is a big deal for the whole field.

Tom: So it's a win for neuroscience and a potential roadmap for AI. That's a pretty good day's work. But I have to ask, what about the limitations? What can't this model do?

Jane: Well, the model is tested on simple synthetic stimuli, like bars. The real world is full of complex textures and natural scenes. The model's neighborhood is also quite small, which limits the range of orientations it can detect. It's great at horizontal, vertical, and diagonal, but finer angles might be a challenge.

Lu: And that's where the future work lies. Scaling this up, testing it on natural images, and seeing if the principles hold. But the foundation is solid. This paper gives us a strong, quantitative proof that the brain could be doing this.

Tom: Alright, so we have a validated theory, a potential path to efficient AI, and a clear list of future challenges. Let's wrap this up in our final segment.

Conclusion: Tom: And we're back for the final stretch. We've been discussing "A Bayes–Markov Neuromorphic Model of Cortical Orientation Selectivity: A Computational Re-implementation and Quantitative Simulation Study," and it's been quite a journey.

Jane: It really has. We started with a question about how the brain detects simple lines, and we ended up talking about Bayesian inference, spiking neurons, and the future of AI. This paper took a two thousand four theory and gave it a modern, quantitative validation.

Tom: And the key results were impressive. The model achieved perfect orientation selectivity, was robust to noise, and showed contrast invariance. And the spiking implementation, especially the Hodgkin-Huxley version, showed that this computation can be done with biologically realistic all-or-none firing.

Jane: Exactly. It's a proof of concept that the brain could be running this kind of local probabilistic inference. It's not just a mathematical abstraction; it's a plausible biological mechanism.

Tom: And for the engineers and AI folks out there, it's a potential roadmap for more efficient, brain-inspired computing. The idea of implementing Bayesian inference with neuromorphic hardware is incredibly exciting.

Jane: It is. But as we discussed, there's still work to be done. The model needs to be tested on more complex stimuli, and the hardware implementation is still a distant dream. But this paper provides a solid foundation to build on.

Tom: Well said. So, we're going to say goodbye to this paper. It was a great read, and it gave us a lot to think about. A big thank you to the authors for their rigorous work.

Jane: Absolutely. And to our listeners, thanks for joining us. We'll be back soon with another paper from the arXiv. Until then, keep your eyes open and your neurons firing.

Tom: See you next time.

Abolfazl Moslemi, Milad Sarabadani, Fatemeh Sefidian, Hossein Peyvandi

Sharif University of Technology · University of Tehran · Islamic Azad University · Pasargad Institute for Advanced Innovative Solutions

q-bio.NC, cs.LG, cs.NE

Submitted: 2026-08-01

Comments: 24 pages. Computational neuroscience; orientation selectivity; neuromorphic modeling

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 46/100

Key concepts

Orientation Selectivity
This refers to the brain's ability to distinguish between different orientations of lines, such as horizontal versus vertical. The paper tests a model's success in performing this task in the visual cortex.
Bayes-Markov Model
This is a theoretical framework where the brain uses Bayesian inference—essentially making probabilistic guesses about the most likely pattern of cortical activity based on raw visual input. The model was re-implemented computationally to test this idea.
Neuromorphic Model
This refers to implementing the theory using spiking neurons instead of simple on-off switches. This biological realism, tested with models like Hodgkin-Huxley, showed that the computation can occur through biologically realistic electrical spikes.
Contrast Invariance
This is a property where a model's selectivity does not change when the visual stimulus is made brighter or dimmer. The paper confirmed this property, which is a feature found in real visual cortex neurons.

Terminology

Summary

Summary

This paper revisits Shirazi’s Bayes–Markov model of cortical orientation selectivity and provides a computational re-implementation with quantitative simulation validation. The original theory formulates the emergence of orientation-selective inhibition as a MAP estimation problem over a two-layer hierarchical Markov random field, solved by a local parallel relaxation algorithm. We preserved this formulation and focused on making its behavior measurable through systematic simulations.

The results support the main qualitative predictions of the original model. The reimplemented model produces sharp orientation tuning, achieves OSI = 1.0 in the biologically motivated implementation, and remains robust to moderate LGN noise. The vectorized implementation makes these simulations computationally feasible without changing the underlying local clique operations. Finally, the spiking SCI-layer implementation shows that the inferred orientation-selective field can be expressed through temporally explicit neural dynamics. The HH model provides a biologically interpretable proof-of-concept all-or-none spiking realization that is consistent with Shirazi’s proposed temporal-coding interpretation.

The main contributions of this work are as follows: (1) We reconstruct the Bayes–Markov formulation of Shirazi’s model, including the two-layer HMRF structure, the MAP objective, and the local GICM-style relaxation rule. (2) We describe the fully LGN-driven version of the model, emphasizing how the original formulation avoids biologically implausible global normalization and fixed reference signals. (3) We provide a vectorized computational implementation that preserves the original local clique-energy operations while making systematic simulation studies feasible. (4) We add a quantitative layer to the model’s previously qualitative evaluation: orientation tuning curves with an orientation selectivity index (OSI), contrast invariance, and robustness to LGN noise across all model variants. (5) We extend the simulation study with a spiking SCI-layer implementation using LIF and Hodgkin–Huxley neurons, showing that the inferred orientation-selective field can be expressed through temporally explicit spike trains.

The simulation study is organized into three principal experiments. The first two revisit the orientation-selectivity and contrast-invariance behavior of the Bayes–Markov model, pairing the original qualitative activity-pattern demonstrations with a quantitative analysis based on the orientation selectivity index and controlled noise. The third introduces a spiking realization of the SCI layer and evaluates it through a battery of experiments spanning baseline selectivity, per-neuron heterogeneity, multi-trial firing-rate stability, and single-neuron gain analysis.

In Simulation 1, a bar stimulus was presented at a range of orientations and propagated through the LGN preprocessing stage and the SCI inference. The SCI layer responds strongly when the stimulus matches the model’s preferred (horizontally tuned) orientation and is suppressed otherwise. The resulting tuning curve peaks sharply at the preferred orientation and is strongly suppressed at the orthogonal orientation, yielding OSI = 1.0. The small secondary responses at oblique orientations are expected from the 3 × 3 neighborhood, which naturally represents the principal orientations 0◦, 45◦, 90◦, and 135◦.

In Simulation 2, contrast invariance was tested qualitatively by presenting a bar of fixed orientation at several contrast levels; the inferred SCI pattern is essentially unchanged across the tested contrast values, supporting the interpretation that the local similarity-based inference depends primarily on the relative structure of the LGN input rather than its absolute magnitude. Robustness was then quantified by adding zero-mean Gaussian noise of standard deviation σ ∈ 0, 0.05, 0.1, 0.2, 0.3, 0.5 to the LGN activation field and measuring the mean response inside the stimulated bar region, repeated for bars at four orientations: 0◦ (preferred), 10◦, 45◦, and 90◦ (orthogonal). At the preferred orientation, the response is preserved for small perturbations: V1ModelBio remains at full activation (+1.0) up to σ ≈ 0.05, falls to ≈ +0.56 at σ = 0.1, and is actively suppressed (negative activation) for σ ≥ 0.2. The spiking models track this behavior in firing-rate terms, declining from ≈ 49 Hz (LIF) and ≈ 67 Hz (HH) at low noise to a few spikes per second beyond σ ≈ 0.1. This identifies a practical noise regime, near σ ≈ 0.1, within which the local inference remains stable. At 90◦ the model is fully and stably suppressed: V1ModelBio stays at −1.0 and both spiking models remain silent (≈ 0 Hz) at every noise level. At 10◦ the response is likewise suppressed throughout (≈ −0.8 rate-coded activation, < 10 Hz spiking). At 45◦ a small residual response is present at zero noise (V1ModelBio ≈ −0.5, HH ≈ 19 Hz), consistent with the weak secondary oblique response expected from the 3 × 3 neighborhood; this residual collapses by σ ≈ 0.1.

In the cross-model comparison, under a shared parameter setting (β = 1.0, γ = 2.0, T = 0.15), V1ModelBio produced the clearest orientation-selective output (OSI = 1.0), whereas the baseline V1Model was more broadly tuned (OSI = 0.33). Both spiking realizations remained sharply selective under the same drive, with OSI = 1.0 for the LIF model and OSI = 0.998 for the HH model.

In Simulation 3, the spiking realization does not change the Bayes–Markov inference step; it is a downstream temporal readout of the inferred MAP field rather than a replacement of the inference algorithm. First, the rate-coded MAP procedure produces an orientation-selective SCI field sMAP. This field is then converted into an input current, ILGN = Ioffset + Iscale sMAP, which drives a grid of spiking neurons. Two spiking neuron models were implemented: a leaky integrate-and-fire (LIF) model and a Hodgkin–Huxley (HH) model. Both spiking implementations preserved the sharp orientation selectivity of the rate-coded MAP field. Driven by the preferred (0◦) stimulus, the LIF model fired at approximately 49 Hz and the HH model at approximately 67 Hz, while both were silent at the orthogonal (90◦) orientation, yielding OSI = 1.0 for the LIF model and OSI = 0.999 for the HH model.

In Experiment 2, per-neuron heterogeneity was added by multiplying each neuron’s MAP-derived input current by a static random gain drawn from U(0.85, 1.15), and the measurement was repeated over ten independent stochastic-noise realizations. The LIF model remains orientation selective but grades under heterogeneity: its pooled CV rises from 0.061 (baseline) to 0.101 (heterogeneous), and the per-trial CV is tight (0.061 ± 0.001 and 0.101 ± 0.001). The HH model instead remains all-or-none: its pooled CV is 0.000 at baseline and only 0.005 under heterogeneity, with per-trial CV 0.004 ± 0.003. In the HH implementation active cells therefore fire at a characteristic rate or remain silent, and this bistability persists across independent noise realizations.

In Experiment 3, the contrast between the two neuron models is explained by their single-neuron frequency–current (f –I) curves. The LIF neuron has a smooth, continuously rising f –I curve: over the heterogeneity drive range (0.935–1.265 in dimensionless input units) its firing rate spans the full interval 0–35 Hz, a spread of 35 Hz. The HH neuron instead has an abrupt spike-initiation threshold followed by a near-plateau firing regime: over the corresponding heterogeneity drive range (6.8–9.2 µA/cm2) its firing rate varies only between 60 and 66.7 Hz, a spread of 6.7 Hz. The all-or-none behavior of the HH population is thus a direct consequence of its voltage-gated channel dynamics, supporting the behavior described in Prediction 1.

The paper also includes a lemma proving that the SCI clique energy can be represented locally through quenching and activating cells, avoiding the need for direct access to abstract SCI states. Several limitations are noted: the current experiments focus on controlled synthetic stimuli rather than large-scale natural image benchmarks, and the 3×3 neighborhood naturally favors a small set of principal orientations, with finer orientation resolution requiring larger neighborhoods or more detailed clique structures.

Improvements for AI systems

Based on this paper, I can identify several specific improvements to AI systems, particularly in computer vision, probabilistic inference, and neuromorphic computing.

Improvement: Implement a two-layer Markov random field (MRF) framework where the hidden layer (analogous to SCI cells) is inferred from an observed layer (analogous to LGN) using maximum a posteriori (MAP) estimation. This replaces simple feedforward feature detectors with a probabilistic inference mechanism that explicitly models uncertainty.

What the improved system can do:

  • Extract orientation-selective features from non-oriented raw inputs without requiring labeled training data or hand-crafted filter banks

  • Bootstrap complex feature selectivity (e.g., orientation) from simpler, non-selective inputs (e.g., contrast) through local cooperative inference

  • Maintain feature selectivity under input noise perturbations up to σ ≈ 0.1, providing robustness that standard convolutional filters lack

Improvement: Replace global optimization (e.g., exhaustive search, global normalization) with a fully parallel, local iterative update rule (Algorithm 1) that uses only neighborhood information. The algorithm uses a two-phase approach: maximum-likelihood initialization followed by iterative refinement with increasing coupling strength.

Improvement: Replace log-sum-exp normalization (which requires knowledge of all possible states) with a local maximum operation. This avoids biologically implausible global normalization and makes the computation fully local.

Improvement: Use local similarity/dissimilarity operators (Eq. 13–14) with a threshold T rather than absolute magnitude comparisons. This makes feature detection dependent on relative structure rather than absolute input intensity.

Improvement: Convert the rate-coded binary output into a spiking realization using Hodgkin–Huxley (HH) neurons. The HH model produces all-or-none firing (pooled CV ≤ 0.005 across 10 trials) compared to graded responses in LIF models (pooled CV = 0.101).

Improvement: Adopt the OSI metric (Eq. 17) as a standardized, quantitative measure of feature selectivity, complementing qualitative visual inspection.

Improvement: Implement a systematic evaluation protocol that includes: (a) controlled noise sweeps (σ ∈ 0, 0.05, 0.1, 0.2, 0.3, 0.5), (b) per-neuron heterogeneity (gain variation ±15%), and (c) multi-trial coefficient of variation (CV) analysis.

Improvement: Use shifted-array operations to vectorize the local clique-energy computations without changing the mathematical model. This was verified to produce zero mismatched pixels compared to pixel-wise implementation.

Summary of Capabilities: The improved AI system would be a probabilistic, locally-computed, contrast-invariant feature extractor that can bootstrap complex selectivity from simple inputs, operate efficiently in parallel, maintain robustness under noise and heterogeneity, and produce stable all-or-none outputs suitable for neuromorphic hardware—all while providing quantitative, reproducible evaluation metrics.

Related papers