1M-neuron model of macaque visual cortex achieving 85% accuracy in decoding orientation from noisy neural population activity — demonstrating robust biological computation without supervised learning.
Spatiotemporal spiking dynamics across 25mm² (16mm² depicted) cortical patch. Top: Population activity waves | Middle: Thalamic input | Bottom: Individual spike trains
Built a biologically-constrained 1M-neuron spiking neural network to understand how the brain processes visual information. The model processes visual stimuli (gratings, bars, dashed bars) through simulated retina → thalamus → cortex pathway and exhibits emergent spatiotemporal dynamics.
Core Question: How do large neural populations reliably decode information (visual orientation) from noisy, spatiotemporal signals?
Answer: Through biologically-plausible literature backed connectivity patterns alone — no gradient descent or supervised training needed.
Task: Decode each neuron's preferred visual orientation from its spiking responses to moving gratings.
Method: Presented gratings at 6 different orientations (0°, 30°, 60°, 90°, 120°, 150°) → measured each neuron's spiking response → assigned the orientation it responded to most strongly.
Result: 85% match with ground truth orientation map.
85% of neurons correctly decoded — Colors represent preferred orientation (red=0°, green=60°, blue=120°, etc.). Clear iso-orientation domains (patches of same color) show the network reliably extracts orientation information from visual signals. 6 populations shown: L4ABE (excitatory layer 4A/B), L4ABI (inhibitory 4A/B), L4CE_ON/OFF (excitatory 4C), L4CI_ON/OFF (inhibitory 4C).
Challenge: Can the network decode orientation from heavily degraded signals?
Noisy thalamic input (Signal to noise ratio 0.5) — a much harder decoding challenge
Orientation maps still retrieved despite extremely heavy noise — demonstrating robust population-level coding
- Robust decoding: Network maintains orientation selectivity at signal to noise ratios of 0.5
- Population coding: Individual neurons can be noisy, but population responses are reliable
- No supervised learning: Orientation selectivity emerges purely from biological connectivity rules
- Biological validity: Matches experimental recordings from macaque V1
- BCI relevance: Same principles apply to decoding information from noisy EEG/neural signals
Key innovation: Achieved through Hebbian-inspired connectivity patterns — no backpropagation or gradient descent.
- 1,000,000 neurons across 4 cortical layers (L4A/B, L4C ON/OFF pathways)
- 1,000,000,000+ synapses with biologically realistic connectivity
- 20,000,000,000+ datapoints processed (0.1ms resolution, 20s simulations)
- 25mm² (16mm² depicted) cortical area (central visual field representation)
- Leaky Integrate-and-Fire neurons (conductance-based, biologically realistic)
- Distance-dependent connectivity (exponentially decaying spatial profiles)
- Orientation-selective patterns (patchy long-range + push-pull short-range)
- Two types of inhibitory neurons (super localized + elliptical)
- Realistic synaptic delays (distance + conduction velocity)
- Retina preprocessing: Center-surround receptive fields (Mexican hat filters)
- ON/OFF pathways: Mimicking biological thalamic ON and OFF cells
- Gabor-based projections: Thalamocortical connections creating orientation selectivity
- Visual stimuli: Moving gratings, bars, dashed patterns and noisy dashed bars
No gradient descent — network topology inspired through biologically-plausible Hebbian-like connectivity rules, demonstrating that biological constraints alone produce robust computation.
This work directly applies to:
Brain-Computer Interfaces (BCIs):
- Understanding population-level neural coding for visual decoding
- Principles for decoding from noisy EEG/neural signals
- Robust signal processing under high noise
Neuromorphic Computing:
- Biologically-inspired architectures for edge AI
- Event-driven spiking computation
- Energy-efficient neural processing
Computational Neuroscience:
- Testing theories of cortical processing
- Understanding orientation selectivity emergence
- Population coding mechanisms
Medical Applications:
- Understanding visual processing disorders
- Neural prosthetics design
- Diagnostic tools for visual system
Why it matters: Real brains work with noisy, sparse, asynchronous spikes. This model shows how robust computation emerges from biological constraints — principles applicable to EEG decoding, neuromorphic hardware, and explainable AI.
- Simulator: NEST Neural Simulation Tool 3.x
- Language: Python 3.8+
- HPC: SLURM job scheduling on compute clusters
- Analysis: NumPy, SciPy, Matplotlib
- Data Scale: ~500GB per full simulation run
- Compute: Multi-node HPC (128+ cores typical)
Title: "Stabilization of the Orientation Map in a Computational Model of L4 in V1 of Macaque Monkey"
📄 Read Full Thesis (PDF) 📄 Final Presentation (PDF)
Abstract: Visual cortex layer 4 exhibits orientation selectivity—neurons respond preferentially to edges of specific angles. This thesis investigates the connectivity patterns that stabilize this "orientation map" under noisy, dynamic input, using a large-scale spiking neural network model constrained by experimental neuroscience data.
Key Contributions:
- 85% orientation decoding accuracy from neural population activity
- Demonstrated noise resistance (50%+ noise tolerance)
- Identified connectivity patterns (patchy, push-pull) stabilizing orientation maps
- Validated against experimental V1 recordings
- Showed emergent spatiotemporal dynamics from local connectivity
Institution: Forschungszentrum Jülich (IAS-6 Computational and Systems Neuroscience) & RWTH Aachen University
Grade: 1.0 (Best) | Period: Nov 2022 – Mar 2024
Silas Theinen
Computational Neuroscientist | Neural Signal Processing
Interested in computational neuroscience, BCIs, or large-scale neural simulations? Let's connect!
If you find this work useful for your research:
@mastersthesis{theinen2024visual,
title={Stabilization of the Orientation Map in a Computational Model of L4 in V1 of Macaque Monkey},
author={Theinen, Silas},
year={2024},
school={RWTH Aachen University and Forschungszentrum J\"ulich},
note={85\% orientation decoding accuracy from 1M-neuron spiking network}
}This work was conducted at the Institute for Advanced Simulation (IAS-6) at Forschungszentrum Jülich, using their HPC infrastructure. Special thanks to the NEST development team for their excellent neural simulation toolkit.
- ✅ 85% accuracy in orientation decoding from neural population activity
- ✅ Robust to 50%+ noise — maintains performance under high noise
- ✅ No supervised learning — emerges from biological connectivity alone
- ✅ 1M neurons, 1B synapses, 20B datapoints — large-scale realistic simulation
- ✅ Biologically validated — matches experimental V1 recordings
- ✅ Applicable to BCIs — principles for decoding noisy neural signals
Note: Due to computational scale (1M+ neurons, 1B+ synapses), full simulation code requires HPC resources. Contact me for implementation details or collaboration opportunities.