MPKNet V6

Bio-inspired vision that runs on a Raspberry Pi

Benchmarks • Architecture • Quick Start • License

76KB model • 33 FPS on RPi5 • 89% Kvasir-v2 • No pretraining • No augmentation • 161× fewer params than MobileNetV3

---

MPKNet V6 implements the parallel visual pathways found in mammalian brains. Instead of stacking more layers, it asks: what if how information flows matters more than parameter count?

	MobileNetV3-S	MPKNet V6	MPKNet V6-Pi
Parameters	2.5M	0.21M	15.5K
Model size	10MB	0.89MB	76KB
FPS (RPi5)	5-8	—	33
Accuracy (Kvasir)	~92%*	89%	82%

MobileNetV3-S accuracy from published benchmarks (not my evaluation). V6/V6-Pi measured on Kvasir-v2 val set (1600 samples). Direct comparison requires same-dataset evaluation with identical training protocol.

161× fewer parameters than MobileNetV3-S. Train in an hour, not a week. Deploy on a $35 Raspberry Pi, not a cloud GPU.

V6 is not about beating SOTA. It's about competitive accuracy at a fraction of the cost.

---

Benchmarks

Image Classification

Dataset	Classes	Resolution	Accuracy	Params	Notes
Kvasir-v2	8	224×224	89%	0.21M	Medical endoscopy (research only)
TinyImageNet	200	64×64	40.6%	0.21M	ResNet18 gets ~41.5% with 52× more params
CIFAR-100	100	32×32	58.8%	0.22M
STL-10	10	96×96	71.7%	0.21M	Only 5K training samples
ImageNet-100	100	224×224	60.8%	0.54M

Video Classification (V6.2 Temporal)

Dataset	Classes	Resolution	Accuracy	Params	Notes
UCF-101	101	112×112	77%	0.58M	8-frame temporal M-pathway, from scratch

V6.2 adds a sequential temporal M-pathway — the M stream processes 8 consecutive frames and computes inter-frame deltas for motion detection, while P sees only the current frame for spatial detail. K gates both. This mirrors the biological role of magnocellular neurons in motion processing.

Edge Deployment

Device	Model	Size	FPS	Accuracy
Raspberry Pi 5 (no heatsink)	V6-Pi	76KB	33	82%
MacBook M3	V6	0.89MB	200+	89%

Finding: Augmentation Hurts at Small Scales

Dataset	No Augmentation	With Augmentation	Change
CIFAR-100 (32×32)	52.8%	46.0%	-6.8%
TinyImageNet (64×64)	40.6%	24.1%	-16.5%
ImageNet-100 (224×224)	60.8%	~62%	+1-2%

This is consistent with NetAug (Cai et al., 2022), which showed regularization hurts tiny models that underfit rather than overfit. At small resolutions, the Fibonacci stride architecture provides sufficient multi-scale coverage that augmentation becomes redundant noise. At 224×224, mild augmentation helps marginally.

---

Quick Start

import torch
from MPKx import MPKNetV6

# Create model
model = MPKNetV6(num_classes=8)  # e.g., Kvasir-v2
print(f"Parameters: {sum(p.numel() for p in model.parameters()):,}")
# Output: Parameters: 210,000

# Inference
x = torch.randn(1, 3, 224, 224)
out = model(x)  # [1, 8]

Training

# Clone and install
git clone https://github.com/DJLougen/MPKnet.git
cd MPKnet
pip install -r requirements.txt

# Train on your dataset
python train.py --dataset kvasir --epochs 100

---

Architecture

MPKNet models the Lateral Geniculate Nucleus (LGN) - the relay station between retina and visual cortex.

The Three Pathways

Pathway	Biological Role	Implementation	What It Captures
M (Magnocellular)	~10% of LGN, motion, global gist	Stride 5 (coarse)	Shape, motion, layout
P (Parvocellular)	~80% of LGN, fine detail, color	Stride 2 (fine)	Texture, edges, color
K (Koniocellular)	~10% of LGN, projects to M and P	Stride 3 (intermediate)	Context-dependent gating

Core Principles

1. Same kernel, different stride - All pathways use 5×5 kernels. Fibonacci strides (2:3:5) differentiate them, producing resolutions that converge toward the golden ratio.
2. Parallel processing - M/P/K run independently until fusion. No cross-talk within pathways.
3. Late fusion only - No pooling within pathways. Global pool only at the end.
4. K modulates, doesn't process - K-pathway generates cross-stream attention gates for M and P (extends Bahdanau attention, FiLM with biological grounding).

What's Novel

• First Fibonacci strides in CNNs - Derived from biological spatial frequency tuning, not empirical search
• First complete M/P/K implementation - Prior work (Magno-Parvo CNN, EVNets, SlowFast) models M/P only
• Biologically-grounded cross-stream gating - K→M/P gating mirrors koniocellular projections in LGN

Why This Works

Biology processes vision with 20 watts. One hypothesis: efficiency comes from the wiring diagram, not raw neuron count.

MPKNet borrows this principle: I restrict where multiplication happens. M and P process in parallel streams before fusion. K modulates both. The math is standard convolutions. The connectivity pattern is inspired by biology.

"It's what you multiply and where you multiply."

---

Ablation Study

Pathway ablations currently running across a variety of datasets. Results forthcoming.

---

Interpretable Failures

Method: I evaluated V6-Pi on Kvasir-v2 validation set (1600 samples), tracking all misclassifications with confidence scores.

Key finding: 63% of errors (183/292) cluster in just two bidirectional pairs.

Per-Class Accuracy

Class	Accuracy	Confusion Pattern
esophagitis	67.9%	→ normal-z-line (58 errors)
dyed-lifted-polyps	70.4%	→ dyed-resection-margins (51 errors)
polyps	76.9%	Scattered across multiple classes
normal-pylorus	99.0%	Nearly perfect

Top Confusion Pairs (with Confidence)

True Class	→ Predicted	Count	Mean Conf	Range
esophagitis	→ normal-z-line	58	68%	50-94%
dyed-lifted-polyps	→ dyed-resection-margins	51	69%	34-97%
dyed-resection-margins	→ dyed-lifted-polyps	40	60%	30-96%
normal-z-line	→ esophagitis	34	61%	48-81%

What it means: The discriminative signal between these pairs is weak enough that errors concentrate here. External context (patient history, procedure timeline, endoscope position) would help, but is unavailable to any vision-only system.

Failure Categories

Type	Count	% of Failures	Meaning
Confident failures (≥80% conf)	44	15%	Model is wrong but sure — miscalibrated
Ambiguous failures (<50% conf)	22	8%	Model knows it doesn't know — honest
Close calls (<15% margin)	69	24%	True class almost won — fixable

Semantic Group Confusion

Direction	Errors	Clinical Impact
pathology to normal	66	Missed disease
normal to pathology	39	False alarm
polyp to procedure	52	Dye similarity
procedure to polyp	42	Dye similarity

Clinical Limitations

This model is a research prototype, not a clinical tool.

Metric	V6-Pi Result	Clinical Requirement
Polyp sensitivity	~75%	>=95% for screening
Pathology to Normal errors	66 cases	Near zero
Confident false negatives	44 @ 88% conf	Unacceptable

Why it's interpretable: Failures cluster in predictable, explainable pairs rather than scattering randomly across 8 classes. You know which cases need human review and why the model failed.

---

Roadmap

MPKNet V6 implements the LGN stage of mammalian vision. What I'm working on next:

Biological extensions:

• Surround suppression - V1-like center-surround for better edge discrimination
• Temporal M pathway - 3D convolutions in M pathway for video (matches M-cell motion sensitivity)

Biological Extensions

• RGC layer - Midget/Parasol/Bistratified cells feeding M/P/K pathways
• Retinotectal pathway - Superior colliculus for saccades
• V1 orientation columns - Edge detection specialization
• Thalamo-cortical loops - Exploring whether attention-like behavior emerges from architecture alone

Applications

• Detection head - YOLO-style head using M/P as multi-scale FPN
• Medical uncertainty - MC Dropout for epistemic uncertainty quantification
• VLM encoder - Lightweight vision encoder for vision-language models
• Webcam eye tracking - Real-time gaze estimation from eye crops
• Thermal glider fire detection - 3D-printed gliders for wildfire monitoring

Citation

@misc{MPKNet,
  author = {Lougen, D.J.},
  title = {MPKNet: A LGN-Inspired Architecture for Efficient Visual Processing},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/DJLougen/MPKnet}
}

Patent pending: US 63/950,391

---

License & Commercial Use

PolyForm Small Business License with Humanitarian Exception.

Use Case	Cost
Academic research	Free
Personal projects	Free
Startups (<$100K revenue)	Free
Non-profits & NGOs	Free
Educational institutions	Free
Low-income region deployment	Free
Commercial (>$100K revenue)	Contact me

Why This License?

A 76KB model on a $35 Raspberry Pi can enable:

• Research prototypes for medical image analysis (not clinical deployment)
• Agricultural monitoring on small farms
• Educational tools in underfunded schools
• Disaster response with limited infrastructure

These use cases should never be paywalled.

Note: For medical applications, see Clinical Limitations. This model is a research tool, not a diagnostic device.

For commercial licensing: d.lougen@mail.utoronto.ca

---

Acknowledgements

Thanks to Paul Dassonville (UO) for introducing me to these cells, and Jay Pratt (U of T) for ongoing collaboration on koniocellular research.

---

Daniel J. Lougen · University of Toronto