About the Project

Inspiration

Modern deep learning models continue to grow larger and more expensive to train, while real-world deployments increasingly demand faster convergence, robustness, and efficient inference on constrained hardware. This project was inspired by recent work on dendritic neural architectures, which suggest that adding structured, error-correcting branches to standard neural networks can improve learning dynamics without changing the core backbone.

Rather than designing a new detector from scratch, I focused on a practical question:

Can dendritic computation be integrated into a production-grade object detector like YOLO to improve convergence and enable efficient edge deployment?

YOLO was chosen because it is widely used in real-world systems and represents realistic engineering constraints, making it a strong testbed for evaluating dendritic optimization beyond toy datasets.


What I Built

This project implements Dendritic YOLO, a modified YOLOv8 pipeline where dendritic convolutional layers are injected into the detection head while keeping the backbone unchanged.

Each dendritic layer augments a standard convolution with multiple lightweight, parallel branches that act as error-correcting pathways. The output of a dendritic convolution is defined as:

$$ y = f(x) + \alpha \cdot \frac{1}{N} \sum_{i=1}^{N} g_i(x) $$

where:

  • \( f(x) \) is the original convolution,
  • \( g_i(x) \) are dendritic branches,
  • \( N \) is the number of branches,
  • \( \alpha \) controls the dendritic contribution.

In this implementation:

  • 6 dendritic branches are used per convolution,
  • branches are depthwise-separable to reduce parameter overhead,
  • dendrites are injected only into YOLO’s detection head (cv3 layers),
  • the backbone is frozen to reduce training cost and isolate dendritic effects.

The full pipeline consists of:

  1. Loading pretrained YOLOv8 weights
  2. Injecting dendritic convolutions into the detection head
  3. Training only the head and dendrites on COCO
  4. Running hyperparameter sweeps for learning rate
  5. Applying 40 percent post-training pruning for edge deployment

What I Learned

This project provided several key insights:

  • Dendritic branches improve convergence speed, achieving higher mAP earlier than the baseline under identical settings.
  • Structural inductive bias matters: small architectural changes can meaningfully affect learning dynamics without increasing model size.
  • Production constraints influence research decisions, especially when working with CPU-only inference, video processing, and memory limits.
  • Full end-to-end retraining is not always necessary: freezing the backbone while training dendrites and the detection head was sufficient to observe gains.

The project also strengthened my understanding of YOLO internals, PyTorch model modification, and the trade-offs between accuracy, latency, and deployability.


Challenges Faced

Several challenges emerged during development:

  • Inference speed on CPU: YOLOv8L is extremely slow on CPU-only environments, requiring careful use of frame skipping and resolution reduction.
  • Video inference stability: containerized environments sometimes produced partially written video outputs, requiring explicit handling of file finalization.
  • Deployment limitations: hosting interactive demos introduced constraints unrelated to model correctness, highlighting the gap between research code and production systems.
  • Balancing rigor and feasibility: trade-offs were required to keep the project both technically meaningful and hackathon-appropriate.

Takeaway

This project demonstrates that dendritic optimization can be applied to large-scale, real-world object detection models like YOLO. With careful integration, dendritic architectures can improve convergence and support aggressive compression without redesigning the entire network.

Overall, the work serves as a proof-of-concept that biologically inspired computation can coexist with modern deep learning pipelines in a practical, engineering-focused setting.

Built With

Share this project:

Updates