EffiNet

# Optimizing Deep Learning Architectures for Enhanced Computational Efficiency

[](https://doi.org/10.1109/ACCESS.2024.0429000)

This repository implements the framework proposed in the paper:  
**"Optimizing Deep Learning Architectures for Enhanced Computational Efficiency"**  
by **Wang Ying** and **Li Hui**.  
The project focuses on building computationally efficient and interpretable deep learning models through a **Dynamic Compositional Architecture (DCA)** and a **Knowledge-Embedded Adaptive Strategy (KEAS)**.

---

## 🧠 Key Concepts

- **Dynamic Compositional Architecture (DCA)**  
  A modular neural network represented as a directed acyclic graph (DAG) with dynamic node activation based on input semantics.

- **Knowledge-Embedded Adaptive Strategy (KEAS)**  
  Integrates symbolic domain knowledge into the model via graph-based constraints, semantic modulation, and optimization-aware adaptation.

- **Efficiency Gains**  
  - ⚡ Up to **40% improvement** in inference speed  
  - 💾 **30% reduction** in memory usage  
  - 🧠 Maintains or improves accuracy compared to SOTA methods

---

## 🛠️ Features

- ✅ Adaptive computation via input-aware path selection
- 🔍 Symbolic interpretability with semantic trace visualization
- 🔁 Multi-objective training incorporating optimization, semantics, and structural regularization
- 📈 Outperforms SOTA models like LUKE, RoBERTa-NER, ELECTRA on benchmark datasets

---

## 📊 Benchmarks

| Dataset            | Accuracy | F1 Score |
|--------------------|----------|----------|
| ImageNet           | 93.89%   | 91.40%   |
| Fashion-MNIST      | 94.31%   | 91.87%   |
| Pascal VOC         | 93.40%   | 91.27%   |
| Tiny ImageNet      | 94.01%   | 91.72%   |

(See detailed tables in the paper for standard deviation and comparisons.)

---

## 📁 Datasets Used

- [ImageNet](http://www.image-net.org/)
- [Fashion-MNIST](https://github.com/zalandoresearch/fashion-mnist)
- [Pascal VOC](http://host.robots.ox.ac.uk/pascal/VOC/)
- [Tiny ImageNet](https://tiny-imagenet.herokuapp.com/)

---

## 🖥️ Installation

```bash
git clone https://github.com/your-username/efficient-dl-architecture.git
cd efficient-dl-architecture
pip install -r requirements.txt