CIFNet: Class Incremental Frugal Network

CIFNet (Class Incremental Frugal Network) is an efficient approach to Class Incremental Learning (CIL) that achieves state-of-the-art accuracy while drastically reducing training time and energy consumption. It is designed for resource-constrained environments such as edge devices, where computational efficiency is critical.

CIFNet introduces a single-step optimization process and a compressed buffer mechanism that stores condensed representations of previous data samples, minimizing both computation and memory usage.

---

🏆 This work was developed as part of my Master’s Thesis in the Master in Artificial Intelligence at the University of A Coruña (MIA-UDC).
It was awarded Best Master’s Thesis in Artificial Intelligence (2025) by the AEPIA (Spanish Association for Artificial Intelligence) at the EVIA Summer School on AI 2025.

---

🚀 Key Features

• ⚡ Single-Step Training — Avoids multiple training epochs to reduce compute time.
• 🧠 Compressed Buffer - Stores low-dimensional representations, minimizing memory footprint.
• 🧩 Resistant to Forgetting — Maintains accuracy on past classes without complex regularization.
• 🌱 Energy Efficient - Optimized for low-resource and real-time applications.

🛠️ Installation

Prerequisites

• Python ≥ 3.9
• PyTorch ≥ 2.7.1
• Other dependencies listed in requirements.txt

Setup

1. Clone the repository:

   git clone https://github.com/AlejandroDopico2/CIFNet.git
   cd CIFNet

2. Install the required dependencies:

   pip install -r requirements.txt

3. Download datasets:

   • MNIST and CIFAR datasets are automatically downloaded via torchvision.
   • For ImageNet, download the dataset manually and place it in the tiny-imagenet or imagenet-100 directory.

🧩 Usage

Training

python main.py -p cfgs/CIFAR10_random.yaml

Configuration files are available in the cfgs/ directory. Modify YAML files to customize hyperparameters and dataset paths.

Example Configuration Files

• CIFAR10_random.yaml: for CIFAR-10
• CIFAR100_random.yaml: for CIFAR-100
• ImageNet100_random.yaml: for ImageNet-100

📊 Results

CIFNet achieves competitive accuracy while significantly reducing training time and energy consumption.

Performance on CIFAR-100

Performance on ImageNet-100

Results for 20 tasks config:

Dataset	Accuracy	Training Time	Energy Consumption
CIFAR-100	59.26%	15 min.	0.063 kWh
ImageNet-100	78.10%	71 min.	0.271 kWh

Sustainable Metrics Comparison (ImageNet-100 T=20)

For detailed experiments and ablation studies, see the paper on arXiv.

🤝 Contributing

Contributions are welcome! To contribute:

1. Fork the repository.
2. Create a new branch for your feature or bugfix.
3. Follow PEP-8 coding standards.
4. Submit a pull request with a clear description of your changes.

To report bugs or request features, open an issue in this repository.

📚 Citation

If you use CIFNet in your research, please cite:

@misc{dopicocastro2025efficientsinglestepframeworkincremental,
      title={Efficient Single-Step Framework for Incremental Class Learning in Neural Networks}, 
      author={Alejandro Dopico-Castro and Oscar Fontenla-Romero and Bertha Guijarro-Berdiñas and Amparo Alonso-Betanzos},
      year={2025},
      eprint={2509.11285},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2509.11285}, 
}

🧑‍💻 Authors

• Alejandro Dopico Castro - GitHub
• Óscar Fontenla Romero
• Bertha Guijarro Berdiñas
• Amparo Alonso Betanzos

License

This project is licensed under the MIT License. See the LICENSE file for details.

🧭 Future Work

• Federated Learning adaptation
• Support for dynamic class addition
• Extended experiments on TinyImageNet