SqueezeNet: AlexNet-level Accuracy with 50x Fewer Parameters

Authors: Iandola, F. N., Han, S., Moskewicz, M. W., et al. Year: 2016 Venue: ICLR ArXiv: https://arxiv.org/abs/1602.07360

Summary

SqueezeNet achieves AlexNet accuracy with 50x fewer parameters using Fire modules. It enables efficient inference on embedded devices and pioneered parameter-efficient architectures.

Key Concepts

• Fire modules (1x1 and 3x3 convolutions)
• Aggressive parameter reduction
• Model compression techniques
• AlexNet-level accuracy on CIFAR-10
• Efficient mobile deployment

Impact

SqueezeNet demonstrated that parameter efficiency was achievable without major accuracy loss, inspiring research into compact architectures.

Related Papers

• MobileNet: Efficient Networks
• ShuffleNet: Channel Shuffle
• EfficientNet: Model Scaling

---

Citation:

@inproceedings{iandola2016squeezenet,
  title={SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and $<$0.5 MB model size},
  author={Iandola, Forrest N and Han, Song and Moskewicz, Matthew W and others},
  booktitle={International Conference on Learning Representations},
  year={2016}
}