SqueezeNet

SqueezeNet delivers AlexNet-level image classification accuracy with 50x fewer parameters, making it the compact standard for resource-constrained edge devices.

SqueezeNet is a high-efficiency Convolutional Neural Network (CNN) for image classification, engineered by researchers from DeepScale, UC Berkeley, and Stanford. Its architectural innovation centers on the 'Fire module,' which strategically uses 1x1 convolution (the 'squeeze' layer) to minimize input channels to the 3x3 filters, drastically reducing the parameter count. This design achieves AlexNet-level accuracy on ImageNet. Critically, the model's parameter count is 50x smaller than AlexNet; when combined with Deep Compression, the file size drops to less than 0.5MB, a 510x reduction. This small footprint makes SqueezeNet essential for deployment on memory-limited hardware like FPGAs and mobile platforms.

https://arxiv.org/abs/1602.07360