SqueezeNet is a deep neural network for image classification released in 2016. It was developed by researchers at DeepScale, the University of California, Berkeley, and Stanford University. The primary design goal was to create a smaller neural network with fewer parameters while maintaining competitive accuracy. The best-performing SqueezeNet model achieved the same accuracy as AlexNet on ImageNet classification, but with a size 510 times smaller than AlexNet's 240 MB parameter file, reducing it to just 5 MB.
SqueezeNet's architecture relies on "fire modules," which combine a squeeze layer (1x1 convolutions) to reduce the number of channels, followed by an expand layer (a mix of 1x1 and 3x3 convolutions) to increase feature maps. This design drastically reduces the parameter count compared to traditional convolutional networks. The model is not a compressed version of AlexNet but an entirely distinct architecture that happens to achieve similar accuracy on the ImageNet validation dataset.
Version History
SqueezeNet was originally released on February 22, 2016, implemented on top of the Caffe deep learning framework. Shortly after, the open-source community ported it to other frameworks. On February 26, 2016, Eddie Bell released a port for Chainer; on March 2, 2016, Guo Haria released a port for Apache MXNet; and on June 3, 2016, Tammy Yang released a port for Keras. In 2017, companies including Baidu, Xilinx, Imagination Technologies, and Synopsys demonstrated SqueezeNet running on low-power platforms such as smartphones, FPGAs, and custom processors. As of 2018, SqueezeNet ships natively in frameworks like PyTorch, Apache MXNet, and Apple CoreML, with third-party implementations available for TensorFlow.
Relationship to Other Networks
AlexNet
SqueezeNet was originally described in the paper "SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size." AlexNet has 240 MB of parameters, while SqueezeNet has just 5 MB. This small size makes SqueezeNet easier to fit into computer memory and transmit over a network. However, SqueezeNet is not a squeezed version of AlexNet; it is an entirely different deep neural network architecture. Both achieve approximately the same accuracy on the ImageNet image classification validation dataset.
Model Compression
Model compression techniques, such as quantization and pruning, can be applied after training. In the SqueezeNet paper, the authors demonstrated that a technique called Deep Compression can reduce the parameter file from 5 MB to 500 KB. Deep Compression has also been applied to other networks like AlexNet and VGG, but SqueezeNet's small baseline makes it particularly suitable for edge deployment.
Variants
Some original SqueezeNet team members continued developing resource-efficient networks, leading to variants like SqueezeNext and other "squeeze"-family models. The open-source community has also extended SqueezeNet to other applications, including semantic segmentation of images and style transfer. These variants and adaptations maintain the core principle of balancing accuracy with parameter efficiency.