GoogLeNet, later renamed Inception v1, is a convolutional neural network (CNN) for computer vision introduced by researchers at Google in 2014. It won the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC14), significantly reducing the error rate in image classification. The architecture was historically important as an early CNN that separates the stem (data ingest), body (data processing), and head (prediction), a design that persists in most modern CNN architectures.
The name "GoogLeNet" was a homage to LeNet, the 1998 CNN by Yann LeCun, as both are CNNs. The team also called it "Inception" after a "we need to go deeper" internet meme, a phrase from the 2010 film Inception. Because later versions were released, the original architecture was renamed "Inception v1". The models and code were released under the Apache 2.0 license on GitHub.
Architecture and Key Innovations
Inception v1 is a deep CNN composed of 22 layers, most of which are "Inception modules". The original paper described Inception modules as a "logical culmination" of the Network in Network approach and work by Arora et al. (2014). Each Inception module applies multiple convolutional filters of different sizes (1x1, 3x3, 5x5) and pooling in parallel, then concatenates their outputs, allowing the network to capture features at various scales.
Due to its depth, Inception v1 suffered from the vanishing gradient problem. The team addressed this by inserting two "auxiliary classifiers" at one-third and two-thirds depth within the network. The loss function was a weighted sum of all three classifiers: L = 0.3 L_aux1 + 0.3 L_aux2 + L_real. These auxiliary classifiers were removed after training was complete. This problem was later more fundamentally solved by the ResNet architecture.
The architecture consists of three parts stacked on top of one another:
- The stem (data ingestion): The first few convolutional layers perform data preprocessing to downscale images to a smaller size.
- The body (data processing): The next many Inception modules perform the bulk of data processing.
- The head (prediction): The final fully-connected layer and softmax produces a probability distribution for image classification.
Inception v2
Inception v2 was released in 2015, in a paper more famous for proposing batch normalization. It had 13.6 million parameters. It improved on Inception v1 by adding batch normalization and removing dropout and local response normalization, which the researchers found became unnecessary when batch normalization was used.
Inception v3
Inception v3 was released in 2016. It improved on Inception v2 by using factorized convolutions. For example, a single 5x5 convolution can be factored into two stacked 3x3 convolutions, both having a receptive field of size 5x5. The 5x5 kernel has 25 parameters compared to 18 in the factorized version, making the factorized version more parameter-efficient. The team found empirically that factorized convolutions helped performance.
It also introduced a form of dimension reduction by concatenating the output from a convolutional layer and a pooling layer. For instance, a tensor of size 35x35x320 can be downscaled by a stride-2 convolution to 17x17x320, and by maxpooling with pool size 2x2 to 17x17x320, which are then concatenated to 17x17x640.
Inception v3 also removed the lowest auxiliary classifier during training, finding that the auxiliary head worked as a form of regularization. Additionally, it proposed label-smoothing regularization: for an image with label c, instead of predicting the one-hot distribution δ_c, the model predicts a smoothed distribution (1 - ε)δ_c + ε/K, where K is the total number of classes.
Inception v4 and Inception ResNet
In 2017, the team released Inception v4, Inception ResNet v1, and Inception ResNet v2. Inception v4 was an incremental update with even more factorized convolutions and other complications empirically found to improve benchmarks. Inception ResNet v1 and v2 are both modifications of Inception v4, where residual connections are added to each Inception module, inspired by the ResNet architecture.
Xception
Xception ("Extreme Inception") was published in 2017. It is a linear stack of depthwise separable convolution layers with residual connections. The design was proposed on the hypothesis that in a CNN, the cross-channel correlations and spatial correlations in the feature maps can be entirely decoupled. Training each Xception network took 3 days on 60 K80 GPUs, or approximately 0.5 petaFLOP-days.
The Inception family has been highly influential in the development of deep learning for computer vision, and its architectural principles continue to inform modern CNN design.