Going Deeper with Convolutions
- Publication
- CVPR 2015Conference
- First submitted to arXiv
- arXiv:1409.4842
Cite this paper · BibTeX
The citation identifies the linked source version. It cites the original paper and its authors.
@misc{classiccnnsinceptionv12014,
title = {{Going Deeper with Convolutions}},
author = {Christian Szegedy and Wei Liu and Yangqing Jia and Pierre Sermanet and Scott Reed and Dragomir Anguelov and Dumitru Erhan and Vincent Vanhoucke and Andrew Rabinovich},
year = {2014},
eprint = {1409.4842},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/1409.4842}
}
GoogLeNet uses the Inception architecture, inspired by the Network in Network model and the online “we need to go deeper” meme.
It is rare to see a paper so earnestly say that an internet meme inspired its research.
But since the authors boldly wrote this, we’ll share it as it is:
The “depth” of the Inception architecture has two meanings: on one hand, it introduces a new organizational structure called the “Inception module,” and on the other hand, it achieves deeper network architectures.
The name "GoogLeNet" is a tribute to Yann LeCun's LeNet-5 network. This specific version of the Inception architecture was used in the ILSVRC14 competition.
Problem Solved
The Inception Module
The core idea of the Inception architecture is to approximate and cover the optimal sparse structure by using existing dense components.
In the module, 1×1, 3×3, and 5×5 convolution filters are used to process features with different spatial distributions. The feature maps from the branches are concatenated along the channel dimension, which serves as the input for the next stage. To avoid excessively large convolution filters, the size of these filters is constrained more for convenience than necessity.
Since pooling operations are crucial in convolutional neural networks, each Inception module includes a parallel pooling path, and the pooling layer's output is concatenated with the output from the convolutional layers, enriching the feature representation further.
To avoid a sharp rise in computation, the architecture uses 1×1 convolutions to reduce channels before the 3×3 and 5×5 convolutions, as well as after pooling. With ReLU, these 1×1 layers also introduce non-linearity. For example, Table 1's Inception (3a) reduces the 5×5 branch from 192 input channels to 16 before producing 32 output channels: the 5×5 weights change from to , though the added 1×1 convolution also has a cost.
As the Inception modules are stacked, the abstraction level of features increases layer by layer, and the concentration of higher-level feature spaces gradually decreases, meaning the ratio of 3×3 and 5×5 convolutions increases as the network deepens.
Configuration Parameters
The table above lists some important configuration parameters of GoogLeNet.
The input crop size of GoogLeNet is , processing RGB images with mean subtraction. "#3×3 reduce" and "#5×5 reduce" indicate the number of 1×1 filters used before the 3×3 and 5×5 convolutions, while "pool proj" refers to the number of 1×1 filters in the projection layer following the max-pooling operation.
The depth of the GoogLeNet architecture, when only considering layers with parameters, is 22 layers, and 27 layers if pooling layers are included. The overall network contains around 100 independent building modules, with the exact number depending on the machine learning infrastructure used. An average pooling layer is used before the classifier, and in practice, an additional linear layer is added, which makes the network easier to adapt to other label sets.
Replacing the fully connected layers with average pooling improved top-1 accuracy by around 0.6 percentage points. However, despite removing the fully connected layers, using Dropout remains essential, with a 40% dropout rate in the main classification branch. The auxiliary classifiers described below use 70%; these are separate settings.
Model Architecture
Due to the depth of the network, effectively backpropagating gradients to all layers is a design concern. To address this, the authors introduced auxiliary classifiers, placed on top of the outputs of Inception modules (4a) and (4d).
The design of the auxiliary classifiers includes a 5×5 average pooling layer, a 1×1 convolution layer (with 128 filters), a fully connected layer with 1024 units, a Dropout layer with a 70% dropout rate, and a linear layer. These classifiers help reinforce the backpropagation of gradient signals during training, improving the discriminative power of lower-level features. During training, the loss from these auxiliary classifiers is weighted by 0.3 and added to the total loss, but the auxiliary networks are removed during inference.
Discussion
Performance at ILSVRC 2014
The task of the ILSVRC 2014 competition was to classify images into one of the 1,000 leaf node categories in the Imagenet hierarchy.
The training set contained around 1.2 million images, the validation set 50,000 images, and the test set 100,000 images. Each image had a ground-truth label; evaluation used top-1 accuracy and top-5 error, while the table here reports top-5 error.
The authors did not use any external data for training during the competition, relying solely on the data provided in the training set. Additionally, several techniques were employed during testing to enhance performance:
- Seven independent versions of the GoogLeNet model were trained, and predictions were made using an ensemble method.
- Multiple crops of the input were used, and the predictions from all independent classifiers were averaged to produce the final result.
The final challenge submission achieved 6.67% top-5 error on both the validation and test sets, ranking first. This was an ensemble of seven models evaluated on 144 crops per image. In the paper's Table 3, one model on one crop has 10.07% top-5 error on the validation set. The 6.67% result therefore includes the full testing setup; it is not the score of a single GoogLeNet on a single crop.
Conclusion
GoogLeNet shows how the Inception design can balance computation and model performance. It also explores the potential of sparse structures, using readily computed dense components as an approximation and pointing toward more refined network designs.
Its scalable design became an important reference for later research, especially where efficiency and performance both matter.
This article is just the beginning of the Inception series. Later iterations, such as Inception-V2, V3, and V4, introduced various improvements and optimizations. We will explore these in due time.