After VGGNet, the next big innovation was the GoogleNet which had won the ILSVRC’14 challenge with an error rate of about 6.7%.
Unlike the previous innovations, which had tried to increase the model capacity by adding more layers, reducing the filter size etc. (such as from AlexNet to VGGNet), GoogleNet had increased the depth using a new type of convolution technique using the Inception module.
The module derives its name from a previous paper by Lin et al and this meme popular in the deep learning community:
Let’s study the key features of GoogleNet architecture.
To summarise, some important features of the GoogleNet architecture are as follows:
- Inception modules stacked on top of each other, total of 22 layers
- Use of 1 x 1 convolutions in the modules
- Parallel convolutions by multiple filters (1×1, 3×3, 5×5)
- Pooling operation of size (3×3)
- No FC layer, except for the last softmax layer for classification
- Number of parameters reduced from 60 million (AlexNet) to 4 million
The details on why the GoogleNet and the inception module work well are beyond the scope of this course, though you are encouraged to read the GoogleNet paper (provided below).
In the next segment, we will look at the architecture of ResNet.
Additional Reading
The previously mentioned GoogleNet paper is given below for your reference:
Report an error