Going deeper with convolutions

This paper proposes a deep convolutional neural network architecture called Inception, designed to achieve high accuracy while carefully managing computational cost. Rather than simply stacking more and larger layers, the design increases both the depth and width of the network while keeping the computational budget constant, with architectural decisions grounded in the Hebbian principle and an intuition for multi-scale processing to improve how computing resources are utilized within the network.

The specific instantiation submitted to the ImageNet Large-Scale Visual Recognition Challenge 2014, a 22-layer network called GoogLeNet, achieved new state-of-the-art results in both the classification and detection tracks of the competition. This demonstrated that carefully engineered, resource-efficient architectural modules could outperform simply making networks deeper or wider in an unconstrained way, influencing subsequent efficient network designs.