Increase only the ratio of pooling branches.

Increase the ratio of both and convolutions.

Increase only the ratio of convolutions.

Keep the ratios of all branches fixed throughout the network.

About 12 times fewer parameters, while being more accurate.

About 12 times fewer parameters, but achieved similar accuracy.

About the same number of parameters, and achieved higher accuracy.

About 12 times more parameters, but achieved higher accuracy.

To reduce the number of channels and therefore lower the cost of the following convolutions.

To merge the branch outputs by element-wise addition.

To halve the spatial resolution of the feature maps.

To enlarge the spatial receptive field without changing the channel count.

It requires a pre-trained model and prevents parallel processing.

It makes each gradient harder to compute and increases computational cost.

It increases the number of parameters and the spatial resolution.

It increases the risk of overfitting and dramatically increases computational cost.

They used a majority vote over the top-1 class predicted by each crop.

They averaged the pre-softmax logits over crops but used only the best classifier.

They selected the maximum probability over crops and then averaged the classifiers.

They averaged the softmax probabilities over both the crops and the individual classifiers.

They are used to see where we can prune the network to reduce the number of parameters without losing much accuracy.

They remove unlikely classes before the final classifier and remain part of the deployed network.

During training, their losses are added to encourage discriminative intermediate features, improve gradient propagation, and provide regularization; the classifiers are discarded at inference time.

During training, their losses are added to encourage discriminative intermediate features, improve gradient propagation, and provide regularization; their predicted classes are averaged with the main classifier at inference time.

The use of external training data in the ILSVRC 2014 competition.

More aggressive data augmentation.

New ideas, algorithms, and improved network architectures.

More powerful hardware, larger datasets, and bigger models.

Their spatial dimensions are kept equal, using appropriate padding, and they are concatenated along the channel dimension.

They are combined by element-wise addition, so all branches must have the same number of channels.

They are cropped to the smallest spatial size and concatenated along a spatial dimension.

They are flattened, concatenated, and then reshaped into feature maps.

It scales as .

It scales as .

It scales as .

It scales as .