ResNet101

ResNet101 is a deep convolutional neural network in the ResNet family, designed for image recognition tasks. Introduced in 2015 by Kaiming He and colleagues at Microsoft Research, it extends the idea of residual learning to deeper networks by using shortcut connections that perform identity mappings, allowing training of networks with many layers without suffering severe vanishing gradients. The 101 in the name indicates the total number of layers, counting 1x1, 3x3, and 1x1 convolutions inside bottleneck blocks and the final fully connected layer.

The architecture uses four stages of bottleneck residual blocks with 3, 4, 23, and 3 blocks respectively, totaling 33 bottleneck blocks. Each bottleneck block consists of a 1x1 convolution to reduce channels, a 3x3 convolution, and a 1x1 convolution to restore channels, with a skip connection that adds the input to the block's output. In the early part of the network a 7x7 convolution with stride 2 and a max-pooling layer reduce spatial dimensions, followed by downsampling between stages.

For alignment of dimensions, projections through 1x1 convolutions are used on the skip path when the number of channels changes. The network ends with a global average pooling layer and a fully connected layer producing 1000-class logits for ImageNet.

ResNet101 is widely used as a feature extractor and pre-trained model; it helped set performance benchmarks at its introduction and remains a common backbone in computer vision. Its residual design has influenced a range of deeper and more efficient architectures that build upon the same core idea.