Residual neural network

In a multilayer neural network model, consider a subnetwork with a certain number of stacked layers (e.g., 2 or 3). Denote the underlying function performed by this subnetwork as ${\displaystyle H(x)}$ , where ${\displaystyle x}$  is the input to the subnetwork. Residual learning re-parameterizes this subnetwork and lets the parameter layers represent a "residual function" ${\displaystyle F(x)=H(x)-x}$ . The output ${\displaystyle y}$  of this subnetwork is then represented as:

${\displaystyle y=F(x)+x}$ 

The operation of "${\displaystyle +\ x}$ " is implemented via a "skip connection" that performs an identity mapping to connect the input of the subnetwork with its output. This connection is referred to as a "residual connection" in later work. The function ${\displaystyle F(x)}$  is often represented by matrix multiplication interlaced with activation functions and normalization operations (e.g., batch normalization or layer normalization). As a whole, one of these subnetworks is referred to as a "residual block".^[1] A deep residual network is constructed by simply stacking these blocks.

Long short-term memory (LSTM) has a memory mechanism that serves as a residual connection.^[4] In an LSTM without a forget gate, an input ${\displaystyle x_{t}}$  is processed by a function ${\displaystyle F}$  and added to a memory cell ${\displaystyle c_{t}}$ , resulting in ${\displaystyle c_{t+1}=c_{t}+F(x_{t})}$ . An LSTM with a forget gate essentially functions as a highway network.

To stabilize the variance of the layers' inputs, it is recommended to replace the residual connections ${\displaystyle x+f(x)}$  with ${\displaystyle x/L+f(x)}$ , where ${\displaystyle L}$  is the total number of residual layers.^[5]

If the function ${\displaystyle F}$  is of type ${\displaystyle F:\mathbb {R} ^{n}\to \mathbb {R} ^{m}}$  where ${\displaystyle n\neq m}$ , then ${\displaystyle F(x)+x}$  is undefined. To handle this special case, a projection connection is used:

${\displaystyle y=F(x)+P(x)}$ 

where ${\displaystyle P}$  is typically a linear projection, defined by ${\displaystyle P(x)=Mx}$  where ${\displaystyle M}$  is a ${\displaystyle m\times n}$  matrix. The matrix is trained via backpropagation, as is any other parameter of the model.

The introduction of identity mappings facilitates signal propagation in both forward and backward paths.^[6]

If the output of the ${\displaystyle \ell }$ -th residual block is the input to the ${\displaystyle (\ell +1)}$ -th residual block (assuming no activation function between blocks), then the ${\displaystyle (\ell +1)}$ -th input is:

${\displaystyle x_{\ell +1}=F(x_{\ell })+x_{\ell }}$ 

Applying this formulation recursively, e.g.:

${\displaystyle {\begin{aligned}x_{\ell +2}&=F(x_{\ell +1})+x_{\ell +1}\\&=F(x_{\ell +1})+F(x_{\ell })+x_{\ell }\end{aligned}}}$ 

yields the general relationship:

${\displaystyle x_{L}=x_{\ell }+\sum _{i=\ell }^{L-1}F(x_{i})}$ 

where ${\textstyle L}$  is the index of a residual block and ${\textstyle \ell }$  is the index of some earlier block. This formulation suggests that there is always a signal that is directly sent from a shallower block ${\textstyle \ell }$  to a deeper block ${\textstyle L}$ .

The residual learning formulation provides the added benefit of mitigating the vanishing gradient problem to some extent. However, it is crucial to acknowledge that the vanishing gradient issue is not the root cause of the degradation problem, which is tackled through the use of normalization. To observe the effect of residual blocks on backpropagation, consider the partial derivative of a loss function ${\displaystyle {\mathcal {E}}}$  with respect to some residual block input ${\displaystyle x_{\ell }}$ . Using the equation above from forward propagation for a later residual block ${\displaystyle L>\ell }$ :^[6]

${\displaystyle {\begin{aligned}{\frac {\partial {\mathcal {E}}}{\partial x_{\ell }}}&={\frac {\partial {\mathcal {E}}}{\partial x_{L}}}{\frac {\partial x_{L}}{\partial x_{\ell }}}\\&={\frac {\partial {\mathcal {E}}}{\partial x_{L}}}\left(1+{\frac {\partial }{\partial x_{\ell }}}\sum _{i=\ell }^{L-1}F(x_{i})\right)\\&={\frac {\partial {\mathcal {E}}}{\partial x_{L}}}+{\frac {\partial {\mathcal {E}}}{\partial x_{L}}}{\frac {\partial }{\partial x_{\ell }}}\sum _{i=\ell }^{L-1}F(x_{i})\end{aligned}}}$ 

This formulation suggests that the gradient computation of a shallower layer, ${\textstyle {\frac {\partial {\mathcal {E}}}{\partial x_{\ell }}}}$ , always has a later term ${\textstyle {\frac {\partial {\mathcal {E}}}{\partial x_{L}}}}$  that is directly added. Even if the gradients of the ${\displaystyle F(x_{i})}$  terms are small, the total gradient ${\textstyle {\frac {\partial {\mathcal {E}}}{\partial x_{\ell }}}}$  resists vanishing due to the added term ${\textstyle {\frac {\partial {\mathcal {E}}}{\partial x_{L}}}}$ .

A basic block is the simplest building block studied in the original ResNet.^[1] This block consists of two sequential 3x3 convolutional layers and a residual connection. The input and output dimensions of both layers are equal.

A bottleneck block^[1] consists of three sequential convolutional layers and a residual connection. The first layer in this block is a 1x1 convolution for dimension reduction (e.g., to 1/2 of the input dimension); the second layer performs a 3x3 convolution; the last layer is another 1x1 convolution for dimension restoration. The models of ResNet-50, ResNet-101, and ResNet-152 are all based on bottleneck blocks.^[1]

The pre-activation residual block^[6] applies activation functions before applying the residual function ${\displaystyle F}$ . Formally, the computation of a pre-activation residual block can be written as:

${\displaystyle x_{\ell +1}=F(\phi (x_{\ell }))+x_{\ell }}$ 

where ${\displaystyle \phi }$  can be any activation (e.g. ReLU) or normalization (e.g. LayerNorm) operation. This design reduces the number of non-identity mappings between residual blocks. This design was used to train models with 200 to over 1000 layers.^[6]

Since GPT-2, transformer blocks have been mostly implemented as pre-activation blocks. This is often referred to as "pre-normalization" in the literature of transformer models.^[7]

Originally, ResNet was designed for computer vision.^[1][8][9]

All transformer architectures include residual connections. Indeed, very deep transformers cannot be trained without them.^[10]

The original ResNet paper made no claim on being inspired by biological systems. However, later research has related ResNet to biologically-plausible algorithms.^[11][12]

A study published in Science in 2023^[13] disclosed the complete connectome of an insect brain (specifically that of a fruit fly larva). This study discovered "multilayer shortcuts" that resemble the skip connections in artificial neural networks, including ResNets.

Residual connections were noticed in neuroanatomy, such as Lorente de No (1938).^{[14]: Fig 3} McCulloch and Pitts (1943) proposed artificial neural networks and considered those with residual connections.^{[15]: Fig 1.h}

In 1961, Frank Rosenblatt described a three-layer multilayer perceptron (MLP) model with skip connections.^{[16]: 313, Chapter 15} The model was referred to as a "cross-coupled system", and the skip connections were forms of cross-coupled connections.

During the late 1980s, "skip-layer" connections were sometimes used in neural networks. Examples include:^[17][18] Lang and Witbrock (1988)^[19] trained a fully connected feedforward network where each layer skip-connects to all subsequent layers, like the later DenseNet (2016). In this work, the residual connection was the form ${\displaystyle x\mapsto F(x)+P(x)}$ , where ${\displaystyle P}$  is a randomly-initialized projection connection. They termed it a "short-cut connection".

Sepp Hochreiter discovered the vanishing gradient problem in 1991^[20] and argued that it explained why the then-prevalent forms of recurrent neural networks did not work for long sequences. He and Schmidhuber later designed the LSTM architecture to solve this problem,^[4][21] which has a "cell state" ${\displaystyle c_{t}}$  that can function as a generalized residual connection. The highway network (2015)^[22][23] applied the idea of an LSTM unfolded in time to feedforward neural networks, resulting in the highway network. ResNet is equivalent to an open-gated highway network.

During the early days of deep learning, there were attempts to train increasingly deep models. Notable examples included the AlexNet (2012), which had 8 layers, and the VGG-19 (2014), which had 19 layers.^[24] However, stacking too many layers led to a steep reduction in training accuracy,^[25] known as the "degradation" problem.^[1] In theory, adding additional layers to deepen a network should not result in a higher training loss, but this is what happened with VGGNet.^[1] If the extra layers can be set as identity mappings, however, then the deeper network would represent the same function as its shallower counterpart. There is some evidence that the optimizer is not able to approach identity mappings for the parameterized layers, and the benefit of residual connections was to allow identity mappings by default.^[26]

In 2014, the state of the art was training deep neural networks with 20 to 30 layers.^[27] The research team for ResNet attempted to train deeper ones by empirically testing various methods for training deeper networks, until they came upon the ResNet architecture.^[28]

DenseNet (2016)^[29] connects the output of each layer to the input to each subsequent layer:

${\displaystyle x_{\ell +1}=F(x_{1},x_{2},\dots ,x_{\ell -1},x_{\ell })}$ 

Stochastic depth^[30] is a regularization method that randomly drops a subset of layers and lets the signal propagate through the identity skip connections. Also known as DropPath, this regularizes training for deep models, such as vision transformers.^[31]

ResNeXt (2017) combines the Inception module with ResNet.^[32][8]

Squeeze-and-Excitation Networks (2018) added squeeze-and-excitation (SE) modules to ResNet.^[33] An SE module is applied after a convolution, and takes a tensor of shape ${\displaystyle \mathbb {R} ^{H\times W\times C}}$  (height, width, channels) as input. Each channel is averaged, resulting in a vector of shape ${\displaystyle \mathbb {R} ^{C}}$ . This is then passed through a multilayer perceptron (with an architecture such as linear-ReLU-linear-sigmoid) before it is multiplied with the original tensor.