The stem of an AnyNet architecture serves as the initial processing stage for RGB input images ($$3$$ channels). It typically applies a $$3 	imes 3$$ convolution with a stride of $$2$$, followed by batch normalization. This sequence halves the spatial resolution of the image (from $$r 	imes r$$ to $$r/2 	imes r/2$$) and generates an initial set of $$c_0$$ channels that serve as the input for the network's body.

AnyNet Stem

The body of an AnyNet architecture conducts the majority of feature extraction by operating on images at progressively decreasing resolutions. Typically designed to process ImageNet inputs (e.g., reducing a $$224 \times 224 \times 3$$ image to a $$7 \times 7 \times c_4$$ representation), the body is divided into multiple stages (e.g., $$4$$ stages). Each stage quarters the total spatial resolution (by halving both height and width) and is constructed from a sequence of ResNeXt blocks. The initial block in a stage applies a stride of $$2$$ to downsample the feature map, which necessitates a $$1 \times 1$$ convolution on the residual branch to match dimensions, while all subsequent blocks in the stage maintain constant resolution and channel counts.

AnyNet Body

The head of an AnyNet architecture converts the feature representations extracted by the network body into the final desired predictions. It employs an entirely standard design consisting of global average pooling to collapse spatial dimensions, followed by a fully connected layer to emit an $$n$$-dimensional vector for $$n$$-class classification.

AnyNet Head

The AnyNet design space provides a generic architectural template for constructing and exploring families of convolutional neural networks. This macro-structure is composed of three primary components: a stem that performs the initial image processing, a body that carries out the bulk of the transformations to build object representations across multiple stages, and a head that converts these representations into the final desired outputs (e.g., via a softmax regressor). Within this generic framework, design choices such as stage depth, channel counts, and block structures (often using ResNeXt blocks) parameterize a vast space of potential network configurations.

Claude

The ResNeXt architecture addresses the trade-off between nonlinearity and dimensionality in standard ResNet designs. Instead of increasing network depth or widening convolutions, ResNeXt increases the number of channels that carry information between blocks while avoiding a quadratic computational penalty. Inspired by the Inception block's strategy of separating information flow into independent groups, ResNeXt applies the exact same transformation across all of its parallel branches. This uniform multi-branch design minimizes the need for manual hyperparameter tuning for each branch.

ResNeXt Architecture

Dive into Deep Learning

S. Xie, R. Girshick, P. Dollar, Z. Tu and K. He. Aggregated Residual Transformations for Deep Neural Networks

The ResNeXt block is a residual block design that incorporates grouped convolutions to process information efficiently across independent branches. It consists of a $$3 \times 3$$ grouped convolution sandwiched between two $$1 \times 1$$ convolutions. The first $$1 \times 1$$ convolution reduces the channel dimensionality, while the second $$1 \times 1$$ convolution restores it, effectively exchanging information across the $$g$$ independent groups. This design minimizes computational overhead: the network only pays an $$\mathcal{O}(c \cdot b)$$ cost for the $$1 \times 1$$ kernels and an $$\mathcal{O}(b^2 / g)$$ cost for the $$3 \times 3$$ grouped kernels, where $$c$$ is the number of channels and $$b$$ is the number of intermediate bottleneck channels. Similar to standard residual blocks, a $$1 \times 1$$ convolution can be applied to the residual connection to match dimensions if needed.

Learn Before

Related

Learn After