Image Classification

Spatial-Reduction Attention, or SRA, is a multi-head attention module used in the Pyramid Vision Transformer architecture which reduces the spatial scale of the key $K$ and value $V$ before the attention operation. This reduces the computational/memory overhead. Details of the SRA in the stage $i$ can be formulated as follows: $$ \text{SRA}(Q, K, V)=\text { Concat }\left(\operatorname{head}\_{0}, \ldots \text { head }\_{N\_{i}}\right) W^{O} $$ $$\text{ head}\_{j}=\text { Attention }\left(Q W\_{j}^{Q}, \operatorname{SR}(K) W\_{j}^{K}, \operatorname{SR}(V) W\_{j}^{V}\right) $$ where Concat $(\cdot)$ is the concatenation operation. $W\_{j}^{Q} \in \mathbb{R}^{C\_{i} \times d\_{\text {head }}}$, $W\_{j}^{K} \in \mathbb{R}^{C\_{i} \times d\_{\text {head }}}$, $W\_{j}^{V} \in \mathbb{R}^{C\_{i} \times d\_{\text {head }}}$, and $W^{O} \in \mathbb{R}^{C\_{i} \times C\_{i}}$ are linear projection parameters. $N\_{i}$ is the head number of the attention layer in Stage $i$. Therefore, the dimension of each head (i.e. $\left.d\_{\text {head }}\right)$ is equal to $\frac{C\_{i}}{N\_{i}} . \text{SR}(\cdot)$ is the operation for reducing the spatial dimension of the input sequence ($K$ or $V$ ), which is written as: $$ \text{SR}(\mathbf{x})=\text{Norm}\left(\operatorname{Reshape}\left(\mathbf{x}, R\_{i}\right) W^{S}\right) $$ Here, $\mathbf{x} \in \mathbb{R}^{\left(H\_{i} W\_{i}\right) \times C\_{i}}$ represents a input sequence, and $R\_{i}$ denotes the reduction ratio of the attention layers in Stage $i .$ Reshape $\left(\mathbf{x}, R\_{i}\right)$ is an operation of reshaping the input sequence $\mathbf{x}$ to a sequence of size $\frac{H\_{i} W\_{i}}{R\_{i}^{2}} \times\left(R\_{i}^{2} C\_{i}\right)$. $W\_{S} \in \mathbb{R}^{\left(R\_{i}^{2} C\_{i}\right) \times C\_{i}}$ is a linear projection that reduces the dimension of the input sequence to $C\_{i}$. $\text{Norm}(\cdot)$ refers to layer normalization.

Reduction-A is an image model block used in the Inception-v4 architecture.

A Selective Kernel unit is a bottleneck block consisting of a sequence of 1×1 convolution, SK convolution and 1×1 convolution. It was proposed as part of the SKNet CNN architecture. In general, all the large kernel convolutions in the original bottleneck blocks in ResNeXt are replaced by the proposed SK convolutions, enabling the network to choose appropriate receptive field sizes in an adaptive manner. In SK units, there are three important hyper-parameters which determine the final settings of SK convolutions: the number of paths $M$ that determines the number of choices of different kernels to be aggregated, the group number $G$ that controls the cardinality of each path, and the reduction ratio $r$ that controls the number of parameters in the fuse operator. One typical setting of SK convolutions is $\text{SK}\left[M, G, r\right]$ to be $\text{SK}\left[2, 32, 16\right]$.

Manifold Mixup is a regularization method that encourages neural networks to predict less confidently on interpolations of hidden representations. It leverages semantic interpolations as an additional training signal, obtaining neural networks with smoother decision boundaries at multiple levels of representation. As a result, neural networks trained with Manifold Mixup learn class-representations with fewer directions of variance. Consider training a deep neural network $f\left(x\right) = f\_{k}\left(g\_{k}\left(x\right)\right)$, where $g\_{k}$ denotes the part of the neural network mapping the input data to the hidden representation at layer $k$, and $f\_{k}$ denotes the part mapping such hidden representation to the output $f\left(x\right)$. Training $f$ using Manifold Mixup is performed in five steps: (1) Select a random layer $k$ from a set of eligible layers $S$ in the neural network. This set may include the input layer $g\_{0}\left(x\right)$. (2) Process two random data minibatches $\left(x, y\right)$ and $\left(x', y'\right)$ as usual, until reaching layer $k$. This provides us with two intermediate minibatches $\left(g\_{k}\left(x\right), y\right)$ and $\left(g\_{k}\left(x'\right), y'\right)$. (3) Perform Input Mixup on these intermediate minibatches. This produces the mixed minibatch: $$ \left(\tilde{g}\_{k}, \tilde{y}\right) = \left(\text{Mix}\_{\lambda}\left(g\_{k}\left(x\right), g\_{k}\left(x'\right)\right), \text{Mix}\_{\lambda}\left(y, y'\right )\right), $$ where $\text{Mix}\_{\lambda}\left(a, b\right) = \lambda \cdot a + \left(1 − \lambda\right) \cdot b$. Here, $\left(y, y' \right)$ are one-hot labels, and the mixing coefficient $\lambda \sim \text{Beta}\left(\alpha, \alpha\right)$ as in mixup. For instance, $\alpha = 1.0$ is equivalent to sampling $\lambda \sim U\left(0, 1\right)$. (4) Continue the forward pass in the network from layer $k$ until the output using the mixed minibatch $\left(\tilde{g}\_{k}, \tilde{y}\right)$. (5) This output is used to compute the loss value and gradients that update all the parameters of the neural network.

Mixture model network (MoNet) is a general framework allowing to design convolutional deep architectures on non-Euclidean domains such as graphs and manifolds. Image and description from: Geometric deep learning on graphs and manifolds using mixture model CNNs

Res2Net is an image model that employs a variation on bottleneck residual blocks. The motivation is to be able to represent features at multiple scales. This is achieved through a novel building block for CNNs that constructs hierarchical residual-like connections within one single residual block. This represents multi-scale features at a granular level and increases the range of receptive fields for each network layer.

A Res2Net Block is an image model block that constructs hierarchical residual-like connections within one single residual block. It was proposed as part of the Res2Net CNN architecture. The block represents multi-scale features at a granular level and increases the range of receptive fields for each network layer. The $3 \times 3$ filters of $n$ channels is replaced with a set of smaller filter groups, each with $w$ channels. These smaller filter groups are connected in a hierarchical residual-like style to increase the number of scales that the output features can represent. Specifically, we divide input feature maps into several groups. A group of filters first extracts features from a group of input feature maps. Output features of the previous group are then sent to the next group of filters along with another group of input feature maps. This process repeats several times until all input feature maps are processed. Finally, feature maps from all groups are concatenated and sent to another group of $1 \times 1$ filters to fuse information altogether. Along with any possible path in which input features are transformed to output features, the equivalent receptive field increases whenever it passes a $3 \times 3$ filter, resulting in many equivalent feature scales due to combination effects. One way of thinking of these blocks is that they expose a new dimension, scale, alongside the existing dimensions of depth, width, and cardinality.

A Ghost BottleNeck is a skip connection block, similar to the basic residual block in ResNet in which several convolutional layers and shortcuts are integrated, but stacks Ghost Modules instead (two stacked Ghost modules). It was proposed as part of the GhostNet CNN architecture. The first Ghost module acts as an expansion layer increasing the number of channels. The ratio between the number of the output channels and that of the input is referred to as the expansion ratio. The second Ghost module reduces the number of channels to match the shortcut path. Then the shortcut is connected between the inputs and the outputs of these two Ghost modules. The batch normalization (BN) and ReLU nonlinearity are applied after each layer, except that ReLU is not used after the second Ghost module as suggested by MobileNetV2. The Ghost bottleneck described above is for stride=1. As for the case where stride=2, the shortcut path is implemented by a downsampling layer and a depthwise convolution with stride=2 is inserted between the two Ghost modules. In practice, the primary convolution in Ghost module here is pointwise convolution for its efficiency.

RegNetY is a convolutional network design space with simple, regular models with parameters: depth $d$, initial width $w\_{0} > 0$, and slope $w\_{a} > 0$, and generates a different block width $u\_{j}$ for each block $j < d$. The key restriction for the RegNet types of model is that there is a linear parameterisation of block widths (the design space only contains models with this linear structure): $$ u\_{j} = w\_{0} + w\_{a}\cdot{j} $$ For RegNetX we have additional restrictions: we set $b = 1$ (the bottleneck ratio), $12 \leq d \leq 28$, and $w\_{m} \geq 2$ (the width multiplier). For RegNetY we make one change, which is to include Squeeze-and-Excitation blocks.

Neural Architecture Search (NAS) learns a modular architecture which can be transferred from a small dataset to a large dataset. The method does this by reducing the problem of learning best convolutional architectures to the problem of learning a small convolutional cell. The cell can then be stacked in series to handle larger images and more complex datasets. Note that this refers to the original method referred to as NAS - there is also a broader category of methods called "neural architecture search".

Just as dropout prevents co-adaptation of activations, DropPath prevents co-adaptation of parallel paths in networks such as FractalNets by randomly dropping operands of the join layers. This discourages the network from using one input path as an anchor and another as a corrective term (a configuration that, if not prevented, is prone to overfitting). Two sampling strategies are: - Local: a join drops each input with fixed probability, but we make sure at least one survives. - Global: a single path is selected for the entire network. We restrict this path to be a single column, thereby promoting individual columns as independently strong predictors.

A GhostNet is a type of convolutional neural network that is built using Ghost modules, which aim to generate more features by using fewer parameters (allowing for greater efficiency). GhostNet mainly consists of a stack of Ghost bottlenecks with the Ghost modules as the building block. The first layer is a standard convolutional layer with 16 filters, then a series of Ghost bottlenecks with gradually increased channels follow. These Ghost bottlenecks are grouped into different stages according to the sizes of their input feature maps. All the Ghost bottlenecks are applied with stride=1 except that the last one in each stage is with stride=2. At last a global average pooling and a convolutional layer are utilized to transform the feature maps to a 1280-dimensional feature vector for final classification. The squeeze and excite (SE) module is also applied to the residual layer in some ghost bottlenecks. In contrast to MobileNetV3, GhostNet does not use hard-swish nonlinearity function due to its large latency.

MobileViT is a vision transformer that is tuned to mobile phone

Neighborhood Attention is a restricted self attention pattern in which each token's receptive field is limited to its nearest neighboring pixels. It was proposed in Neighborhood Attention Transformer as an alternative to other local attention mechanisms used in Hierarchical Vision Transformers. NA is in concept similar to stand alone self attention (SASA), in that both can be implemented with a raster scan sliding window operation over the key value pair. However, NA would require a modification to handle corner pixels, which helps maintain a fixed receptive field size and an increased number of relative positions. The primary challenge in experimenting with both NA and SASA has been computation. Simply extracting key values for each query is slow, takes up a large amount of memory, and is eventually intractable at scale. NA was therefore implemented through a new CUDA extension to PyTorch, NATTEN.

ShuffleNet V2 Block is an image model block used in the ShuffleNet V2 architecture, where speed is the metric optimized for (instead of indirect ones like FLOPs). It utilizes a simple operator called channel split. At the beginning of each unit, the input of $c$ feature channels are split into two branches with $c - c'$ and $c'$ channels, respectively. Following G3, one branch remains as identity. The other branch consists of three convolutions with the same input and output channels to satisfy G1. The two $1\times1$ convolutions are no longer group-wise, unlike the original ShuffleNet. This is partially to follow G2, and partially because the split operation already produces two groups. After convolution, the two branches are concatenated. So, the number of channels keeps the same (G1). The same “channel shuffle” operation as in ShuffleNet is then used to enable information communication between the two branches. The motivation behind channel split is that alternative architectures, where pointwise group convolutions and bottleneck structures are used, lead to increased memory access cost. Additionally more network fragmentation with group convolutions reduces parallelism (less friendly for GPU), and the element-wise addition operation, while they have low FLOPs, have high memory access cost. Channel split is an alternative where we can maintain a large number of equally wide channels (equally wide minimizes memory access cost) without having dense convolutions or too many groups.

AugMix mixes augmented images through linear interpolations. Consequently it is like Mixup but instead mixes augmented versions of the same image.

MetaFormer is a general architecture abstracted from Transformers by not specifying the token mixer.

Many machine learning tasks such as multiple instance learning, 3D shape recognition, and few-shot image classification are defined on sets of instances. Since solutions to such problems do not depend on the order of elements of the set, models used to address them should be permutation invariant. We present an attention-based neural network module, the Set Transformer, specifically designed to model interactions among elements in the input set. The model consists of an encoder and a decoder, both of which rely on attention mechanisms. In an effort to reduce computational complexity, we introduce an attention scheme inspired by inducing point methods from sparse Gaussian process literature. It reduces the computation time of self-attention from quadratic to linear in the number of elements in the set. We show that our model is theoretically attractive and we evaluate it on a range of tasks, demonstrating the state-of-the-art performance compared to recent methods for set-structured data.

A Dual Path Network (DPN) is a convolutional neural network which presents a new topology of connection paths internally. The intuition is that ResNets enables feature re-usage while DenseNet enables new feature exploration, and both are important for learning good representations. To enjoy the benefits from both path topologies, Dual Path Networks share common features while maintaining the flexibility to explore new features through dual path architectures. We formulate such a dual path architecture as follows: $$x^{k} = \sum\limits\_{t=1}^{k-1} f\_t^{k}(h^t) \text{,} $$ $$ y^{k} = \sum\limits\_{t=1}^{k-1} v\_t(h^t) = y^{k-1} + \phi^{k-1}(y^{k-1}) \text{,} \\\\ $$ $$ r^{k} = x^{k} + y^{k} \text{,} \\\\ $$ $$ h^k = g^k \left( r^{k} \right) \text{,} $$ where $x^{k}$ and $y^{k}$ denote the extracted information at $k$-th step from individual path, $v_t(\cdot)$ is a feature learning function as $f_t^k(\cdot)$. The first equation refers to the densely connected path that enables exploring new features. The second equation refers to the residual path that enables common features re-usage. The third equation defines the dual path that integrates them and feeds them to the last transformation function in the last equation.

Random Erasing is a data augmentation method for training the convolutional neural network (CNN), which randomly selects a rectangle region in an image and erases its pixels with random values. In this process, training images with various levels of occlusion are generated, which reduces the risk of over-fitting and makes the model robust to occlusion. Random Erasing is parameter learning free, easy to implement, and can be integrated with most of the CNN-based recognition models. Random Erasing is complementary to commonly used data augmentation techniques such as random cropping and flipping, and can be implemented in various vision tasks, such as image classification, object detection, semantic segmentation. In the Albumentations library, there is a generalization of RandomErasing called CoarseDropout, which allows masking an arbitrary number of regions of rectangular shape. It could be applied to images, segmentation masks, and key points. Documentation for CoarseDropout

Compress an ensemble of models into a single one by averaging their weights (under certain pre-conditions).

Inception-ResNet-v2 Reduction-B is an image model block used in the Inception-ResNet-v2 architecture.

Inception-ResNet-v2-B is an image model block for a 17 x 17 grid used in the Inception-ResNet-v2 architecture. It largely follows the idea of Inception modules - and grouped convolutions - but also includes residual connections.

Inception-ResNet-v2-C is an image model block for an 8 x 8 grid used in the Inception-ResNet-v2 architecture. It largely follows the idea of Inception modules - and grouped convolutions - but also includes residual connections.

ProxylessNAS directly learns neural network architectures on the target task and target hardware without any proxy task. Additional contributions include: - Using a new path-level pruning perspective for neural architecture search, showing a close connection between NAS and model compression. Memory consumption is saved by one order of magnitude by using path-level binarization. - Using a novel gradient-based approach (latency regularization loss) for handling hardware objectives (e.g. latency). Given different hardware platforms: CPU/GPU/Mobile, ProxylessNAS enables hardware-aware neural network specialization that’s exactly optimized for the target hardware.

A CoordConv layer is a simple extension to the standard convolutional layer. It has the same functional signature as a convolutional layer, but accomplishes the mapping by first concatenating extra channels to the incoming representation. These channels contain hard-coded coordinates, the most basic version of which is one channel for the $i$ coordinate and one for the $j$ coordinate. The CoordConv layer keeps the properties of few parameters and efficient computation from convolutions, but allows the network to learn to keep or to discard translation invariance as is needed for the task being learned. This is useful for coordinate transform based tasks where regular convolutions can fail.

Decaying Momentum, or Demon, is a stochastic optimizer motivated by decaying the total contribution of a gradient to all future updates. By decaying the momentum parameter, the total contribution of a gradient to all future updates is decayed. A particular gradient term $g\_{t}$ contributes a total of $\eta\sum\_{i}\beta^{i}$ of its "energy" to all future gradient updates, and this results in the geometric sum, $\sum^{\infty}\_{i=1}\beta^{i} = \beta\sum^{\infty}\_{i=0}\beta^{i} = \frac{\beta}{\left(1-\beta\right)}$. Decaying this sum results in the Demon algorithm. Letting $\beta\_{init}$ be the initial $\beta$; then at the current step $t$ with total $T$ steps, the decay routine is given by solving the below for $\beta\_{t}$: $$ \frac{\beta\_{t}}{\left(1-\beta\_{t}\right)} = \left(1-t/T\right)\beta\_{init}/\left(1-\beta\_{init}\right)$$ Where $\left(1-t/T\right)$ refers to the proportion of iterations remaining. Note that Demon typically requires no hyperparameter tuning as it is usually decayed to $0$ or a small negative value at time $T$. Improved performance is observed by delaying the decaying. Demon can be applied to any gradient descent algorithm with a momentum parameter.

Feature Matching is a regularizing objective for a generator in generative adversarial networks that prevents it from overtraining on the current discriminator. Instead of directly maximizing the output of the discriminator, the new objective requires the generator to generate data that matches the statistics of the real data, where we use the discriminator only to specify the statistics that we think are worth matching. Specifically, we train the generator to match the expected value of the features on an intermediate layer of the discriminator. This is a natural choice of statistics for the generator to match, since by training the discriminator we ask it to find those features that are most discriminative of real data versus data generated by the current model. Letting $\mathbf{f}\left(\mathbf{x}\right)$ denote activations on an intermediate layer of the discriminator, our new objective for the generator is defined as: $ ||\mathbb{E}\_{x\sim p\_{data} } \mathbf{f}\left(\mathbf{x}\right) − \mathbb{E}\_{\mathbf{z}∼p\_{\mathbf{z}}\left(\mathbf{z}\right)}\mathbf{f}\left(G\left(\mathbf{z}\right)\right)||^{2}\_{2} $. The discriminator, and hence $\mathbf{f}\left(\mathbf{x}\right)$, are trained as with vanilla GANs. As with regular GAN training, the objective has a fixed point where G exactly matches the distribution of training data.

Progressive Neural Architecture Search, or PNAS, is a method for learning the structure of convolutional neural networks (CNNs). It uses a sequential model-based optimization (SMBO) strategy, where we search the space of cell structures, starting with simple (shallow) models and progressing to complex ones, pruning out unpromising structures as we go. At iteration $b$ of the algorithm, we have a set of $K$ candidate cells (each of size $b$ blocks), which we train and evaluate on a dataset of interest. Since this process is expensive, PNAS also learns a model or surrogate function which can predict the performance of a structure without needing to train it. We then expand the $K$ candidates of size $b$ into $K' \gg K$ children, each of size $b+1$. The surrogate function is used to rank all of the $K'$ children, pick the top $K$, and then train and evaluate them. We continue in this way until $b=B$, which is the maximum number of blocks we want to use in a cell.

Inception v2 is the second generation of Inception convolutional neural network architectures which notably uses batch normalization. Other changes include dropping dropout and removing local response normalization, due to the benefits of batch normalization.

ShuffleNet V2 Downsampling Block is a block for spatial downsampling used in the ShuffleNet V2 architecture. Unlike the regular ShuffleNet V2 block, the channel split operator is removed so the number of output channels is doubled.

Inception-A is an image model block used in the Inception-v4 architecture.

Inception-B is an image model block used in the Inception-v4 architecture.

Inception-C is an image model block used in the Inception-v4 architecture.

Inception-v4 is a convolutional neural network architecture that builds on previous iterations of the Inception family by simplifying the architecture and using more inception modules than Inception-v3.