**CSPDenseNet** is a convolutional neural network and object detection backbone where we apply the Cross Stage Partial Network (CSPNet) approach to [DenseNet](https://paperswithcode.com/method/densenet). The CSPNet partitions the feature map of the base layer into two parts and then merges them through a cross-stage hierarchy. The use of a split and merge strategy allows for more gradient flow through the network.

CSPDenseNet

CSPNet: A New Backbone that can Enhance Learning Capability of CNN

**DFDNet**, or **DFDNet**, is a deep face dictionary network for face restoration to guide the restoration process of degraded observations. Given a LQ image $I\_{d}$, the DFDNet selects the dictionary features that have the most similar structure with the input. Specially, we re-norm the whole dictionaries via component AdaIN (termed as CAdaIN) based on the input component to eliminate the distribution or style diversity. The selected dictionary features are then utilized to guide the restoration process via dictionary feature transformation.

DFDNet

Blind Face Restoration via Deep Multi-scale Component Dictionaries

**ScheduledDropPath** is a modified version of [DropPath](https://paperswithcode.com/method/droppath). In DropPath, each path in the cell is stochastically dropped with some fixed probability during training. In ScheduledDropPath, each path in the cell is dropped out with a probability that is linearly increased over the course of training.

ScheduledDropPath

Learning Transferable Architectures for Scalable Image Recognition

The **Routing Transformer** is a [Transformer](https://paperswithcode.com/method/transformer) that endows self-attention with a sparse routing module based on online k-means. Each attention module considers a clustering of the space: the current timestep only attends to context belonging to the same cluster. In other word, the current time-step query is routed to a limited number of context through its cluster assignment.

Routing Transformer

Efficient Content-Based Sparse Attention with Routing Transformers

This method proposes first discretizing observations and calculating the action distribution distance under comparable cases (intersection states).

Playstyle Distance

An Unsupervised Video Game Playstyle Metric via State Discretization

**Sarsa** is an on-policy TD control algorithm:

$$Q\left(S\_{t}, A\_{t}\right) \leftarrow Q\left(S\_{t}, A\_{t}\right) + \alpha\left[R_{t+1} + \gamma{Q}\left(S\_{t+1}, A\_{t+1}\right) - Q\left(S\_{t}, A\_{t}\right)\right] $$

This update is done after every transition from a nonterminal state $S\_{t}$. if $S\_{t+1}$ is terminal, then $Q\left(S\_{t+1}, A\_{t+1}\right)$ is defined as zero.

To design an on-policy control algorithm using Sarsa, we estimate $q\_{\pi}$ for a behaviour policy $\pi$ and then change $\pi$ towards greediness with respect to $q\_{\pi}$.

Source: Sutton and Barto, Reinforcement Learning, 2nd Edition

Sarsa

**Retrace** is an off-policy Q-value estimation algorithm which has guaranteed convergence for a target and behaviour policy $\left(\pi, \beta\right)$. With off-policy rollout for TD learning, we must use importance sampling for the update:

$$ \Delta{Q}^{\text{imp}}\left(S\_{t}, A\_{t}\right) = \gamma^{t}\prod\_{1\leq{\tau}\leq{t}}\frac{\pi\left(A\_{\tau}\mid{S\_{\tau}}\right)}{\beta\left(A\_{\tau}\mid{S\_{\tau}}\right)}\delta\_{t} $$

This product term can lead to high variance, so Retrace modifies $\Delta{Q}$ to have importance weights truncated by no more than a constant $c$:

$$ \Delta{Q}^{\text{imp}}\left(S\_{t}, A\_{t}\right) = \gamma^{t}\prod\_{1\leq{\tau}\leq{t}}\min\left(c, \frac{\pi\left(A\_{\tau}\mid{S\_{\tau}}\right)}{\beta\left(A\_{\tau}\mid{S\_{\tau}}\right)}\right)\delta\_{t} $$

Retrace

Safe and Efficient Off-Policy Reinforcement Learning

Context-Aware Visual Attention-based end-to-end pipeline for Webpage Object Detection (_CoVA_) aims to learn function _f_ to predict labels _y = [$y_1, y_2, ..., y_N$]_ for a webpage containing _N_ elements. The input to CoVA consists of:
1. a screenshot of a webpage,
2. list of bounding boxes _[x, y, w, h]_ of the web elements, and
3. neighborhood information for each element obtained from the DOM tree.

This information is processed in four stages:
1. the graph representation extraction for the webpage,
2. the Representation Network (_RN_),
3. the Graph Attention Network (_GAT_), and
4. a fully connected (_FC_) layer.

The graph representation extraction computes for every web element _i_ its set of _K_ neighboring web elements _$N_i$_. The _RN_ consists of a Convolutional Neural Net (_CNN_) and a positional encoder aimed to learn a visual representation _$v_i$_ for each web element _i &isin; {1, ..., N}_. The _GAT_ combines the visual representation _$v_i$_ of the web element _i_ to be classified and those of its neighbors, i.e., _$v_k$ &forall;k &isin; $N_i$_ to compute the contextual representation _$c_i$_ for web element _i_. Finally, the visual and contextual representations of the web element are concatenated and passed through the _FC_ layer to obtain the classification output.

CoVA

CoVA: Context-aware Visual Attention for Webpage Information Extraction

A **Dimension-wise Convolution**, or **DimConv**, is a type of [convolution](https://paperswithcode.com/method/convolution) that can encode depth-wise, width-wise, and height-wise information independently. To achieve this, DimConv extends depthwise convolutions to all dimensions of the input tensor $X \in \mathbb{R}^{D\times{H}\times{W}}$, where $W$, $H$, and $D$ corresponds to width, height, and depth of $X$. DimConv has three branches, one branch per dimension. These branches apply $D$ depth-wise convolutional kernels $k\_{D} \in \mathbb{R}^{1\times{n}\times{n}}$ along depth, $W$ width-wise convolutional kernels $k\_{W} \in \mathbb{R}^{n\times{1}\times{1}}$ along width, and $H$ height-wise convolutional kernels $k\_{H} \in \mathbb{R}^{n\times{1}\times{n}}$ kernels along height
to produce outputs $Y\_{D}$, $Y\_{W}$, and $Y\_{H} \in \mathbb{R}^{D\times{H}\times{W}}$ that
encode information from all dimensions of the input tensor. The outputs of these independent branches are concatenated along the depth dimension, such that the first spatial plane of $Y\_{D}$, $Y\_{W}$, and $Y\_{H}$ are put together and so on, to produce the output $Y\_{Dim} = ${$Y\_{D}$, $Y\_{W}$, $Y\_{H}$} $\in \mathbb{R}^{3D\times{H}\times{W}}$.

DimConv

DiCENet: Dimension-wise Convolutions for Efficient Networks

**Synchronized Batch Normalization (SyncBN)** is a type of [batch normalization](https://paperswithcode.com/method/batch-normalization) used for multi-GPU training. Standard batch normalization only normalizes the data within each device (GPU). SyncBN normalizes the input within the whole mini-batch.

SyncBN

Context Encoding for Semantic Segmentation

**Libra R-CNN** is an object detection model that seeks to achieve a balanced training procedure. The authors motivation is that training in past detectors has suffered from imbalance during the training process, which generally consists in three levels – sample level, feature level, and objective level. To mitigate the adverse effects, Libra R-CNN integrates three novel components: IoU-balanced
sampling, [balanced feature pyramid](https://paperswithcode.com/method/balanced-feature-pyramid), and [balanced L1 loss](https://paperswithcode.com/method/balanced-l1-loss), respectively for reducing the imbalance at sample, feature, and objective level.

Libra R-CNN

Libra R-CNN: Towards Balanced Learning for Object Detection

[HHO](https://aliasgharheidari.com/HHO.html) is a popular swarm-based, gradient-free optimization algorithm with several active and time-varying phases of exploration and exploitation. This algorithm initially published by the prestigious Journal of Future Generation Computer Systems (FGCS) in 2019, and from the first day, it has gained increasing attention among researchers due to its flexible structure, high performance, and high-quality results. The main logic of the HHO method is designed based on the cooperative behaviour and chasing styles of Harris' hawks in nature called "surprise pounce". Currently, there are many suggestions about how to enhance the functionality of HHO, and there are also several enhanced variants of the HHO in the leading Elsevier and IEEE transaction journals.

From the algorithmic behaviour viewpoint, there are several effective features in HHO :
Escaping energy parameter has a dynamic randomized time-varying nature, which can further improve and harmonize the exploratory and exploitive patterns of HHO. This factor also supports HHO to conduct a smooth transition between exploration and exploitation.
Different exploration mechanisms with respect to the average location of hawks can increase the exploratory trends of HHO throughout initial iterations.
Diverse LF-based patterns with short-length jumps enrich the exploitative behaviours of HHO when directing a local search.
The progressive selection scheme supports search agents to progressively advance their position and only select a better position, which can improve the superiority of solutions and intensification powers of HHO throughout the optimization procedure.
HHO shows a series of searching strategies and then, it selects the best movement step. This feature has also a constructive influence on the exploitation inclinations of HHO.
The randomized jump strength can assist candidate solutions in harmonising the exploration and exploitation leanings.
The application of adaptive and time-varying components allows HHO to handle difficulties of a feature space including local optimal solutions, multi-modality, and deceptive optima.

🔗 The source codes of HHO are publicly available at https://aliasgharheidari.com/HHO.html

Harris Hawks optimization (HHO)

The **Normalized Linear Combination of Activations**, or **NormLinComb**, is a type of activation function that has trainable parameters and uses the normalized linear combination of other activation functions.

$$NormLinComb(x) = \frac{\sum\limits_{i=0}^{n} w_i \mathcal{F}_i(x)}{\mid \mid W \mid \mid}$$

NormLinComb

Trainable Activations for Image Classification

**AdaSqrt** is a stochastic optimization technique that is motivated by the observation that methods like [Adagrad](https://paperswithcode.com/method/adagrad) and [Adam](https://paperswithcode.com/method/adam) can be viewed as relaxations of [Natural Gradient Descent](https://paperswithcode.com/method/natural-gradient-descent).

The updates are performed as follows:

$$ t \leftarrow t + 1 $$

$$ \alpha\_{t} \leftarrow \sqrt{t} $$

$$ g\_{t} \leftarrow \nabla\_{\theta}f\left(\theta\_{t-1}\right) $$

$$ S\_{t} \leftarrow S\_{t-1} + g\_{t}^{2} $$

$$ \theta\_{t+1} \leftarrow \theta\_{t} + \eta\frac{\alpha\_{t}g\_{t}}{S\_{t} + \epsilon} $$

AdaSqrt

Second-order Information in First-order Optimization Methods

**IoU-Balanced Sampling** is hard mining method for object detection. Suppose we need to sample $N$ negative samples from $M$ corresponding candidates. The selected probability for each sample under random sampling is:

$$ p = \frac{N}{M} $$

To raise the selected probability of hard negatives, we evenly split the sampling interval into $K$ bins according to IoU. $N$ demanded negative samples are equally distributed to each bin. Then we select samples from them uniformly. Therefore, we get the selected probability under IoU-balanced sampling:

$$ p\_{k} = \frac{N}{K}*\frac{1}{M\_{k}}\text{ , } k\in\left[0, K\right)$$

where $M\_{k}$ is the number of sampling candidates in the corresponding interval denoted by $k$. $K$ is set to 3 by default in our experiments.

The sampled histogram with IoU-balanced sampling is shown by green color in the Figure to the right. The IoU-balanced sampling can guide the distribution of training samples close to the one of hard negatives.

IoU-Balanced Sampling

The **NVAE Generative Residual Cell** is a skip connection block used as part of the [NVAE](https://paperswithcode.com/method/nvae) architecture for the generator. The residual cell expands the number of channels $E$ times before applying the [depthwise separable convolution](https://paperswithcode.com/method/depthwise-separable-convolution), and then maps it back to $C$ channels. The design motivation was to help model long-range correlations in the data by increasing the receptive field of the network, which explains the expanding path but also the use of depthwise convolutions to keep a handle on parameter count.

NVAE Generative Residual Cell

NVAE: A Deep Hierarchical Variational Autoencoder

**IFNet** is an architecture for video frame interpolation that adopts a coarse-to-fine strategy with progressively increased resolutions: it iteratively updates intermediate flows and soft fusion mask via successive [IFBlocks](https://paperswithcode.com/method/ifblock). Conceptually, according to the iteratively updated flow fields, we can move corresponding pixels from two input frames to the same location in a latent intermediate frame and use a fusion mask to combine pixels from two input frames. Unlike most previous optical flow models, IFBlocks do not contain expensive operators like cost volume or forward warping and use 3 × 3 [convolution](https://paperswithcode.com/method/convolution) and deconvolution as building blocks.

IFNet

RIFE: Real-Time Intermediate Flow Estimation for Video Frame Interpolation

**FoveaBox** is anchor-free framework for object detection. Instead of using predefined anchors to enumerate possible locations, scales and aspect ratios for the search of the objects, FoveaBox directly learns the object existing possibility and the bounding box coordinates without anchor reference. This is achieved by: (a) predicting category-sensitive semantic maps for the object existing possibility, and (b) producing category-agnostic bounding box for each position that potentially contains an object. The scales of target boxes are naturally associated with feature pyramid representations for each input image

It is a single, unified network composed of a backbone network and two task-specific subnetworks. The backbone is responsible for computing a convolutional feature map over an entire input image and is an off-the-shelf convolutional network. The first subnet performs per pixel classification on the backbone’s output; the second subnet performs bounding box prediction for the corresponding
position.

Viet-Anh on Software

AI Glossary

Tech Terminology

CSPDenseNet

DFDNet

ScheduledDropPath

Routing Transformer

Playstyle Distance

Sarsa

Retrace

Context-aware Visual Attention-based (CoVA) webpage object detection pipeline

Dimension-wise Convolution

Synchronized Batch Normalization

Libra R-CNN

Harris Hawks optimization

Normalized Linear Combination of Activations

AdaSqrt

IoU-Balanced Sampling

NVAE Generative Residual Cell

IFNet

FoveaBox

Viet-Anh on Software