기법

A Stochastic Dueling Network, or SDN, is an architecture for learning a value function $V$. The SDN learns both $V$ and $Q$ off-policy while maintaining consistency between the two estimates. At each time step it outputs a stochastic estimate of $Q$ and a deterministic estimate of $V$.

Modules used in GAN's style transfer.

Transformer is a type of self-attention-based neural networks originally applied for NLP tasks. Recently, pure transformer-based models are proposed to solve computer vision problems. These visual transformers usually view an image as a sequence of patches while they ignore the intrinsic structure information inside each patch. In this paper, we propose a novel Transformer-iN-Transformer (TNT) model for modeling both patch-level and pixel-level representation. In each TNT block, an outer transformer block is utilized to process patch embeddings, and an inner transformer block extracts local features from pixel embeddings. The pixel-level feature is projected to the space of patch embedding by a linear transformation layer and then added into the patch. By stacking the TNT blocks, we build the TNT model for image recognition. Image source: Han et al.

Weight Standardization is a normalization technique that smooths the loss landscape by standardizing the weights in convolutional layers. Different from the previous normalization methods that focus on activations, WS considers the smoothing effects of weights more than just length-direction decoupling. Theoretically, WS reduces the Lipschitz constants of the loss and the gradients. Hence, WS smooths the loss landscape and improves training. In Weight Standardization, instead of directly optimizing the loss $\mathcal{L}$ on the original weights $\hat{W}$, we reparameterize the weights $\hat{W}$ as a function of $W$, i.e. $\hat{W}=\text{WS}(W)$, and optimize the loss $\mathcal{L}$ on $W$ by SGD: $$ \hat{W} = \Big[ \hat{W}\_{i,j}~\big|~ \hat{W}\_{i,j} = \dfrac{W\_{i,j} - \mu\_{W\_{i,\cdot}}}{\sigma\_{W\_{i,\cdot}+\epsilon}}\Big] $$ $$ y = \hat{W}*x $$ where $$ \mu_{W\_{i,\cdot}} = \dfrac{1}{I}\sum\_{j=1}^{I}W\_{i, j},~~\sigma\_{W\_{i,\cdot}}=\sqrt{\dfrac{1}{I}\sum\_{i=1}^I(W\_{i,j} - \mu\_{W\_{i,\cdot}})^2} $$ Similar to Batch Normalization, WS controls the first and second moments of the weights of each output channel individually in convolutional layers. Note that many initialization methods also initialize the weights in some similar ways. Different from those methods, WS standardizes the weights in a differentiable way which aims to normalize gradients during back-propagation. Note that we do not have any affine transformation on $\hat{W}$. This is because we assume that normalization layers such as BN or GN will normalize this convolutional layer again.

Bi-attention employs the attention-in-attention (AiA) mechanism to capture second-order statistical information: the outer point-wise channel attention vectors are computed from the output of the inner channel attention.

BigGAN-deep is a deeper version (4x) of BigGAN. The main difference is a slightly differently designed residual block. Here the $z$ vector is concatenated with the conditional vector without splitting it into chunks. It is also based on residual blocks with bottlenecks. BigGAN-deep uses a different strategy than BigGAN aimed at preserving identity throughout the skip connections. In G, where the number of channels needs to be reduced, BigGAN-deep simply retains the first group of channels and drop the rest to produce the required number of channels. In D, where the number of channels should be increased, BigGAN-deep passes the input channels unperturbed, and concatenates them with the remaining channels produced by a 1 × 1 convolution. As far as the network configuration is concerned, the discriminator is an exact reflection of the generator. There are two blocks at each resolution (BigGAN uses one), and as a result BigGAN-deep is four times deeper than BigGAN. Despite their increased depth, the BigGAN-deep models have significantly fewer parameters mainly due to the bottleneck structure of their residual blocks.

Bilateral grid is a new data structure that enables fast edge-aware image processing. It enables edge-aware image manipulations such as local tone mapping on high resolution images in real time. Source: Chen et al. Image source: Chen et al.

Our tracker, CenterTrack, applies a detection model to a pair of images and detections from the prior frame. Given this minimal input, CenterTrack localizes objects and predicts their associations with the previous frame. That's it. CenterTrack is simple, online (no peeking into the future), and real-time.

D4PG, or Distributed Distributional DDPG, is a policy gradient algorithm that extends upon the DDPG. The improvements include a distributional updates to the DDPG algorithm, combined with the use of multiple distributed workers all writing into the same replay table. The biggest performance gain of other simpler changes was the use of $N$-step returns. The authors found that the use of prioritized experience replay was less crucial to the overall D4PG algorithm especially on harder problems.

Dialogue-Adaptive Pre-training Objective (DAPO) is a pre-training objective for dialogue adaptation, which is designed to measure qualities of dialogues from multiple important aspects, like Readability, Consistency and Fluency which have already been focused on by general LM pre-training objectives, and those also significant for assessing dialogues but ignored by general LM pre-training objectives, like Diversity and Specificity.

DynamicConv is a type of convolution for sequential modelling where it has kernels that vary over time as a learned function of the individual time steps. It builds upon LightConv and takes the same form but uses a time-step dependent kernel: $$ \text{DynamicConv}\left(X, i, c\right) = \text{LightConv}\left(X, f\left(X\_{i}\right)\_{h,:}, i, c\right) $$

An Eligibility Trace is a memory vector $\textbf{z}\_{t} \in \mathbb{R}^{d}$ that parallels the long-term weight vector $\textbf{w}\_{t} \in \mathbb{R}^{d}$. The idea is that when a component of $\textbf{w}\_{t}$ participates in producing an estimated value, the corresponding component of $\textbf{z}\_{t}$ is bumped up and then begins to fade away. Learning will then occur in that component of $\textbf{w}\_{t}$ if a nonzero TD error occurs before the trade falls back to zero. The trace-decay parameter $\lambda \in \left[0, 1\right]$ determines the rate at which the trace falls. Intuitively, they tackle the credit assignment problem by capturing both a frequency heuristic - states that are visited more often deserve more credit - and a recency heuristic - states that are visited more recently deserve more credit. $$E\_{0}\left(s\right) = 0 $$ $$E\_{t}\left(s\right) = \gamma\lambda{E}\_{t-1}\left(s\right) + \textbf{1}\left(S\_{t} = s\right) $$ Source: Sutton and Barto, Reinforcement Learning, 2nd Edition

FLAVA aims at building a single holistic universal model that targets all modalities at once. FLAVA is a language vision alignment model that learns strong representations from multimodal data (image-text pairs) and unimodal data (unpaired images and text). The model consists of an image encode transformer to capture unimodal image representations, a text encoder transformer to process unimodal text information, and a multimodal encode transformer that takes as input the encoded unimodal image and text and integrates their representations for multimodal reasoning. During pretraining, masked image modeling (MIM) and mask language modeling (MLM) losses are applied onto the image and text encoders over a single image or a text piece, respectively, while contrastive, masked multimodal modeling (MMM), and image-text matching (ITM) loss are used over paired image-text data. For downstream tasks, classification heads are applied on the outputs from the image, text, and multimodal encoders respectively for visual recognition, language understanding, and multimodal reasoning tasks It can be applied to broad scope of tasks from three domains (visual recognition, language understanding, and multimodal reasoning) under a common transformer model architecture.

A Global Context Network, or GCNet, utilises global context blocks to model long-range dependencies in images. It is based on the Non-Local Network, but it modifies the architecture so less computation is required. Global context blocks are applied to multiple layers in a backbone network to construct the GCNet.

GPT-NeoX is an autoregressive transformer decoder model whose architecture largely follows that of GPT-3, with a few notable deviations. The model has 20 billion parameters with 44 layers, a hidden dimension size of 6144, and 64 heads. The main difference with GPT-3 is the change in tokenizer, the addition of Rotary Positional Embeddings, the parallel computation of attention and feed-forward layers, and a different initialization scheme and hyperparameters.

To avoid the problem caused by low-frequent entity-relation pairs, our MBS uses the estimated probabilities from a trained model $\mathbf{\theta}'$ to calculate frequencies for each triplet and query. By using $\mathbf{\theta}'$, the NS loss in KGE with MBS is represented as follows: \begin{align} &\ell_{mbs}(\mathbf{\theta};\mathbf{\theta}') \nonumber \ =&-\frac{1}{|D|}\sum_{(x,y) \in D} \Bigl[A_{mbs}(\mathbf{\theta}')\log(\sigma(s_{\mathbf{\theta}}(x,y)+\gamma))\nonumber\ &+\frac{1}{\nu}sum_{y_{i}\sim p_n(y_{i}|x)}^{\nu}B_{mbs}(\mathbf{\theta}')\log(\sigma(-s_{\mathbf{\theta}}(x,y_i)-\gamma))\Bigr], \end{align}

NAS-FPN is a Feature Pyramid Network that is discovered via Neural Architecture Search in a novel scalable search space covering all cross-scale connections. The discovered architecture consists of a combination of top-down and bottom-up connections to fuse features across scales

Network Dissection is an interpretability method for CNNs that evaluates the alignment between individual hidden units and a set of visual semantic concepts. By identifying the best alignments, units are given human interpretable labels across a range of objects, parts, scenes, textures, materials, and colors. The measurement of interpretability proceeds in three steps: - Identify a broad set of human-labeled visual concepts. - Gather the response of the hidden variables to known concepts. - Quantify alignment of hidden variable−concept pairs.

A Noisy Linear Layer is a linear layer with parametric noise added to the weights. This induced stochasticity can be used in reinforcement learning networks for the agent's policy to aid efficient exploration. The parameters of the noise are learned with gradient descent along with any other remaining network weights. Factorized Gaussian noise is the type of noise usually employed. The noisy linear layer takes the form: $$y = \left(b + Wx\right) + \left(b\_{noisy}\odot\epsilon^{b}+\left(W\_{noisy}\odot\epsilon^{w}\right)x\right) $$ where $\epsilon^{b}$ and $\epsilon^{w}$ are random variables.

Deep Neural Networks (DNNs) are typically trained by backpropagation in a batch learning setting, which requires the entire training data to be made available prior to the learning task. This is not scalable for many real-world scenarios where new data arrives sequentially in a stream form. We aim to address an open challenge of "Online Deep Learning" (ODL) for learning DNNs on the fly in an online setting. Unlike traditional online learning that often optimizes some convex objective function with respect to a shallow model (e.g., a linear/kernel-based hypothesis), ODL is significantly more challenging since the optimization of the DNN objective function is non-convex, and regular backpropagation does not work well in practice, especially for online learning settings.

One-Shot Aggregation is an image model block that is an alternative to Dense Blocks, by aggregating intermediate features. It is proposed as part of the VoVNet architecture. Each convolution layer is connected by two-way connection. One way is connected to the subsequent layer to produce the feature with a larger receptive field while the other way is aggregated only once into the final output feature map. The difference with DenseNet is that the output of each layer is not routed to all subsequent intermediate layers which makes the input size of intermediate layers constant.

ProphetNet is a sequence-to-sequence pre-training model that introduces a novel self-supervised objective named future n-gram prediction and the proposed n-stream self-attention mechanism. Instead of optimizing one-step-ahead prediction in the traditional sequence-to-sequence model, the ProphetNet is optimized by $n$-step ahead prediction that predicts the next $n$ tokens simultaneously based on previous context tokens at each time step. The future n-gram prediction explicitly encourages the model to plan for the future tokens and further help predict multiple future tokens.

In relation-aware global attention (RGA) stresses the importance of global structural information provided by pairwise relations, and uses it to produce attention maps. RGA comes in two forms, spatial RGA (RGA-S) and channel RGA (RGA-C). RGA-S first reshapes the input feature map $X$ to $C\times (H\times W)$ and the pairwise relation matrix $R \in \mathbb{R}^{(H\times W)\times (H\times W)}$ is computed using \begin{align} Q &= \delta(W^QX) \end{align} \begin{align} K &= \delta(W^KX) \end{align} \begin{align} R &= Q^TK \end{align} The relation vector $r_i$ at position $i$ is defined by stacking pairwise relations at all positions: \begin{align} r_i = [R(i, :); R(:,i)] \end{align} and the spatial relation-aware feature $y_i$ can be written as \begin{align} Y_i = [g^c_\text{avg}(\delta(W^\varphi x_i)); \delta(W^\phi r_i)] \end{align} where $g^c_\text{avg}$ denotes global average pooling in the channel domain. Finally, the spatial attention score at position $i$ is given by \begin{align} a_i = \sigma(W_2\delta(W_1y_i)) \end{align} RGA-C has the same form as RGA-S, except for taking the input feature map as a set of $H\times W$-dimensional features. RGA uses global relations to generate the attention score for each feature node, so provides valuable structural information and significantly enhances the representational power. RGA-S and RGA-C are flexible enough to be used in any CNN network; Zhang et al. propose using them jointly in sequence to better capture both spatial and cross-channel relationships.

RepPoints is a representation for object detection that consists of a set of points which indicate the spatial extent of an object and semantically significant local areas. This representation is learned via weak localization supervision from rectangular ground-truth boxes and implicit recognition feedback. Based on the richer RepPoints representation, the authors develop an anchor-free object detector that yields improved performance compared to using bounding boxes.

Self-Attention Network (SANet) proposes two variations of self-attention used for image recognition: 1) pairwise self-attention which generalizes standard dot-product attention and is fundamentally a set operator, and 2) patchwise self-attention which is strictly more powerful than convolution.

The Simple Neural Attention Meta-Learner, or SNAIL, combines the benefits of temporal convolutions and attention to solve meta-learning tasks. They introduce positional dependence through temporal convolutions to make the model applicable to reinforcement tasks - where the observations, actions, and rewards are intrinsically sequential. They also introduce attention in order to provide pinpoint access over an infinitely large context. SNAIL is constructing by combining the two: we use temporal convolutions to produce the context over which we use a causal attention operation.

TrOCR is an end-to-end Transformer-based OCR model for text recognition with pre-trained CV and NLP models. It leverages the Transformer architecture for both image understanding and wordpiece-level text generation. It first resizes the input text image into $384 × 384$ and then the image is split into a sequence of 16 patches which are used as the input to image Transformers. Standard Transformer architecture with the self-attention mechanism is leveraged on both encoder and decoder parts, where wordpiece units are generated as the recognized text from the input image.

Triplet attention comprises of three branches each responsible for capturing crossdimension between the spatial dimensions and channel dimension of the input. Given an input tensor with shape (C × H × W), each branch is responsible for aggregating cross-dimensional interactive features between either the spatial dimension H or W and the channel dimension C.

Wasserstein GAN + Gradient Penalty, or WGAN-GP, is a generative adversarial network that uses the Wasserstein loss formulation plus a gradient norm penalty to achieve Lipschitz continuity. The original WGAN uses weight clipping to achieve 1-Lipschitz functions, but this can lead to undesirable behaviour by creating pathological value surfaces and capacity underuse, as well as gradient explosion/vanishing without careful tuning of the weight clipping parameter $c$. A Gradient Penalty is a soft version of the Lipschitz constraint, which follows from the fact that functions are 1-Lipschitz iff the gradients are of norm at most 1 everywhere. The squared difference from norm 1 is used as the gradient penalty.

AlphaStar is a reinforcement learning agent for tackling the game of Starcraft II. It learns a policy $\pi\_{\theta}\left(a\_{t}\mid{s\_{t}}, z\right) = P\left[a\_{t}\mid{s\_{t}}, z\right]$ using a neural network for parameters $\theta$ that receives observations $s\_{t} = \left(o\_{1:t}, a\_{1:t-1}\right)$ as inputs and chooses actions as outputs. Additionally, the policy conditions on a statistic $z$ that summarizes a strategy sampled from human data such as a build order [1]. AlphaStar uses numerous types of architecture to incorporate different types of features. Observations of player and enemy units are processed with a Transformer. Scatter connections are used to integrate spatial and non-spatial information. The temporal sequence of observations is processed by a core LSTM. Minimap features are extracted with a Residual Network. To manage the combinatorial action space, the agent uses an autoregressive policy and a recurrent pointer network. The agent is trained first with supervised learning from human replays. Parameters are subsequently trained using reinforcement learning that maximizes the win rate against opponents. The RL algorithm is based on a policy-gradient algorithm similar to actor-critic. Updates are performed asynchronously and off-policy. To deal with this, a combination of $TD\left(\lambda\right)$ and V-trace are used, as well as a new self-imitation algorithm (UPGO). Lastly, to address game-theoretic challenges, AlphaStar is trained with league training to try to approximate a fictitious self-play (FSP) setting which avoids cycles by computing a best response against a uniform mixture of all previous policies. The league of potential opponents includes a diverse range of agents, including policies from current and previous agents. Image Credit: Yekun Chai #### References 1. Chai, Yekun. "Deciphering AlphaStar on StarCraft II." (2019). https://cyk1337.github.io/notes/2019/07/21/RL/DRL/Decipher-AlphaStar-on-StarCraft-II/ #### Code Implementation 1. https://github.com/opendilab/DI-star

Ape-X is a distributed architecture for deep reinforcement learning. The algorithm decouples acting from learning: the actors interact with their own instances of the environment by selecting actions according to a shared neural network, and accumulate the resulting experience in a shared experience replay memory; the learner replays samples of experience and updates the neural network. The architecture relies on prioritized experience replay to focus only on the most significant data generated by the actors. In contrast to Gorila, Ape-X uses a shared, centralized replay memory, and instead of sampling uniformly, it prioritizes, to sample the most useful data more often. All communications are batched with the centralized replay, increasing the efficiency and throughput at the cost of some latency. And by learning off-policy, Ape-X has the ability to combine data from many distributed actors, by giving the different actors different exploration policies, broadening the diversity of the experience they jointly encounter.

A Bottleneck Transformer Block is a block used in Bottleneck Transformers that replaces the spatial 3 × 3 convolution layer in a Residual Block with Multi-Head Self-Attention (MHSA).

Domain Adaptive Neighborhood Clustering via Entropy Optimization (DANCE) is a self-supervised clustering method that harnesses the cluster structure of the target domain using self-supervision. This is done with a neighborhood clustering technique that self-supervises feature learning in the target. At the same time, useful source features and class boundaries are preserved and adapted with a partial domain alignment loss that the authors refer to as entropy separation loss. This loss allows the model to either match each target example with the source, or reject it as unknown.

In the field of scene segmentation, encoder-decoder structures cannot make use of the global relationships between objects, whereas RNN-based structures heavily rely on the output of the long-term memorization. To address the above problems, Fu et al. proposed a novel framework, the dual attention network (DANet), for natural scene image segmentation. Unlike CBAM and BAM, it adopts a self-attention mechanism instead of simply stacking convolutions to compute the spatial attention map, which enables the network to capture global information directly. DANet uses in parallel a position attention module and a channel attention module to capture feature dependencies in spatial and channel domains. Given the input feature map $X$, convolution layers are applied first in the position attention module to obtain new feature maps. Then the position attention module selectively aggregates the features at each position using a weighted sum of features at all positions, where the weights are determined by feature similarity between corresponding pairs of positions. The channel attention module has a similar form except for dimensional reduction to model cross-channel relations. Finally the outputs from the two branches are fused to obtain final feature representations. For simplicity, we reshape the feature map $X$ to $C\times (H \times W)$ whereupon the overall process can be written as \begin{align} Q,\quad K,\quad V &= W_qX,\quad W_kX,\quad W_vX \end{align} \begin{align} Y^\text{pos} &= X+ V\text{Softmax}(Q^TK) \end{align} \begin{align} Y^\text{chn} &= X+ \text{Softmax}(XX^T)X \end{align} \begin{align} Y &= Y^\text{pos} + Y^\text{chn} \end{align} where $W_q$, $W_k$, $W_v \in \mathbb{R}^{C\times C}$ are used to generate new feature maps. The position attention module enables DANet to capture long-range contextual information and adaptively integrate similar features at any scale from a global viewpoint, while the channel attention module is responsible for enhancing useful channels as well as suppressing noise. Taking spatial and channel relationships into consideration explicitly improves the feature representation for scene segmentation. However, it is computationally costly, especially for large input feature maps.

Deep Graph Infomax (DGI), a general approach for learning node representations within graph-structured data in an unsupervised manner. DGI relies on maximizing mutual information between patch representations and corresponding high-level summaries of graphs—both derived using established graph convolutional network architectures. The learnt patch representations summarize subgraphs centered around nodes of interest, and can thus be reused for downstream node-wise learning tasks. In contrast to most prior approaches to unsupervised learning with GCNs, DGI does not rely on random walk objectives, and is readily applicable to both transductive and inductive learning setups. Description and image from: DEEP GRAPH INFOMAX