[ { "original_id": "030cf20c-f0d2-4fb7-a22f-4ddb1e9417c8", "query_id": 20, "date": "2021-07-07T16:39:28Z", "year": "2021", "category": "cs.LG", "content": "~\\\\\nExternal uncertainty quantification approaches do not affect the models' predictions, since the evaluation of the uncertainty is separated from the underlying prediction task. Furthermore, several external approaches can be applied to already trained networks at the same time without affecting each other. Raghu et al. argued that when both tasks, the prediction and the uncertainty quantification, are done by one single method, the uncertainty estimation is biased by the actual prediction task. Therefore, they recommended a \"direct uncertainty prediction\" and suggested to train two neural networks, one for the actual prediction task and a second one for the prediction of the uncertainty on the first network's predictions. Similarly, Ramalho and Miranda introduced an additional neural network for uncertainty estimation. But in contrast to , the representation space of the training data is considered and the density around a given test sample is evaluated. The additional neural network uses this training data density in order to predict whether the main network's estimate is expected to be correct or false. Hsu et al. detected out-of-distribution examples in classification tasks at test-time by predicting total probabilities for each class, additional to the categorical distribution given by the softmax output. The class wise total probability is predicted by applying the sigmoid function to the network's logits. Based on these total probabilities, OOD examples can be identified as those with low class probabilities for all classes. \\\\\nIn contrast to this, Oberdiek et al. took the sensitivity of the model, i.e. the model's slope, into account by using gradient metrics for the uncertainty quantification in classification tasks. Lee et al. applied a similar idea but made use of back propagated gradients. In their work they presented state of the art results on out-of-distribution and corrupted input detection.", "prefix_titles": [ [ "title", "A Survey of Uncertainty in Deep Neural Networks" ], [ "section", "Uncertainty Estimation" ], [ "subsection", "Single Deterministic Methods" ], [ "subsubsection", "External Uncertainty Quantification Approaches" ] ], "prefix_titles_query": "What are the current external uncertainty quantification approaches for single deterministic methods in deep neural networks?", "cites": [ 4617, 4592, 7130, 4620, 4621 ], "cite_extract_rate": 1.0, "origin_cites_number": 5 }, { "original_id": "07c7f2da-ef2c-475e-823a-a86caf9c3b67", "query_id": 31, "date": "2021-07-07T16:39:28Z", "year": "2021", "category": "cs.LG", "content": "\\label{sec:uncertainty_measures}\nIn Section \\ref{sec:uncertainty_quantification_methods}, we presented different methods for modeling and predicting different types of uncertainty in neural networks. In order to evaluate these approaches, measures have to be applied on the derived uncertainties. In the following, we present different measures for quantifying the different predicted types of uncertainty.\nIn general, the correctness and trustworthiness of these uncertainties is not automatically given. In fact, there are several reasons why evaluating the quality of the uncertainty estimates is a challenging task.\n\\begin{itemize}\n \\setlength\\itemsep{0.5em}\n \\item First, the quality of the uncertainty estimation depends on the underlying method for estimating uncertainty. This is exemplified in the work undertaken by Yao et al. , which shows that different approximates of Bayesian inference (e.g. Gaussian and Laplace approximates) result in different qualities of uncertainty estimates.\n \\item Second, there is a lack of ground truth uncertainty estimates and defining ground truth uncertainty estimates is challenging. For instance, if we define the ground truth uncertainty as the uncertainty across human subjects, we still have to answer questions as \"How many subjects do we need?\" or \"How to choose the subjects?\".\n \\item Third, there is a lack of a unified quantitative evaluation metric . To be more specific, the uncertainty is defined differently in different machine learning tasks such as classification, segmentation, and regression. For instance, prediction intervals or standard deviations are used to represent uncertainty in regression tasks, while entropy (and other related measures) are used to capture uncertainty in classification and segmentation tasks.\n\\end{itemize}", "prefix_titles": [ [ "title", "A Survey of Uncertainty in Deep Neural Networks" ], [ "section", "Uncertainty Measures and Quality" ] ], "prefix_titles_query": "What are the different measures of uncertainty in deep neural networks and how do they impact the quality of the models?", "cites": [ 4684, 4685, 3288 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "09365f96-9cba-420c-afc8-a3c8d6b4d1b6", "query_id": 32, "date": "2021-07-07T16:39:28Z", "year": "2021", "category": "cs.LG", "content": "For classification tasks, the network's softmax output already represents a measure of confidence. But since the raw softmax output is neither very reliable nor can it represent all sources of uncertainty , further approaches and corresponding measures were developed.", "prefix_titles": [ [ "title", "A Survey of Uncertainty in Deep Neural Networks" ], [ "section", "Uncertainty Measures and Quality" ], [ "subsection", "Evaluating Uncertainty in Classification Tasks" ] ], "prefix_titles_query": "What are the methods for evaluating uncertainty in classification tasks within deep neural networks and how do these methods impact the quality of the model's predictions?", "cites": [ 4593, 1624 ], "cite_extract_rate": 1.0, "origin_cites_number": 2 }, { "original_id": "d193f79c-349d-447b-816c-8fa8897f56dd", "query_id": 37, "date": "2021-07-07T16:39:28Z", "year": "2021", "category": "cs.LG", "content": "As already discussed above, removing the model uncertainty and receiving an accurate estimation of the data uncertainty leads to a well calibrated predictor. Following several works based on deep ensembles and BNNs, also compared their performance to other methods based on the resulting calibration. Lakshminarayanan et al. and Mehrtash et al. reported an improved calibration by applying deep ensembles compared to single networks. However, Rahaman and Thiery showed that for specific configurations as the usage of mixup-regularization, deep ensembles can even increase the calibration error. On the other side they showed that applying temperature scaling on the averaged predictions can give a significant improvement on the calibration. \\\\\nFor the Bayesian approaches, showed that restricting the Bayesian approximation to the weights of the last fully connected layer of a DNN is already enough to improve the calibration significantly. Zhang et al. and Laves et al. showed that confidence estimates computed with MC dropout can be poorly calibrated. To overcome this, Zhang et al. proposed structured dropout, which consists of dropping channel, blocks or layers, to promote model diversity and reduce calibration errors. \\\\", "prefix_titles": [ [ "title", "A Survey of Uncertainty in Deep Neural Networks" ], [ "section", "Calibration" ], [ "subsection", "Calibration Methods" ], [ "subsubsection", "Calibration with Uncertainty Estimation Approaches" ] ], "prefix_titles_query": "What are the recent advancements in calibration methods for deep neural networks focusing on uncertainty estimation approaches?", "cites": [ 3300, 8832, 4658, 4692, 4676, 4695, 4694, 3288 ], "cite_extract_rate": 1.0, "origin_cites_number": 8 }, { "original_id": "a96c9b0e-9d71-4bee-bf0c-0428d07e4d1b", "query_id": 132, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "The Transformer~ is an innovative architecture that facilitates the transfer of weighted representation knowledge between various neural units. It relies solely on attention mechanisms and doesn't use recurrent or convolutional architectures. The attention mechanism is a crucial component of the Transformer as it assigns weights to all the encoded input representations and learns the most important part of the input data. The output of the attention is obtained by taking the weighted sum of the values, and the weights are calculated using the compatibility function of the query with the corresponding key~. Numerous attention mechanisms~ have been developed in large models. For instance, in natural language processing, self-attention is created to connect various positions in a single sequence for generating a representation of the same sequence. Transformer leverages a mask matrix to provide an attention mechanism based on self-attention, in which the mask matrix specifies which words can ``see'' each other.\nTransformer is an important structure for PFMs in NLP, CV, and GL areas. For NLP, the Transformer can help solve the long-range dependency issues when processing sequential input data. For example, the GPT-3~ is a generative model based on the transformer. For CV, the Vision Transformer (ViT)~ is proposed to represent an image to a series of image patches, which is similar to a series of word embeddings. For GL, the Graph Transformer Networks (GTN)~ are employed to learn new graph structures and powerful node representations without domain knowledge. Transformers become scalable enough to drive ground-breaking capabilities for PFMs thanks to the transformer structures to achieve higher parallelization. The ViT-22B model~, for instance, has about 22B parameters, and the largest language models can have upwards of 100B parameters (e.g., GPT-3 has 175B and PaLM~ has 540B parameters).", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "Basic Components" ], [ "subsection", "Transformer for PFMs" ] ], "prefix_titles_query": "What are the fundamental components and roles of Transformer architectures in the development and functioning of Pretrained Foundation Models from BERT to ChatGPT?", "cites": [ 679, 1183, 8558, 732, 1554, 7095, 38 ], "cite_extract_rate": 1.0, "origin_cites_number": 7 }, { "original_id": "8b882e93-190e-4264-b9b4-fa6f2a75f940", "query_id": 134, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "There are many pretraining tasks created for CV to learn the feature space, which is based on SSL. It utilizes pretext tasks that contain human-designed labels, like jigsaw puzzles or the comparison of various patches from images. This enables the generalization of learned representations to a range of downstream tasks.\n\\textbf{Specific Pretext Task.}\nA pretext task also referred to as a predefined task, is created for the encoder networks to perform during the pretraining phase. The network is trained by predicting the answer to a special pretext task. Based on particular features of the data, pseudo labels are generated for the fictitious task. Then, using guided learning techniques, the encoder networks are trained to solve the pretext task. For example, inpainting aims to pretrain models by predicting the missed center part.\n\\textbf{Frame Order Learning Task.} Learning frame order from videos involves frame processing through time steps, which can serve as the pretraining task for CV. This issue usually relates to completing pretextual exercises that can aid in the acquisition of visual temporal representations.\n\\textbf{Data Generation Task.} The representational capabilities within the generative adversarial networks (GANs) can also be used in the pretraining tasks. Projecting data back into the latent space, as demonstrated by BiGANs~, is helpful for auxiliary supervised discrimination tasks by acting as feature representations.\n\\textbf{Data Reconstruction Task.} Since the images can be divided into patches inspired by the natural language, some pretraining tasks for NLP can also be used in CV, e.g., the autoencoder-based masked prediction. The original image is first divided into a few patches and discrete visual tokens are used to encode each patch. The visual tokens from the masked patches are outputted in the second stage to match the corresponding visual tokens from the fixed tokenizer.\n\\textbf{Miscellaneous.} To train the PFMs in CV, additional pretraining tasks are suggested. For instance, based on contrastive learning, encoder networks are used for pretraining on various data augmentation. The parameters are trained by maximizing the distance between negative pairs (e.g., pairs with different labels) and minimizing the distance between positive pairs (e.g., pairs with the same labels). To pretrain the parameters of the backbone network, the DeepClustering~ method divides the representations into various clusters and labels these clusters as supervised signals.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "Basic Components" ], [ "subsection", "Pretraining Tasks for PFMs" ], [ "subsubsection", "Pretraining Tasks for CV" ] ], "prefix_titles_query": "What are the various pretraining tasks used for Pretrained Foundation Models (PFMs) in computer vision (CV)?", "cites": [ 9095, 9094 ], "cite_extract_rate": 1.0, "origin_cites_number": 2 }, { "original_id": "d5594cf1-ec15-4e99-82d4-c0aff6c92c8b", "query_id": 136, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "Most of the popular pretraining models need lots of pretraining data, which imposes huge requirements on the hardware, making it challenging to retrain, and only fine-tuning can be done to the model.\nTo solve these problems, some models appear. For example, ERNIE Tiny released by Baidu is a miniaturized ERNIE~, that reduces the number of layers and increases the prediction speed by 4.3 times with a slight decrease in accuracy.\nLan et al. propose the ALBERT~ to reduce memory consumption and training speed.\nHowever, it is undeniable that no matter what kind of compression is done for these large-scale models, the performance of the models in these tasks will deteriorate sharply. It requires paying attention to the efficient representation of high-level semantic and grammatical information and lossless compression in future works.\nBy using word-embedded parameter factorization and hidden parameter sharing between layers, ALBERT significantly reduces the number of parameters of the model without performance loss. It proposes the training task of SOP, which predicts the order of the two sentences to improve the performance.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Natural Language Processing" ], [ "subsection", "Boosting Methods" ], [ "paragraph", "Boosting on Model Performance" ] ], "prefix_titles_query": "What are the various boosting methods used to enhance the performance of pretrained foundation models in natural language processing, particularly from BERT to ChatGPT?", "cites": [ 1150, 2475 ], "cite_extract_rate": 1.0, "origin_cites_number": 2 }, { "original_id": "0411fac8-d211-46f3-867e-f60b611650a8", "query_id": 137, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "ERNIE(Baidu)~ is mainly composed of two parts, the Transformer encoder and task embedding. In the Transformer encoder, the self-attention mechanism is used to capture the context information of each token and generate context representation embedding. \nTask embedding is a technique that applies different characteristics to a task.\nERNIE 2.0~ introduces multi-task learning to realize the pretraining of lexical, grammar, and semantics.\nERNIE 2.0 uses seven different pretraining tasks, covering three aspects: word level, sentence level, and semantic level. It uses continual learning, making the knowledge in the previous training task retained and enabling the model to acquire long-distance memory.\nIt uses a Transformer encoder and introduces task embedding, enabling the model to distinguish different tasks in the continual learning process.\nUniLM~ uses three pretraining tasks: unidirectional LM, bidirectional LM, and encoder-decoder LM. \nIt can simultaneously complete three kinds of target tasks in the pretraining stage through the self-attention layer mask mechanism.\nIn the training stage, UniLM adopts the small-segment mask strategy proposed by SpanBERT, and the loss function is composed of the loss functions of the above three pretraining tasks. To maintain the contribution consistency on all loss functions, the three pretraining tasks are trained simultaneously.\nModeling and parameter sharing of multiple tasks make LMs achieve good generalization ability in Natural Language Understanding (NLU) and NLG tasks.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Natural Language Processing" ], [ "subsection", "Boosting Methods" ], [ "paragraph", "Boosting on Model Performance" ], [ "paragraph", "Boosting for Multi-task Learning" ] ], "prefix_titles_query": "What are the recent advancements in boosting methods for improving model performance and multi-task learning in pretrained foundation models for natural language processing?", "cites": [ 2476, 50, 2475 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "f0b30ce0-79ed-4853-b7a2-3527355f4a4c", "query_id": 140, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "Chain-of-thought (CoT) prompting is a technique for improving the reasoning ability of LLMs by prompting them to generate a series of intermediate steps that lead to the final answer of a multi-step problem. The CoT is a series of intermediate reasoning steps, which can significantly improve the ability of LLMs to perform complex reasoning . Besides, fine-tuning with CoT shows slightly more harmless compared to without CoT . CoT prompting is an emergent property of model scale, meaning it works better with larger and more powerful language models. It is also possible to fine-tune models on CoT reasoning datasets to enhance this capability further and stimulate better interpretability.\nIn a CoT prompting experiment, a prompt is provided to the model that outlines a multi-step problem. The prompt might pose a question such as ``After selling 30 out of his 100 chickens and 10 out of his 20 pigs, how many animals does a farmer have left?'' The model then generates a sequence of intermediate reasoning steps, for example, ``The farmer has 100-30=70 chickens remaining'' and ``The farmer has 20-10=10 pigs remaining,'' before generating the final answer, such as ``The farmer has 70+10=80 animals remaining.'' CoT prompting has demonstrated its efficacy in improving the performance of LLMs on various reasoning tasks, such as arithmetic, symbolic reasoning, and common sense. It is a promising technique that can enhance the ability of language models to reason about complicated problems.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Natural Language Processing" ], [ "subsection", "Instruction-Aligning Methods" ], [ "paragraph", "Supervised Fine-Tuning (SFT)" ], [ "paragraph", "Chain-of-Thoughts" ] ], "prefix_titles_query": "What are the instruction-aligning methods used for pretrained foundation models in natural language processing, specifically focusing on supervised fine-tuning (SFT) and chain-of-thoughts approaches?", "cites": [ 7468, 8472, 1578 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "3bd63707-4357-48a2-914a-922ccb66761b", "query_id": 142, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "\\label{sec:Learning by Generation}\nAlthough many existing applications are popular after the development of the GAN-based approach, the representation abilities inside the GANs are not entirely exploited due to the absence of a feature encoder. \nThus, Bidirectional Generative Adversarial Networks (BiGANs)~ is proposed to project data back into the latent space, which is useful for auxiliary supervised discrimination tasks via serving as feature representations. \nBased on BiGANs, BigBiGAN~ first achieves the SOTA in unsupervised representation learning on ImageNet by adding an encoder and modifying the discriminator. As shown in Fig. \\ref{fig:bigbigan} from~, the traditional components of GANs (encoder $\\mathcal{E}$ and generator $\\mathcal{G}$) are used to produce data-latent pairs, denoted as $(\\textbf{x}\\sim P_{\\textbf{x}},\\hat{\\textbf{z}}\\sim\\mathcal{E}(\\textbf{x}))$ and $(\\hat{\\textbf{x}}\\sim\\mathcal{G}(\\textbf{z}),\\textbf{z}\\sim P_{\\textbf{z}})$. The final loss $\\ell$ is defined as the sum of data-specific term $s_{\\textbf{x}},s_{\\textbf{z}}$ and data-joint term $s_{\\textbf{xz}}$. The introduced discriminator $\\mathcal{D}$ (Adversarially Learned Inference (ALI)~, or BiGAN~) learns to discriminate between pairs from the raw data, latent distribution and encoded vector. \n\\begin{figure*}[t]\n \\centering\n \\includegraphics[width=0.8\\linewidth]{figures/BigBiGAN_zip.pdf}\n \\caption{The structure of the BigBiGAN framework~.}\n \\label{fig:bigbigan}\n\\end{figure*}", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Computer Vision" ], [ "subsection", "Learning by Generation" ] ], "prefix_titles_query": "What are the advancements and applications of pretrained foundation models in computer vision, specifically focusing on the learning by generation approach?", "cites": [ 7100, 9095, 2509 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "9c50d080-4943-4b08-8efe-735c19238add", "query_id": 144, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "\\label{sec:graph_gic}\nThe essential motivation of pretraining based on graph information completion (GIC) is to mask part of the information of the input graph data and recover the masked information based on the unmasked graph data, so as to pretrain the graph embedding, as shown in Fig. \\ref{fig:GIC_GPP}. \nSimilar ideas appeared earlier in the field of image and text processing.\nFor instance, in image processing, information such as image pixels and colors are recovered to pretrain the image encoder; in text processing, many methods implement pretraining of word embeddings and encoders by recovering part of the information in a sentence based on context words.\nThese methods inspire the design of graph completion tasks on graph PFMs.\n\\begin{figure*}[!t]\n \\centering\n \\subfigure[Graph Information Completion (GIC).]{\n \\includegraphics[width=.4\\linewidth]{fig_graph/GIC.pdf}}\n \\centering\n \\subfigure[Graph Property Prediction (GPP).]{\n \\includegraphics[width=.4\\linewidth]{fig_graph/GPP.pdf}}\n\t\\caption{Graph Information Completion (GIC) and Graph Property Prediction (GPP).}\n\t\\label{fig:GIC_GPP}\n\\end{figure*}\nAmong them, You et al.~ are inspired by image inpainting, and first propose to cover them by removing the features of the target nodes, and then recover/predict the features of the masked nodes.\nIn order to recover/predict the masked information, GraphCompetion~ is achieved by providing GCNs with unmasked node features (limited to the 2-layer GCNs of the second-order neighbors of each target node).\nThe purpose of GraphCompetion's pretraining is to help the model better perform feature representation and teach the model to extract features from the context.\nYou et al.~ propose the attribute mask task (namely, AttributeMask), which masks node attributes randomly, and then requires the self-supervising module to reconstruct the masked attributes.\nJin et al.~ think deeply about SSL on graph data, and propose the edge mask task (namely, EdgeMask), seeking to develop self-supervision in pairs based not only on a single node itself but on the connection between two nodes in the graph.\nIn particular, EdgeMask randomly masks some edges and then asks the model to reconstruct the masked edges.\nIn short, EdgeMask is expected to help GNN learn local connectivity information.\nHu et al.~ propose a PFM that masks node and edge attributes and then predicts this masked information based on the adjacent structure.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Graph Learning" ], [ "subsection", "Learning by Graph Information Completion" ] ], "prefix_titles_query": "Can you provide recent research papers on pretrained foundation models utilized for graph learning, specifically focusing on techniques for learning by graph information completion?", "cites": [ 7582, 1440, 7335 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "4ccfac6c-9d70-41dd-b30a-e80b2492db26", "query_id": 145, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "Based on the early homogeneity assumption, a mass of graph models tends to project contextual nodes to similar positions in semantic space.\nSuch consistency of the context in the graph is also applied to the pretraining graph model, which attempts to adjust the node representation by capturing the distribution characteristics of the nodes in the context, as shown in Fig. \\ref{fig:GCA} (a).\nRandom walk is an efficient method to acquire context. \nIt can capture the distribution characteristics of different perspectives in the context by designing a variety of walk strategies.\nThe DeepWalk~ adopts a truncated random walk strategy to represent the node context as the form of a sequence of nodes.\nBy introducing the idea of NLP into the network embedding model, DeepWalk regards the node sequence as a ``sentence''\nand models it based on the skip-gram model, providing an unsupervised and scalable training method for node representation.\nFurthermore, on the basis of DeepWalk, node2vec~ uses two different parameter-controlled random walk strategies to obtain deviated node sequences to fully capture the context information.\nDifferent from randomly sampling nodes from the context, some recent methods directly consider the relationship between the node's k-order neighbor distribution (as positive examples) and non-adjacent nodes (as negative examples), and use this to train the graph model.\nLINE~ respectively proposes first- and second-order proximity to describe the local similarity between pairs of nodes in the graph from different perspectives, and uses it to optimize node representation.\nMeanwhile, LINE uses negative sampling and edge sampling techniques to optimize the second-order traversal and excessive training storage overhead.\nVGAE~ introduces a variational autoencoder to encode graph structure data, and model the node first-order neighbor through a GCN encoder and a simple inner product decoder.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Graph Learning" ], [ "subsection", "Learning by Graph Consistency Analysis " ], [ "paragraph", "Context Consistency" ] ], "prefix_titles_query": "What are the recent advancements in pretrained foundation models (PFMs) for graph learning, with a specific focus on learning by graph consistency analysis, particularly in the context of context consistency?", "cites": [ 1010, 218, 282, 229 ], "cite_extract_rate": 1.0, "origin_cites_number": 4 }, { "original_id": "86076f2a-243d-4e11-8ad7-99d604259a6e", "query_id": 146, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "In the field of NLP and CV, contrastive learning as an efficient self-supervised mechanism is widely used in the pretraining of models.\nIn fact, the internal comparison mechanism of such methods is based on the mutual information estimation of the original graph data and the augmented graph data to maintain the consistency of the data itself, as shown in Fig. \\ref{fig:GCA} (b).\nInspired by contrastive learning, some studies have begun to generate augmented samples of original data samples in the graph model. \nAmong them, two augmented samples from the same original sample are regarded as positive pairs, and two augmented samples from different original samples are regarded as negative pairs.\nFor node-level tasks, GCC~ devises the pretext task as subgraph instance discrimination in and across networks.\nAnd GCC also enhances the ability of GNNs to learn the intrinsic and transferable structural representations by introducing contrastive learning.\nSpecifically, GCC samples subgraphs from the whole graph as augmentations via random walk with restart and artificially designs positional node embedding as node initial features.\nAs a novel graph representation learning model, GCA~ incorporates various priors for topological and semantic aspects of the graph to achieve adaptive contrastive augmentation. \nSpecifically, GCA devises an enhancement scheme based on node centrality measures to highlight important connection structures, while corrupting node features by adding noise to specific nodes to lead the pretraining model to recognize underlying semantic information.\nFor graph-level tasks, some studies have attempted to introduce more diverse contrastive learning strategies.\nAmong them, You et al.~ introduce four common graph augmentation tasks (i.e., node dropping, edge perturbation, attribute masking, and subgraph sampling) into the GL model based on underlying prior and propose a unified comparative learning framework: GraphCL.\nMeanwhile, GraphCL discusses in depth the role of data augmentation in comparative learning and gives experimental demonstration that joint multiple augmentation strategies can improve model performance.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Graph Learning" ], [ "subsection", "Learning by Graph Consistency Analysis " ], [ "paragraph", "Context Consistency" ], [ "paragraph", "Self Consistency" ] ], "prefix_titles_query": "What are the advancements and methodologies in pretrained foundation models (PFMs) for graph learning, specifically focusing on context consistency and self consistency analysis?", "cites": [ 2531, 2532, 1184 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "38bce72c-4b48-41da-8192-684816c0192e", "query_id": 152, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "In ViLBERT~, the text and image modes are first encoded separately, and their outputs go through a standard attention module. This module is based on the Transformer structure. Still, in the self-attention mechanism, each module uses its query to calculate attention with the value and key of another module to integrate the information between different modules.\nThe model is pretrained on two tasks. The first task is the mask task, which is the same as BERT. On the image side, the goal of the task is that when the region image is masked, the classification distribution of the output of the model can be as consistent as possible with the output distribution of the model used to extract the region features (such as Faster-RCNN). The second task is the language image matching task. DALL-E is a series of deep learning models developed by OpenAI to generate images from natural language prompts. The first version of DALL-E uses a transformer-based architecture, similar to the one used in the GPT LMs, to process the textual prompts and generate image-like representations. The model is trained on a dataset of images and their associated textual descriptions based on GPT-3. DALL-E 2~ is the improved version by employing contrastive language-image pretraining (CLIP) for capturing semantic association between image-text pairs and GLIDE diffusion model for text-conditional image synthesis. Furthermore, GPT-4 is proposed by OpenAI recently. It is a large-scale multimodal model which adopts RLHF and demonstrates human-level performance on various professional and academic benchmarks.\nBased on the multi-modal data containing more available information than previous single-modality data, thus the performance of these models gets enhanced by combining with the SSL on the benchmark dataset. Cross and Learn~ is the first method that reveals crossmodal information as an alternative source of supervision and obtains powerful feature representations from combining crossmodal loss and diversity loss in both RGB and optical flow modalities. \nDifferent from the existing methods that learn feature representations from only a single task in cross-domain datasets, Ren and Lee et al.~ propose a novel deep multi-task network to learn more generalizable visual representations to overcome the domain difference and further utilize the cross-domain information in different tasks. In that paper, the cross-domain datasets are real and synthetic datasets generated by a GAN-based network, while the multiple tasks are the predictions of the surface normal, depth, and instance contour in RGB images. This model performs better than any previous single-task-based SSL methods by learning general-purpose visual representations from cross-domain multi-task feature learning. Tian et al.~ believe that a powerful representation is one that models cross-view factors from the perspective of humans view to understand the world. They propose Contrastive Multiview Coding (CMC) to learn a video representation by maximizing mutual information between different views of the same scene.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Other Data Modality" ], [ "subsection", "PFMs for Multimodal" ], [ "paragraph", "Single-Stream Model" ], [ "paragraph", "Cross-Stream Model" ] ], "prefix_titles_query": "What are the developments and distinctions between single-stream and cross-stream models for multimodal pretrained foundation models?", "cites": [ 7585, 1639, 1278, 2558, 2557, 2559, 2525 ], "cite_extract_rate": 1.0, "origin_cites_number": 7 }, { "original_id": "ec1a273d-353b-4af3-af4b-b4673cd4d2a9", "query_id": 153, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "Code generation with LLMs involves using pretrained language models to automatically generate code based on natural language descriptions of a desired program. This approach has the potential to greatly improve the efficiency of software development by reducing the need for manual coding and allowing developers to focus on higher-level tasks.\nThe technique involves training large-scale language models on vast amounts of natural language text and then fine-tuning the models on specific programming tasks. By inputting natural language descriptions of code, the model can generate code snippets that are syntactically and semantically correct. Code generation with LLMs has been applied in various programming domains, including web development, NLP, and data analysis. The models used for code generation include GPT-4, T5, and Codex, among others. For example, Andrei et al.~ have investigated and assessed the fine-tuning of transformer models for personalized code generation. Specifically, they have evaluated the effectiveness of various personalization techniques in the domain of generating unit tests for Java methods and learning to personalize for a specific software project. Shailja et al.~ assess the capacity of LLMs to generate Verilog that is useful. To achieve this, pretrained LLMs are fine-tuned on Verilog datasets collected from GitHub and Verilog textbooks. An evaluation framework is then constructed, consisting of test benches for functional analysis and a flow for testing the syntax of Verilog code generated in response to problems of varying degrees of difficulty. An open-source CodeGen LLM that has undergone fine-tuning has been shown to outperform the current leading commercial Codex LLM. The CodeGen~ is a group of LLMs that have up to 16.1B parameters and can handle both natural language and programming language data. Additionally, they have released the training library JAX FORMER as open-source. Their work demonstrates that the model can perform as well as the previous state-of-the-art zero-shot Python code generation on HumanEval, showcasing the practical applications of the trained model. Synchromesh, introduced in the study by Poesia et al.~, adopts a novel approach called Target Similarity Tuning (TST) to retrieve a small set of examples from a training bank. Then, Synchromesh utilizes these examples to train a pretrained language model and generates programs by applying Constrained Semantic Decoding (CSD). CSD is a general framework that can restrict the output to valid programs in the target language. In this work, the authors show that the combined use of CSD and TST results in significant improvements in prediction accuracy, as well as preventing runtime errors.\nHowever, there are still some limitations to code generation with LLMs, such as the models' tendency to generate overly verbose or inefficient code and their inability to handle complex programming tasks. Nevertheless, the technology has shown significant promise and has the potential to revolutionize the software development industry.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Other Data Modality" ], [ "subsection", "PFM for Code Generation" ] ], "prefix_titles_query": "What is the evolution and application of pretrained foundation models for code generation across different data modalities?", "cites": [ 2561, 2563, 2560, 2562 ], "cite_extract_rate": 1.0, "origin_cites_number": 4 }, { "original_id": "d6377901-42cc-4855-b17d-4fd5c8c828f7", "query_id": 154, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "Model compression requires less computing resources and memory. \nIt is a potential approach to reduce the model size and enhance computation efficiency. The model compression strategy can be divided into two ways: parameter compression and structure compression.\nThe methods of parameter compression include parameter pruning, parameter quantization, low-rank decomposition, and parameter sharing. Parameter pruning refers to designing evaluation criteria for model parameters to delete redundant parameters based on a sizeable PFM. For example, Compressing BERT~ prunes BERT before training while maintaining the performance equivalent to that of the original model. Parameter quantization is the quantization of model parameters from 32-bit full-precision floating-point numbers to lower-order numbers. For example, Q8BERT~ uses 8-bit quantization to compress parameters fourfold with little impact on model performance. Low-rank decomposition is to reduce the dimension of a high-dimensional parameter vector into a sparse low-dimensional vector. Parameter sharing refers to the structured matrix or clustering methods to map model parameters and reduce the number of parameters. For example, the ALBERT~ uses decomposition-embedded parameterization and cross-layer parameter sharing to reduce the parameters in the model.\nStructure compression refers to compact networks and knowledge distillation. A compact network means reducing the number of parameters and calculations by designing a new compact network structure. \nKnowledge distillation refers to the transfer of knowledge from the larger teacher model to the smaller student model through the use of a soft label, etc. DistilBERT~, for example, uses the knowledge distillation method to compress BERT, reducing the size of the BERT model by 40\\% while retaining 97\\% of its language comprehension.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "Other Advanced Topics on PFMs" ], [ "subsection", "Model Compression" ] ], "prefix_titles_query": "What are the recent advancements and techniques in model compression for pretrained foundation models, and how have they evolved from BERT to ChatGPT?", "cites": [ 856, 7579, 1150, 7580 ], "cite_extract_rate": 1.0, "origin_cites_number": 4 }, { "original_id": "a32d62ef-465a-457c-9f5e-a566801d437d", "query_id": 155, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "There are still many methods to manipulate the predicted results of the pretraining model employing a backdoor attack.~ demonstrates that it is possible to construct a weight poisoning attack in which pretrained weights are injected. After the fine-tuning stage, the backdoor is exposed. Attackers manipulate model predictions easily by injecting arbitrary keywords.~ shows that PFMs in NLP can be manipulated by modifying the model corpus. The ``meaning'' of new words or existing words can be controlled by changing their weight parameters.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "Other Advanced Topics on PFMs" ], [ "subsection", "Security and Privacy" ], [ "paragraph", "Generation Adversarial Samples" ], [ "paragraph", "Backdoor Attacks" ] ], "prefix_titles_query": "What are the recent advancements and challenges in addressing security and privacy concerns, specifically focusing on generation adversarial samples and backdoor attacks, in pretrained foundation models?", "cites": [ 2566, 2567 ], "cite_extract_rate": 1.0, "origin_cites_number": 2 }, { "original_id": "cc275360-66c7-4eeb-acdb-c935357457b1", "query_id": 157, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "Generative Adversarial Networks (GANs)~ have provided a paradigm to learn representations for unlabelled data, and spawn many GAN-based approaches on downstream tasks. In image translation, pix2pix software~ first proposes the conditional adversarial networks as a solution to the image-to-image translation problems, and achieves reasonable results on real-world datasets. \nMarkovian Generative Adversarial Networks (MGANs)~ is a method to generate texture synthesis, which can be applied to style transfer and video\nstylization. CycleGAN~ provides a learning algorithm to translate an original image from the source domain to a target domain without containing pairs of images in datasets for supervised learning. StyleGAN~ is a style-based generator to serve as an alternative architecture for traditional GANs. Pixel Recurrent Neural Networks (PixelRNN)~ aims to complete images by modeling full dependencies between the color channels. DiscoGAN~ is designed to learn relations between different domains. \nGANs have also provided a novel direction to study data synthesis because it perfectly simulates the distribution of the original data. Laplacian Pyramid of Adversarial Networks (LAPGAN)~ uses a cascade of convolutional networks to generate images in a coarse-to-fine fashion. Similarly, Stacked Generative Adversarial Networks (SGAN)~ decompose variations into multiple levels and gradually resolve uncertainties by stacking several GANs in a top-down way.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "Traditional Learning Methods" ], [ "subsection", "Traditional Image Learning" ], [ "subsubsection", "Generation-Based Networks" ] ], "prefix_titles_query": "What are the advancements and applications of generation-based networks in traditional image learning methods, particularly in the context of pretrained foundation models from BERT to ChatGPT?", "cites": [ 8567, 1252, 2574, 7217, 8566, 7022, 2573, 157, 896 ], "cite_extract_rate": 1.0, "origin_cites_number": 9 }, { "original_id": "7f51ca8f-a2ef-4915-9892-afbc3e061179", "query_id": 160, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "\\myparagraph{\\textbf{Pretext Tasks}} \n posits a mechanism based on approximate conditional independence (CI) to connect pretext and downstream task data distributions, which suggests that pretext tasks can self-supervisedly learn the representations from unlabelled data that reduce the sample complexity of downstream supervised tasks. The experiments both on CV and NLP task supports this theory. Representation Learning via Invariant Causal Mechanisms (R\\textsc{e}LIC)~ also provides a theoretical understanding from the perspective that the explicit invariance constraints across augmentations can yield improved generalization guarantees.\n\\myparagraph{\\textbf{Multi-View Redundancy}}\nFrom the perspective of a multi-view setting, understands contrastive learning as exploiting multiple views of data for representation learning.\nThis theory provides a theoretical analysis that the linear functions of these representations from pretraining are still competitive compared with the non-linear optimal predictor of the label. In other words, the linear functions of the learned representations are nearly optimal on downstream prediction tasks whenever the different views provide redundant information about the label.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs Theory" ], [ "subsection", "Different Perspectives" ] ], "prefix_titles_query": "What are the different theoretical perspectives in the development and application of pretrained foundation models from BERT to ChatGPT?", "cites": [ 8562, 2582, 2583 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "de3557a8-d0ff-477f-bcd8-0aea88f57ef6", "query_id": 161, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "\\myparagraph{\\textbf{Contrastive Learning}}\nAlthough experimental results show us that previous designs such as contrastive loss or momentum updating can produce impressive performance in SSL. However, one of the most important questions that remain in SSL is why these methods can maintain representation consistency during the pretraining process. A naive view is the minimization between positive pairs can boost invariance learning, while the maximization between negative pairs contributes to avoiding representational collapse. shows that contrastive learning can achieve competitive bound via intra-class concentration, thus leading to the reduction of sample complexity on downstream tasks from the benefit of transferred representations. This research also provides a framework that can be utilized both on the guarantees of the quality of learning representations during the pretraining phase and the future assumptions added to the framework that allow tighter guarantees.\n\\myparagraph{\\textbf{Non-Contrastive Learning}}\nWhile contrastive learning shows an effect by capturing the similarity and dissimilarity among the unlabelled examples, and further converging to an average local optimum which represents the general representations, recent non-contrastive SSL methods such as BYOL and SimSiam also shows the SOTA performance without the design of comparison between negative pairs. Based on the analysis of the eigenspaces, Tian et al.~ study the behavior of non-contrastive SSL training and prove that the effects are from both the predictor and stop-gradient signal. Based on this theory, a novel and simple \\textbf{DirectPred} method is proposed as a by-product of this theoretical exploration.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs Theory" ], [ "subsection", "Different Categories" ] ], "prefix_titles_query": "What are the different categories of pretrained foundation models discussed in the context of their theoretical development from BERT to ChatGPT?", "cites": [ 2529, 2584 ], "cite_extract_rate": 1.0, "origin_cites_number": 2 }, { "original_id": "3df97b8f-533d-47bf-a8d7-c1e2d6830a59", "query_id": 164, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "Model-based Pretraining RL aims to first pretrain a generative world model to capture the underlying structure of the environment and then leverage the world model as a state encoder or simulator during fine-tuning. World Models is the first work that proposes to learn a compressed spatial and temporal representation of the environment in an unsupervised manner using a simple Variational Autoencoder, which greatly improves the sample efficiency compared to training from scratch. However, learning the world model without being aware of the environment's dynamic might lead to ignorance of some key information in the environment. Dreamer proposed to learn latent dynamics by approximating the representation, transition, and reward model. They then train RL agents purely by imagination in a latent space, which is more efficient since it brings a low memory footprint and enables fast predictions of thousands of imagined trajectories in parallel.\nFurthermore, DreamerPro proposes a reconstruction-free approach based on prototypical representations to migrate the task-irrelevant visual distractions problem in the latent dynamics modeling. DreamerPro significantly outperforms previous SOTA methods when there are complex background distractions. To verify whether learning accurate world models for the real world is promising, Daydreamer applies Dreamer to the real-world physical robots problem and empirically demonstrates significant learning efficiency gains.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "PFMs for Reinforcement Learning" ], [ "paragraph", "Model-based Pretraining RL" ] ], "prefix_titles_query": "What are the advancements and applications of pretrained foundation models in model-based reinforcement learning?", "cites": [ 2595, 2598, 2596, 2597, 7592 ], "cite_extract_rate": 1.0, "origin_cites_number": 5 }, { "original_id": "387c3436-5b01-4b33-bab1-b87ca6857923", "query_id": 167, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "This is a fundamental task of NLP to identify people, places, organizations, and other entities in text.\nIt is a crucial primary tool for many NLP tasks, including information extraction, question answering, semantic parsing, machine translation, etc.\n\\myparagraph{CoNLL 2003~}\nIt consists of newswire text from the Reuters RCV1\ncorpus. It contains four different entity types (Location, Organization, Person, and Miscellaneous) and includes 1,393 English news articles, and 909 German news articles.\n\\myparagraph{OntoNotes 5.0~}\nThe dataset consists of 174,5K English, 900K Chinese, and 300K Arabic text data. It comes from telephone conversations, news agencies, radio news, radio conversations, and blogs. It has 18 entity classes containing 11 types, seven values, and 2,945,000 text data.\n\\myparagraph{MSRA~}\nThis is a Chinese dataset that is obtained from the news domain. It has three types of entities and is used as a shared task on SIGNAN back in 2006.", "prefix_titles": [ [ "title", "A Comprehensive Survey on Pretrained Foundation Models: A History from BERT to ChatGPT" ], [ "section", "Datasets" ], [ "subsection", "Downstream Tasks and Datasets on NLP" ], [ "paragraph", "Sentiment Analysis (SA)" ], [ "paragraph", "Named Entity Recognition (NER)" ] ], "prefix_titles_query": "What papers discuss the datasets used for downstream tasks like Sentiment Analysis and Named Entity Recognition in the context of pretrained foundation models from BERT to ChatGPT?", "cites": [ 7, 8385, 7596 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "8469161e-625b-422b-90e0-a6cca9638a04", "query_id": 422, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "From the perspective of the \"Real \\& Fake\", generator is counterfeiter designed to deceive the discriminator, while discriminator is police designed to distinguish between real and fake samples. The motivation of GANs is to train the generator based on the loss of the discriminator. {Compared to supervised classification tasks, discriminator formally needs to perform only bi-classification tasks, which is easy to implement. Therefore, discriminator is very easy to overfit.} Furthermore, unlike the direct training objective of the classification task (Minimizing cross-entropy loss), the objective of GANs training is indirect. Hence, only one-dimensional output of the discriminator does not provide a complete representation on truth or false of samples. Some studies have shown that the present discriminators contain some significant deficiencies in the frequency domain and attribute domain , which are evidence of the lacking discrimination for discriminators. Excessive shortage of discrimination makes the generator lack incentives from the discriminator to learn useful information of the data. {In addition to discriminator overfitting and lacking discrimination of discriminators, discriminator forgetting is another challenge for GANs.} To alleviate these situations, many regularization methods and additional supervision tasks have been proposed in the literature, which can be divided into three categories: \\textit{Data Augmentation and Preprocessing}, \\textit{Consistency Regularization}, and \\textit{Self-supervision}. {All of them are based on data augmentation and are orthogonal to each other. As shown in Table \\ref{table:SOTA }, The state-of-the-art GANs always adopt two or even all of the above regularization.}", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Regularization and Normalization of \"Real \\& Fake\"" ] ], "prefix_titles_query": "What techniques have been used for the regularization and normalization of \"real & fake\" data in GANs, according to recent research?", "cites": [ 6998, 72 ], "cite_extract_rate": 1.0, "origin_cites_number": 2 }, { "original_id": "a5e04be5-4733-477b-9081-c52ce2a3be68", "query_id": 424, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "In context of semi-supervised and unsupervised learning, consistency regularization has been widely used in . It is motivated by the fact that models should produce consistent predictions given input and their semantics-preserving augmentations, such as image rotating, and adversarial attacks. It is pertinent to note that the supervision of GANs training is weak. To increase the discrimination of discriminator, some consistency regularization techniques have also been used. Due to different goals, we divide these into two parts: \\textit{Image Consistency} and \\textit{Network Consistency} as demonstrated in Figure \\ref{fig:consistency}.", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Regularization and Normalization of \"Real \\& Fake\"" ], [ "subsection", "Consistency Regularization" ] ], "prefix_titles_query": "What are the latest research advancements in consistency regularization for \"real & fake\" regularization and normalization in GANs?", "cites": [ 115, 114, 7000, 7192 ], "cite_extract_rate": 1.0, "origin_cites_number": 4 }, { "original_id": "4e2e9490-faa8-4a2f-9dbd-66ae6c225118", "query_id": 425, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "Self-supervised learning aims to learn representations from the data itself without explicit manual supervision. Recently, some self-supervised studies provide competitive results on ImageNet classification and the representations learned from which transfer well to downstream tasks. Self-supervised learning outperforms its supervised pre-training counterpart in many tasks, such as detection and segmentation, sometimes surpassing it by large margins. This suggests that self-supervised learning obtains more representational features and significantly improve the representation of networks. Based on this, self-supervised learning is introduced into the training of GANs, and we divide them into two categories according to different self-supervision tasks: \\textit{Predictive Self-supervised Learning} and \\textit{Contrastive Self-supervised Learning}.", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Regularization and Normalization of \"Real \\& Fake\"" ], [ "subsection", "Self-Supervision" ] ], "prefix_titles_query": "What are the recent advancements in self-supervision techniques for regularizing and normalizing real and fake data in Generative Adversarial Networks (GANs)?", "cites": [ 123, 122, 7000 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "2267f79d-3862-4c72-bdff-9670534c3017", "query_id": 426, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "{According to the perspective of \\textbf{\"Real $\\&$ Fake\"}, many regularization and normalization technologies inspired from supervised learning have been proposed to GANs training. The key point of them is improving the representation and generalizability of the discriminator. \\textit{Data Augmentation and Preprocessing} is a basic operation containing many types such as spatial augmentation, visual augmentation, frequency augmentation, and noise augmentation. Among them, combining adaptive strategies and all augmentation has achieved the most remarkable achievement and has been employed as default operations in most GANs training. \\textit{Consistency Regularization} and \\textit{Self-supervision} are designed additional tasks based on data augmentation, which further improve the efficiency of data augmentation and extract more useful information under stronger data augmentation beyond the existing yet limited practices. Currently, combining contrastive self-supervised learning with adaptive data augmentation has achieved state of the art in GANs training.}", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Regularization and Normalization of \"Real \\& Fake\"" ], [ "subsection", "Summary" ] ], "prefix_titles_query": "What are the key findings in the summary section regarding the regularization and normalization techniques applied to \"real & fake\" data in GANs?", "cites": [ 130, 108, 125 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "fdfb947c-602c-4735-9040-96d30810700e", "query_id": 427, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "From the perspective of \"Fitting distribution\", generator is considered as a distribution mapping function and the optimal discriminator is considered to be the distribution divergence. Wasserstein distance is a popular and important in GANs, and it corresponds to the optimal transport of the generator. To solve the dual problem of Wasserstein distance, Lipschitz continuity is introduced into the GANs training. The Wasserstein-based GANs (WGAN and WGAN-GP) have achieved remarkable results during the training. However, some studies suggest that the success of WGAN-GP is not due to the Wasserstein distance and the Lipschitz constraint of discriminator may improve the performance and stability of GANs training regardless of the statistical distance used as a loss function. Therefore, the Lipschitz continuity of discriminator is an essential condition during GANs training. {Weight clipping is a simple and the first solution to enforce a Lipschitz constraint, which clamps the weights of discriminator to a fixed box after each gradient update.} Furthermore, {gradient penalty}, {weight normalization}, and {weight regularization} are widely applied in GANs training for fulfilling Lipschitz continuity as summarized in subsequent subsections.", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Regularization and Normalization of \"Fitting distribution\" " ] ], "prefix_titles_query": "How do regularization and normalization techniques influence the fitting distribution in Generative Adversarial Networks (GANs)?", "cites": [ 64, 135, 136, 6997 ], "cite_extract_rate": 1.0, "origin_cites_number": 4 }, { "original_id": "1d7d66d8-7065-4eb8-b5e7-40c9b5f20e5c", "query_id": 428, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "Gradient penalty is a simple and direct way to fulfill Lipschitz continuity. Specifically, K-Lipschitz continuity of the function $f$ can be accessed by $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla f(\\hat{x})||_2-{\\rm K})^2$. According to the optimal transport theory mentioned on the Section \\ref{sect:A-1} of the Supplementary Online-only Material, gradient penalty can be used for the approximation of $W_c(\\mu,\\upsilon)$ in WGANs, named WGAN-GP . Specifically, WGAN-GP fulfills the 1-Lipschitz continuity of the discriminator by $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||_2-{\\rm 1})^2$, which limits the gradient of the discriminator to 1. Although WGAN-GP solves the instability of GANs training to some extent, the assumption of optimal transport is a constrained linear programming problem. Overly strict restriction reduces the exploratory of the discriminator. In contrast, the optimal transport with the regular term mentioned is an unconstrained optimization problem. Like optimal transport corresponds to 1-Lipschitz continuity, the optimal transport with the regular term corresponds to k-Lipschitz continuity (${\\rm k}\\leq1$) of the discriminator, named WGAN-LP , which is implemented by $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[\\left(\\mathop{\\max}\\{0,||\\nabla D_{\\theta}(\\hat{x})||_2-1\\}\\right)^2\\right]$. WGAN-LP achieves better performance by using a weaker regularization term which enforces the Lipschitz constraint of the discriminator. \n\tWGAN-GP and WGAN-LP introduce Wasserstein distance into GANs framework. Due to the gap between limited input samples and the strict Lipschitz constraint on the whole input sample domain, the approximation of the Wasserstein distance is a challenging task. To this end, WGAN-div introduces a Wasserstein divergence into GANs training. The objective of WGAN-div can be smoothly derived as:\n\t\\begin{equation}\n\t\t\\mathbb{E}_{y\\sim q(y)}[\\varphi(y)]-\\mathbb{E}_{x\\sim p(x)}[\\varphi(x)]+k \\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[ ||\\varphi(\\hat{x})||^p\\right].\n\t\\end{equation}\n\tThe objective of WGAN-div is similar to WGAN-GP and WGAN-LP. It can be considered as achieving 0-Lipschitz continuity of discriminator by adopting $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[||\\nabla D_{\\theta}(\\hat{x})||^p\\right]$. \n\tGenerally, Wasserstein distance and Wasserstein divergence are reliable ways of measuring the difference between fake and real data distribution, which leads to the stable training of WGAN-based algorithms. However, a recent study shows that the c-transform method achieves better estimation of Wasserstein divergence but leads to worse performance compared to the gradient penalty method. The results demonstrate that the success of WGAN-based methodologies cannot truly be attributed to approximate the Wasserstein distance and the gradient penalty methods improve the performance indeed. Furthermore, some studies also demonstrate that gradient penalty methods of discriminator, such as 1-GP, k-GP (${\\rm k}\\leq 1$), and 0-GP stabilize the training and improve the performance of GANs remarkably regardless of the loss functions. Based on these observations, stabilizing GANs training using gradient penalty is widely applied in the research community for various losses of GANs. In the rest of this section, we discuss gradient penalty methods regardless of the loss function by dividing them into three parts: \\textit{1-GP}: $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}(||\\nabla D_{\\theta}(\\hat{x})||-{\\rm 1})^p$, \\textit{k-GP (${\\rm k}\\leq1$)}: $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[\\left(\\mathop{\\max}\\{0,||\\nabla D_{\\theta}(\\hat{x})||-1\\}\\right)^p\\right]$, and \\textit{0-GP}: $\\mathop{\\min}\\mathbb{E}_{\\hat{x}\\sim\\pi}\\left[||\\nabla D_{\\theta}(\\hat{x})||^p\\right]$, where $\\pi$ is the distribution of different image space (entire image space or part of image space) and $||\\cdot||$ represents the norm of the gradient. Generally, the loss function of the discriminator with GP can be formulated as:\n\t\\begin{equation}\\label{GP}\n\t\t\\mathcal{L}_{D}=f(\\phi,\\theta)+\\lambda\\mathcal{L}_{GP},\n\t\\end{equation}\n\twhere $f(\\phi,\\theta)$ is the uniform loss function defined in Eq (\\ref{EQ:eqn1}) and $\\mathcal{L}_{GP}$ is the gradient penalty regularization.", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Regularization and Normalization of \"Fitting distribution\" " ], [ "subsection", "Gradient Penalty" ] ], "prefix_titles_query": "What are the recent advancements and methodologies regarding the use of gradient penalty for regularization and normalization in Generative Adversarial Networks (GANs) when fitting distributions?", "cites": [ 137, 136, 88, 138, 93, 135, 8317, 6997 ], "cite_extract_rate": 1.0, "origin_cites_number": 8 }, { "original_id": "44c32216-8be6-4e5c-aa2f-ca241ea4139f", "query_id": 429, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "Assuming the objectives of GANs are convex-concave, some studies have proposed the global convergence of GANs . However, these theoretical convergence analyses are only applicable to the GANs with the optimal discriminator. Therefore, some studies focus on analyzing the local convergence of GANs. According to Nagarajan et al. and Mescheder et al. , under some assumptions, GANs dynamics are locally convergent. However, if these assumptions are not satisfied, especially if the data distributions are not continuous, GANs dynamics do not always converge locally unless some regularization techniques are used.\n\tWe review {Jacobian regularization} techniques in this section, which minimize the Jacobian matrix to achieve local convergence. With the same motivation, Mescheder et al. propose a simplified gradient penalties method, named zero-centered gradient penalties (zc-GP), that guarantees the local convergence under suitable assumptions. Since it is similar to 0-GP, we cover it in Section 4.", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Regularization and Normalization of \"Training dynamics\"" ] ], "prefix_titles_query": "What are the recent advancements in regularization and normalization techniques specifically for improving the training dynamics of GANs?", "cites": [ 98, 54, 99, 7190, 8323 ], "cite_extract_rate": 1.0, "origin_cites_number": 5 }, { "original_id": "f33ea61f-e2e1-4dfc-9f9a-3c4da4ad1e64", "query_id": 430, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "Conditional-based layer normalization is only used for the generator of the conditional generation. It aims to introduce conditional information to each layer of the generator, which helps to improve the quality of the generated images. $\\gamma(c)$ and $\\beta(c)$ in Eq (44) are calculated with different features or class labels as input to the neural network in different methods. Miyato et al. and Zhang et al. used the Conditional Batch Normalization (CBN) to encode class labels, thereby improving the quality of conditional generation. Huang et al. and Karras et al. used the Adaptive Instance Normalization (AdaIN) with target images to improve the accuracy of style transfer. Park et al. used the Spatially-Adaptive (de) Normalization (SPADE) with semantic segmentation image to incorporate semantic information into all layers. Wang et al. used the Attentive Normalization (AN) to model long-range dependent attention, which is similar to self-attention GAN .\n\tIn summary, the main difference between these conditional-based normalizations is the content of conditional inputs (c in Eq (45)). As the information of inputs is gradually enriched, the performance of conditional generation is gradually improved. The related summaries are shown in Table \\ref{table:layer normalization}.", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Regularization and Normalization of \"Other methods\"" ], [ "subsection", "Layer Normalization" ], [ "subsubsection", "Conditional-based layer Normalization" ] ], "prefix_titles_query": "What are the recent advancements and techniques in conditional-based layer normalization for regularization and normalization in GANs?", "cites": [ 154, 8324, 150, 151, 157, 7194 ], "cite_extract_rate": 1.0, "origin_cites_number": 6 }, { "original_id": "fb4135aa-ff8b-439d-9a28-a790ffddd379", "query_id": 431, "date": "2020-08-19T12:52:10Z", "year": "2020", "category": "cs.LG", "content": "{In this section, to provide a side view to the selection of regularization and normalization techniques, we investigate the applications of regularization and normalization techniques frequently employed in state-of-the-art and popular GANs. We select six methods (One per year from 2017-2022) categorized into two classes according to different tasks: Unconditional Generation and Conditional Generation. The selected methods and analysis are shown in Table \\ref{table:SOTA }. PGGAN is a popular GAN model in recent years, which grows the size of both the generator and discriminator progressively. PGGAN empowers high-resolution image generation. Since PGGAN was proposed in 2017, only some simple regularization techniques were applied: WGAN-GP , BN , and LN ; BigGAN is a popular conditional generative adversarial networks, which uses many regularization and normalization techniques, such as zc-GP, SN , Off-Diagonal OR , and CBN ; AutoGAN is the first study introducing the Neural architecture search (NAS) to GANs. It defines the search space for the generator architecture and adopts Inception score as the reward to discover the best architecture. The main focus of AutoGAN is architecture, so AutoGAN only comprises SN ; StyleGAN2 is the most popular architecture of GANs, which produces photorealistic images with large varieties and is widely used in image generation, such as Image Completion , Image-to-Image Translation . StyleGAN2-ADA proposes a novelty adaptive data augmentation methods. Combining StyleGAN2 and adaptive data augmentation, StyleGAN2-ADA obtains impressive performance in image generation, particularly in data-effficient generation. Furthermore, InsGen combines StyleGAN2-ADA with contrastive learning, acquiring state of the art on many generation tasks and datasets. Recently, StyleGAN-XL scales StyleGAN to large diverse datasets and sets a new state-of-the-art on large-scale image synthesis. In summary, many regularization and normalization techniques have been used in state-of-the-art GANs with zc-GP and SN being more attractive to researchers. Data augmentation is a striking method and orthogonal to other ongoing researches on training, architecture, and regularization. Therefore, popular augmentation strategies, such as ADA, have been employed as default operations GANs training. Furthermore, self-supervision has been used to further improve the performance of GANs, which is also orthogonal to other methods.}\n\t\\begin{table}\n\t\t\\caption{The applications of the Regularization and Normalization techniques used in SOTA GANs.}\n\t\t\\label{table:SOTA }\n\t\t\\centering\n\t\t\\footnotesize\n\t\t\\begin{tabular}{c | c | c |c | c | c|c }\t\n\t\t\t\\hline\n\t\t\t\\hline\n\t\t\t{ Method}& Task&\\makecell[c]{Gradient \\\\Penalty}&\\makecell[c]{Data augmentation\\\\ and preprocessing}&\\makecell{Self\\\\supervision}&\\makecell[c]{Weight\\\\ normalization }&\\makecell[c]{Layer\\\\ normalization}\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{PGGAN\\\\(2017)}&\\makecell{Unconditinal\\\\ Generation}&WGAN-GP&\\makecell{None}&None&\\makecell{None}&BN: G,LN: D\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{BigGAN\\\\ (2018 )}&\\makecell{Conditinal \\\\Generation}&zc-GP&None&None&\\makecell{SN: G, D}&CBN\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{AutoGAN\\\\ (2019 )}&\\makecell{Unconditinal\\\\ Generation}&None&None&None&\\makecell{SN:D}&None\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{StyleGAN2-ADA\\\\(2020)}&\\makecell{Unconditinal \\\\Generation}&zc-GP&\\makecell{Adaptive}&None&None&IN\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{InsGen\\\\(2021)}&\\makecell{Unconditinal \\\\Generation}&zc-GP&\\makecell{Adaptive}&contrastive&None&IN\\\\\n\t\t\t\\hline\n\t\t\t\\makecell{StyleGAN-XL\\\\(2022)}&\\makecell{Conditinal \\\\Generation}&None&\\makecell{Translation\\\\Cutout}&None&SN:D&IN\\\\\n\t\t\t\\hline\n\t\t\\end{tabular}\n\t\\end{table}", "prefix_titles": [ [ "title", "A Systematic Survey of Regularization and Normalization in GANs" ], [ "section", "Applications of Regularization and Normalization in SOTA GANs" ] ], "prefix_titles_query": "What are the recent applications of regularization and normalization techniques in state-of-the-art GAN architectures?", "cites": [ 56, 164, 8324, 108, 99, 130, 161, 162, 163, 62, 157, 78, 8317 ], "cite_extract_rate": 1.0, "origin_cites_number": 13 }, { "original_id": "126f4ecf-20b1-42a6-a316-22f2b7ead365", "query_id": 723, "date": "2021-05-10T14:07:49Z", "year": "2021", "category": "cs.CL", "content": "The environment of the Policy Learning framework has been a long-existing problem. built a user simulator to model the user feedback as the reward signal of an environment. They modeled a stack-like user agenda to iteratively change the user goal and thus shifting the dialogue states. While using a user simulator for environment modeling seems to be promising for that it involves less human interaction,~ argued that training a user simulator required a large amount of annotated data. proposed Multi-Agent Dialog Policy Learning, where they have two agents interact with each other, performing both user and agent, learning policy simultaneously. Furthermore, they incorporated a role-specific reward to facilitate role-based response generation and here both agents also acted as the environment of the other one.", "prefix_titles": [ [ "title", "Recent Advances in Deep Learning Based Dialogue Systems: A Systematic Survey" ], [ "section", "Task-oriented Dialogue Systems" ], [ "subsection", "Research Challenges and Hot Topics" ], [ "subsubsection", "Training Environment for PL" ] ], "prefix_titles_query": "What are the recent advancements and research challenges in training environments for task-oriented dialogue systems within deep learning-based dialogue systems?", "cites": [ 7483, 1915, 1923 ], "cite_extract_rate": 1.0, "origin_cites_number": 3 }, { "original_id": "fb7b9548-89f9-498a-b8cc-7136406cea76", "query_id": 735, "date": "2021-05-10T14:07:49Z", "year": "2021", "category": "cs.CL", "content": "\\label{Interactive Training}\nInteractive training, also called human-in-loop training, is a unique training method for dialogue systems. Annotated data is fixed and limited, not being able to cover all dialogue settings. Also, it takes a long time to train a good system. But in some industrial products, the dialogue systems need not be perfect when accomplishing their tasks. Thus, interactive training is desirable because the dialogue systems can improve themselves via interactions with users anywhere and anytime, which is a more flexible and cheap way to finetune the parameters. \nTraining schemes with the above intuition have been developed in recent years.~ introduced a reinforcement learning-based online learning framework. The agent interacted with a human dialogue partner and the partner provided feedback as a reward signal. first trained the agent with two-stage supervised learning, and then used an interaction-based reinforcement learning to finetune. Every time the user chose the best one from K responses generated by the pretrained model and then responded to this selected response. Instead of learning through being passively graded,~ proposed a model that actively asked questions to seek improvement. Active learning was applicable to both offline and online learning settings. argued that most conversation samples an agent saw happened after it was pretrained and deployed. Thus, they proposed a framework to train the agent from the real conversations it participated in. The agent evaluated the satisfaction score of the user from the user's response of each turn and explicitly requested the user feedback when it thought that a mistake has been made. The user feedback was further used for learning. placed the interactive learning in a cooperative game and tried to learn a long-term implicit strategy via Reinforce algorithm. Some of these work has been adopted by industry products and is a very promising direction for study.", "prefix_titles": [ [ "title", "Recent Advances in Deep Learning Based Dialogue Systems: A Systematic Survey" ], [ "section", "Open-Domain Dialogue Systems" ], [ "subsection", "Interactive Training" ] ], "prefix_titles_query": "What are the recent advancements in interactive training techniques for open-domain dialogue systems based on deep learning?", "cites": [ 1914, 2009, 7079, 8515, 7459 ], "cite_extract_rate": 1.0, "origin_cites_number": 5 }, { "original_id": "dc290065-ab0a-4da2-9665-8ecf9cd04ce4", "query_id": 775, "date": "2021-07-07T16:39:28Z", "year": "2021", "category": "cs.LG", "content": "Compared to many other principles, single deterministic methods are computational efficient in training and evaluation. For training, only one network has to be trained and often the approaches can even be applied on pre-trained networks. Depending on the actual approach, only a single or at most two forward passes have to be fulfilled for evaluation. The underlying networks could contain more complex loss functions, which slows down the training process or external components that have to be trained and evaluated additionally . But in general, this is still more efficient than the number of predictions needed for ensembles based methods (Section \\ref{sec:ensemble.methods}), Bayesian methods (Section \\ref{sec:bayesian_methods}), and test-time data augmentation methods (Section \\ref{sec:test_time_augmentation}).\nA drawback of single deterministic neural network approaches is the fact that they rely on a single opinion and can therefore become very sensitive to the underlying network architecture, training procedure, and the training data.\n\\input{tables/table_bayesian_methods}", "prefix_titles": [ [ "title", "A Survey of Uncertainty in Deep Neural Networks" ], [ "section", "Uncertainty Estimation" ], [ "subsection", "Single Deterministic Methods" ], [ "subsubsection", "Summing Up Single Deterministic Methods" ] ], "prefix_titles_query": "What are the key findings and advancements in single deterministic methods for uncertainty estimation in deep neural networks?", "cites": [ 4617, 4613 ], "cite_extract_rate": 1.0, "origin_cites_number": 2 }, { "original_id": "6f4b628d-0f75-4948-a123-10026b7e2d15", "query_id": 786, "date": "2023-02-18T20:51:09Z", "year": "2023", "category": "cs.AI", "content": "The modeling method with a contextual LM can be regarded as the autoencoding model. However, due to the inconsistency in the training stage and fine-tuning stage, the performance of the autoencoding model is poor in the Natural Language Generation (NLG) task.\nPermuted LM aims to combine the advantages of the autoregressive LM and the autoencoder LM.\nIt improves the defects of the two models to a great extent and can be used as a basic idea for the construction of future pretraining target tasks.\nFor a given input sequence $T=[w_{1},w_{2}...\n,w_{N}]$, the formal representation of the target function of the permuted LM is as follows\n\\begin{equation}\n\\max _{\\theta} \\mathbb{E}_{z \\sim Z_{N}}\\left[\\sum_{t=1}^{N} \\log p_{\\theta}\\left(x_{z_{T=t}} \\mid x_{z_{T