Unsupervised Pretraining Research Articles

From Recognition to Prediction: Leveraging Sequence Reasoning for Action Anticipation

The action anticipation task refers to predicting what action will happen based on observed videos, which requires the model to have a strong ability to summarize the present and then reason about the future. Experience and common sense suggest that there is a significant correlation between different actions, which provides valuable prior knowledge for the action anticipation task. However, previous methods have not effectively modeled this underlying statistical relationship. To address this issue, we propose a novel end-to-end video modeling architecture that utilizes attention mechanisms, named Anticipation via Recognition and Reasoning (ARR). ARR decomposes the action anticipation task into action recognition and sequence reasoning tasks and effectively learns the statistical relationship between actions by next action prediction (NAP). In comparison to existing temporal aggregation strategies, ARR is able to extract more effective features from observable videos to make more reasonable predictions. In addition, to address the challenge of relationship modeling that requires extensive training data, we propose an innovative approach for the unsupervised pre-training of the decoder, which leverages the inherent temporal dynamics of video to enhance the reasoning capabilities of the network. Extensive experiments on the Epic-kitchen-100, EGTEA Gaze+, and 50salads datasets demonstrate the efficacy of the proposed methods. The code is available at https://github.com/linuxsino/ARR .

A Deep Learning Application of Capsule Endoscopic Gastric Structure Recognition Based on a Transformer Model.

Gastric structure recognition systems have become increasingly necessary for the accurate diagnosis of gastric lesions in capsule endoscopy. Deep learning, especially using transformer models, has shown great potential in the recognition of gastrointestinal (GI) images according to self-attention. This study aims to establish an identification model of capsule endoscopy gastric structures to improve the clinical applicability of deep learning to endoscopic image recognition. A total of 3343 wireless capsule endoscopy videos collected at Nanfang Hospital between 2011 and 2021 were used for unsupervised pretraining, while 2433 were for training and 118 were for validation. Fifteen upper GI structures were selected for quantifying the examination quality. We also conducted a comparison of the classification performance between the artificial intelligence model and endoscopists by the accuracy, sensitivity, specificity, and positive and negative predictive values. The transformer-based AI model reached a relatively high level of diagnostic accuracy in gastric structure recognition. Regarding the performance of identifying 15 upper GI structures, the AI model achieved a macroaverage accuracy of 99.6% (95% CI: 99.5-99.7), a macroaverage sensitivity of 96.4% (95% CI: 95.3-97.5), and a macroaverage specificity of 99.8% (95% CI: 99.7-99.9) and achieved a high level of interobserver agreement with endoscopists. The transformer-based AI model can accurately evaluate the gastric structure information of capsule endoscopy with the same performance as that of endoscopists, which will provide tremendous help for doctors in making a diagnosis from a large number of images and improve the efficiency of examination.

Source-free collaborative domain adaptation via multi-perspective feature enrichment for functional MRI analysis

Resting-state functional MRI (rs-fMRI) is increasingly employed in multi-site research to analyze neurological disorders, but there exists cross-site/domain data heterogeneity caused by site effects such as differences in scanners/protocols. Existing domain adaptation methods that reduce fMRI heterogeneity generally require accessing source domain data, which is challenging due to privacy concerns and/or data storage burdens. To this end, we propose a source-free collaborative domain adaptation (SCDA) framework using only a pretrained source model and unlabeled target data. Specifically, a multi-perspective feature enrichment method (MFE) is developed to dynamically exploit target fMRIs from multiple views. To facilitate efficient source-to-target knowledge transfer without accessing source data, we initialize MFE using parameters of a pretrained source model. We also introduce an unsupervised pretraining strategy using 3,806 unlabeled fMRIs from three large-scale auxiliary databases. Experimental results on three public and one private datasets show the efficacy of our method in cross-scanner and cross-study prediction.

Investigating Contrastive Pair Learning's Frontiers in Supervised, Semisupervised, and Self-Supervised Learning.

In recent years, contrastive learning has been a highly favored method for self-supervised representation learning, which significantly improves the unsupervised training of deep image models. Self-supervised learning is a subset of unsupervised learning in which the learning process is supervised by creating pseudolabels from the data themselves. Using supervised final adjustments after unsupervised pretraining is one way to take the most valuable information from a vast collection of unlabeled data and teach from a small number of labeled instances. This study aims firstly to compare contrastive learning with other traditional learning models; secondly to demonstrate by experimental studies the superiority of contrastive learning during classification; thirdly to fine-tune performance using pretrained models and appropriate hyperparameter selection; and finally to address the challenge of using contrastive learning techniques to produce data representations with semantic meaning that are independent of irrelevant factors like position, lighting, and background. Relying on contrastive techniques, the model efficiently captures meaningful representations by discerning similarities and differences between modified copies of the same image. The proposed strategy, involving unsupervised pretraining followed by supervised fine-tuning, improves the robustness, accuracy, and knowledge extraction of deep image models. The results show that even with a modest 5% of data labeled, the semisupervised model achieves an accuracy of 57.72%. However, the use of supervised learning with a contrastive approach and careful hyperparameter tuning increases accuracy to 85.43%. Further adjustment of the hyperparameters resulted in an excellent accuracy of 88.70%.

RS-Dseg: semantic segmentation of high-resolution remote sensing images based on a diffusion model component with unsupervised pretraining

Semantic segmentation plays a crucial role in interpreting remote sensing images, especially in high-resolution scenarios where finer object details, complex spatial information and texture structures exist. To address the challenge of better extracting semantic information and ad-dressing class imbalance in multiclass segmentation, we propose utilizing diffusion models for remote sensing image semantic segmentation, along with a lightweight classification module based on a spatial-channel attention mechanism. Our approach incorporates unsupervised pretrained components with a classification module to accelerate model convergence. The diffusion model component, built on the UNet architecture, effectively captures multiscale features with rich contextual and edge information from images. The lightweight classification module, which leverages spatial-channel attention, focuses more efficiently on spatial-channel regions with significant feature information. We evaluated our approach using three publicly available datasets: Postdam, GID, and Five Billion Pixels. In the test of three datasets, our method achieved the best results. On the GID dataset, the overall accuracy was 96.99%, the mean IoU was 92.17%, and the mean F1 score was 95.83%. In the training phase, our model achieved good performance after only 30 training cycles. Compared with other models, our method reduces the number of parameters, improves the training speed, and has obvious performance advantages.

From RNN techniques to pre-trained models: emphasis on the use in Arabic machine translation

<p>Neural Machine Translation (NMT) has gained increasing attention in recent years due to its promising performance over conventional approaches such as Statistical Machine Translation. Nevertheless, when applied to languages with different structures, such as the pair (English, Arabic) that interests us in this work, its efficiency is degraded.<br />Alternatively, the use of unsupervised pre-training of large neural models has recently made a significant leap forward in the field of Natural Language Processing (NLP). By warm starting from published checkpoints, NLP practitioners have been able to push the state-of-the-art boundaries on multiple benchmarks while simultaneously saving considerable computational time. The emphasis so far has been mainly on natural language understanding problems. In this paper, we show the efficacy of pre-trained checkpoints for Arabic Machine Translation. We have evolved a transformer-based sequence-to-sequence model which is compatible with pre-trained checkpoints publicly available for Arabic Bidirectional Encoder Representations from Transformers (AraBERT) and Arabic Generative Pre-trained Transformer (AraGPT) and conducted a thorough empirical study on the usefulness of initializing our Arabic MT model, both encoder and decoder, with these checkpoints. Our models yield new state-of-the-art results in Arabic MT.</p>

QDETRv: Query-Guided DETR for One-Shot Object Localization in Videos

In this work, we study one-shot video object localization problem that aims to localize instances of unseen objects in the target video using a single query image of the object. Toward addressing this challenging problem, we extend a popular and successful object detection method, namely DETR (Detection Transformer), and introduce a novel approach –query-guided detection transformer for videos (QDETRv). A distinctive feature of QDETRv is its capacity to exploit information from the query image and spatio-temporal context of the target video, which significantly aids in precisely pinpointing the desired object in the video. We incorporate cross-attention mechanisms that capture temporal relationships across adjacent frames to handle the dynamic context in videos effectively. Further, to ensure strong initialization for QDETRv, we also introduce a novel unsupervised pretraining technique tailored to videos. This involves training our model on synthetic object trajectories with an analogous objective as the query-guided localization task. During this pretraining phase, we incorporate recurrent object queries and loss functions that encourage accurate patch feature reconstruction. These additions enable better temporal understanding and robust representation learning. Our experiments show that the proposed model significantly outperforms the competitive baselines on two public benchmarks, VidOR and ImageNet-VidVRD, extended for one-shot open-set localization tasks.

Abstract 4911: PAIRWISE: Deep learning-powered personalized medicine approach identifies actionable drug combinations in B cell lymphoma

Abstract Combination drug therapies offer promise for improving cancer treatment. However, identifying actionable drug combinations for specific cancer subtypes and individual patients remains challenging due to the vast number of possibilities. Existing approaches like high-throughput screening (HTS) and rule-based drug additions have limitations in their ability to predict clinically translatable synergistic combinations. Therefore, advanced machine learning focused on patient-centric modelling have the potential to optimize combination therapies. PAIRWISE is a patient-centric deep learning-powered drug combination recommendation platform. Given gene expression along with chemical structures and targets from a set of two drugs being applied, the output is the binary combination outcome. PAIRWISE introduces a new approach of representing tumor genomics using the unsupervised pretraining design on 11K TCGA tumors. The power of PAIRWISE is derived from knowledge-guided transfer learning by learning from in vitro experiments with drug-drug-cell interactions and cellular heterogeneity. PAIRWISE can accurately predict combination synergy with area under the curve (AUC) > 0.85 across 10 tissue types in a large dataset of 200K in vitro drug combinations and significantly outperform all state-of-the-art methods. Diffuse large B cell lymphoma (DLBCL) is the most common B cell non-Hodgkin lymphoma and remains incurable in around 40% of patients. Finding optimal combination therapy for each DLBCL patient is a critical unmet need. As independent validation, we applied PAIRWISE to an external HTS study of Bruton Tyrosine Kinase inhibitor (BTKi) combinations (Griner et al., 2014), confirming the reproducibility and accuracy of high-confidence predictions (AUC = 0.72). PAIRWISE was also applied to transcriptomic data of 29 DLBCL cell lines (Bal et al., 2022) and predicted BTKi-containing combinations across ABC subtypes with higher synergy scores than GCB (P = 0.02). This is consistent with expected better overall efficacy of BTK inhibitors in ABC subtype (Younes et al., 2019; Davis et al., 2010; Wilson et al., 2015). To predict opportunities in DLBCL, PAIRWISE was used to identify BTKi partners for an additional 211 anti-cancer drugs using 562 published ex vivo DLBCL samples (Schmitz et al., 2018). PAIRWISE effectively identified novel DLBCL sub-population with significantly higher response scores for predicted BTKi combinations and favorable survival. Notably, patients predicted as responsive had significantly better outcomes upon first-line chemotherapy treatments than non-responsive ones (P < 0.0001). Collectively, we demonstrate that PAIRWISE is a highly efficient and generalizable framework that enables confident predictions of pairwise drug combinations at disease subtype level, with the potential of informing treatment options and clinical trial design. Citation Format: Chengqi Xu, Jake Cohen-Setton, Zhouyang Shen, Marta Milo, Ilkay US, Ben Sidders, Jude Fitzgibbon, Heng Pan, Olivier Elemento, Krishna Bulusu. PAIRWISE: Deep learning-powered personalized medicine approach identifies actionable drug combinations in B cell lymphoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 4911.

Determination and investigation of shear wave velocity based on one deep/machine learning technique

Rock physics plays an important role in the oil, gas, and water industries by providing essential data for reservoir management. This study focuses on predicting Shear Wave Velocity (Vs), a key geophysical parameter, using the Deep Belief Network (DBN) and Random Forest (RF) algorithms. The dataset is divided for training, validation, and testing, revealing that the DBN model surpasses the RF and empirical models in accuracy. With a low RMSE value of 0.0398 km/s in testing and an R2 of 0.9775, the DBN model demonstrates precision in Vs prediction. The analysis identifies specific parameters—Measure Depth (MD), Caliper (CALL), Gamma Ray (GR), Shallow Electrical Resistivity (RES-S), Medium Electrical Resistivity (RES-M), Deep Electrical Resistivity (RES-D), Compressional Wave Velocity (DCT), Neutron Porosity (NPHI), and Photoelectric Coefficient Index (PEF)—as significantly influencing Vs, while CALI and GR have a relatively lesser effect. The results of the Vs prediction indicated that the DBN algorithm excels in deep feature extraction, unsupervised pretraining, and efficient handling of complex data. Ultimately, the DBN algorithm proves to be a superior and time-saving alternative for Vs prediction and analysis in subsurface reservoirs.

CD-MAE: Contrastive Dual-Masked Autoencoder Pre-Training Model for PCB CT Image Element Segmentation

Element detection is an important step in the process of the non-destructive testing of printed circuit boards (PCB) based on computed tomography (CT). Compared with the traditional manual detection method, the image semantic segmentation method based on deep learning greatly improves efficiency and accuracy. However, semantic segmentation models often require a large amount of data for supervised training to generalize better model performance. Unlike natural images, the PCB CT image annotation task is more time-consuming and laborious than the semantic segmentation task. In order to reduce the cost of labeling and improve the ability of the model to utilize unlabeled data, unsupervised pre-training is a very reasonable and necessary choice. The masked image reconstruction model represented by a masked autoencoder is pre-trained on the unlabeled data, learning a strong feature representation ability by recovering the masked image, and shows a good generalization ability in various downstream tasks. In the PCB CT image element segmentation task, considering the characteristics of the image, it is necessary to use a model with strong feature robustness in the pre-training stage to realize the representation learning on a large number of unlabeled PCB CT images. Based on the above purposes, we proposed a contrastive dual-masked autoencoder (CD-MAE) pre-training model, which can learn more robust feature representation on unlabeled PCB CT images. Our experiments show that the CD-MAE outperforms the baseline model and fully supervised models in the PCB CT element segmentation task.

Bidirectional encoders to state-of-the-art: a review of BERT and its transformative impact on natural language processing

First developed in 2018 by Google researchers, Bidirectional Encoder Representations from Transformers (BERT) represents a breakthrough in natural language processing (NLP). BERT achieved state-of-the-art results across a range of NLP tasks while using a single transformer-based neural network architecture. This work reviews BERT's technical approach, performance when published, and significant research impact since release. We provide background on BERT's foundations like transformer encoders and transfer learning from universal language models. Core technical innovations include deeply bidirectional conditioning and a masked language modeling objective during BERT's unsupervised pretraining phase. For evaluation, BERT was fine-tuned and tested on eleven NLP tasks ranging from question answering to sentiment analysis via the GLUE benchmark, achieving new state-of-the-art results. Additionally, this work analyzes BERT's immense research influence as an accessible technique surpassing specialized models. BERT catalyzed adoption of pretraining and transfer learning for NLP. Quantitatively, over 10,000 papers have extended BERT and it is integrated widely across industry applications. Future directions based on BERT scale towards billions of parameters and multilingual representations. In summary, this work reviews the method, performance, impact and future outlook for BERT as a foundational NLP technique. We provide background on BERT's foundations like transformer encoders and transfer learning from universal language models. Core technical innovations include deeply bidirectional conditioning and a masked language modeling objective during BERT's unsupervised pretraining phase. For evaluation, BERT was fine-tuned and tested on eleven NLP tasks ranging from question answering to sentiment analysis via the GLUE benchmark, achieving new state-of-the-art results. Additionally, this work analyzes BERT's immense research influence as an accessible technique surpassing specialized models. BERT catalyzed adoption of pretraining and transfer learning for NLP. Quantitatively, over 10,000 papers have extended BERT and it is integrated widely across industry applications. Future directions based on BERT scale towards billions of parameters and multilingual representations. In summary, this work reviews the method, performance, impact and future outlook for BERT as a foundational NLP technique.

Deeply Unsupervised Patch Re-Identification for Pre-Training Object Detectors.

Unsupervised pre-training aims at learning transferable features that are beneficial for downstream tasks. However, most state-of-the-art unsupervised methods concentrate on learning global representations for image-level classification tasks instead of discriminative local region representations, which limits their transferability to region-level downstream tasks, such as object detection. To improve the transferability of pre-trained features to object detection, we present Deeply Unsupervised Patch Re-ID (DUPR), a simple yet effective method for unsupervised visual representation learning. The patch Re-ID task treats individual patch as a pseudo-identity and contrastively learns its correspondence in two views, enabling us to obtain discriminative local features for object detection. Then the proposed patch Re-ID is performed in a deeply unsupervised manner, appealing to object detection, which usually requires multi-level feature maps. Extensive experiments demonstrate that DUPR outperforms state-of-the-art unsupervised pre-trainings and even the ImageNet supervised pre-training on various downstream tasks related to object detection.

Unsupervised Representation Learning for Large-Scale Wafer Maps in Micro-Electronic Manufacturing

Recognition of wafer map defect patterns is essential for evaluating the reliability of micro-electronic manufacturing. Due to the difficulty of labeling, the available large-scale wafer maps are raw data without labeling. The scarcity of labeled samples reduces the defect pattern recognition accuracy of popular deep convolutional neural networks. To overcome this problem, we propose a wafer map deep clustering (WMDC) model. It learns generic representations from unlabeled datasets in an unsupervised manner. A prototype metric loss during training helps to learn the semantic features of the categories. We improve the recognition accuracy of the model when trained using scarce labeled data by transferring the weights of unsupervised pretraining. Experiments on WM811K and MixedWM38 wafer datasets demonstrate that the WMDC model is capable of obtaining robust prior representations from the unlabeled wafer maps. Accuracies of 97.43% and 98.74% are obtained when fine-tuning using scarce labeled data from both datasets, respectively.

Low-Light Outdoor Smoke Detection Based on Unsupervised Pre-Training

Low-Light Outdoor Smoke Detection Based on Unsupervised Pre-Training

Quality-Driven Regularization for Deep Learning Networks and Its Application to Industrial Soft Sensors.

The growth of data collection in industrial processes has led to a renewed emphasis on the development of data-driven soft sensors. A key step in building an accurate, reliable soft sensor is feature representation. Deep networks have shown great ability to learn hierarchical data features using unsupervised pretraining and supervised fine-tuning. For typical deep networks like stacked auto-encoder (SAE), the pretraining stage is unsupervised, in which some important information related to quality variables may be discarded. In this article, a new quality-driven regularization (QR) is proposed for deep networks to learn quality-related features from industrial process data. Specifically, a QR-based SAE (QR-SAE) is developed, which changes the loss function to control the weights of the different input variables. By choosing an appropriate inductive bias for the weight matrix, the model provides quality-relevant information for predictive modeling. Finally, the proposed QR-SAE is used to predict the quality of a real industrial hydrocracking process. Comparative experiments show that QR-SAE can extract quality-related features and achieve accurate prediction performance.

Large-Scale Unsupervised Audio Pre-Training for Video-to-Speech Synthesis

Large-Scale Unsupervised Audio Pre-Training for Video-to-Speech Synthesis

DNA language models are powerful predictors of genome-wide variant effects.

The expanding catalog of genome-wide association studies (GWAS) provides biological insights across a variety of species, but identifying the causal variants behind these associations remains a significant challenge. Experimental validation is both labor-intensive and costly, highlighting the need for accurate, scalable computational methods to predict the effects of genetic variants across the entire genome. Inspired by recent progress in natural language processing, unsupervised pretraining on large protein sequence databases has proven successful in extracting complex information related to proteins. These models showcase their ability to learn variant effects in coding regions using an unsupervised approach. Expanding on this idea, we here introduce the Genomic Pre-trained Network (GPN), a model designed to learn genome-wide variant effects through unsupervised pretraining on genomic DNA sequences. Our model also successfully learns gene structure and DNA motifs without any supervision. To demonstrate its utility, we train GPN on unaligned reference genomes of Arabidopsis thaliana and seven related species within the Brassicales order and evaluate its ability to predict the functional impact of genetic variants in A. thaliana by utilizing allele frequencies from the 1001 Genomes Project and a comprehensive database of GWAS. Notably, GPN outperforms predictors based on popular conservation scores such as phyloP and phastCons. Our predictions for A. thaliana can be visualized as sequence logos in the UCSC Genome Browser (https://genome.ucsc.edu/s/gbenegas/gpn-arabidopsis). We provide code (https://github.com/songlab-cal/gpn) to train GPN for any given species using its DNA sequence alone, enabling unsupervised prediction of variant effects across the entire genome.

Generation of a Melanoma and Nevus Data Set From Unstandardized Clinical Photographs on the Internet

Artificial intelligence (AI) training for diagnosing dermatologic images requires large amounts of clean data. Dermatologic images have different compositions, and many are inaccessible due to privacy concerns, which hinder the development of AI. To build a training data set for discriminative and generative AI from unstandardized internet images of melanoma and nevus. In this diagnostic study, a total of 5619 (CAN5600 data set) and 2006 (CAN2000 data set; a manually revised subset of CAN5600) cropped lesion images of either melanoma or nevus were semiautomatically annotated from approximately 500 000 photographs on the internet using convolutional neural networks (CNNs), region-based CNNs, and large mask inpainting. For unsupervised pretraining, 132 673 possible lesions (LESION130k data set) were also created with diversity by collecting images from 18 482 websites in approximately 80 countries. A total of 5000 synthetic images (GAN5000 data set) were generated using the generative adversarial network (StyleGAN2-ADA; training, CAN2000 data set; pretraining, LESION130k data set). The area under the receiver operating characteristic curve (AUROC) for determining malignant neoplasms was analyzed. In each test, 1 of the 7 preexisting public data sets (total of 2312 images; including Edinburgh, an SNU subset, Asan test, Waterloo, 7-point criteria evaluation, PAD-UFES-20, and MED-NODE) was used as the test data set. Subsequently, a comparative study was conducted between the performance of the EfficientNet Lite0 CNN on the proposed data set and that trained on the remaining 6 preexisting data sets. The EfficientNet Lite0 CNN trained on the annotated or synthetic images achieved higher or equivalent mean (SD) AUROCs to the EfficientNet Lite0 trained using the pathologically confirmed public data sets, including CAN5600 (0.874 [0.042]; P = .02), CAN2000 (0.848 [0.027]; P = .08), and GAN5000 (0.838 [0.040]; P = .31 [Wilcoxon signed rank test]) and the preexisting data sets combined (0.809 [0.063]) by the benefits of increased size of the training data set. The synthetic data set in this diagnostic study was created using various AI technologies from internet images. A neural network trained on the created data set (CAN5600) performed better than the same network trained on preexisting data sets combined. Both the annotated (CAN5600 and LESION130k) and synthetic (GAN5000) data sets could be shared for AI training and consensus between physicians.

Unsupervised pre-training of graph transformers on patient population graphs.

Pre-training has shown success in different areas of machine learning, such as Computer Vision, Natural Language Processing (NLP), and medical imaging. However, it has not been fully explored for clinical data analysis. An immense amount of clinical records are recorded, but still, data and labels can be scarce for data collected in small hospitals or dealing with rare diseases. In such scenarios, pre-training on a larger set of unlabeled clinical data could improve performance. In this paper, we propose novel unsupervised pre-training techniques designed for heterogeneous, multi-modal clinical data for patient outcome prediction inspired by masked language modeling (MLM), by leveraging graph deep learning over population graphs. To this end, we further propose a graph-transformer-based network, designed to handle heterogeneous clinical data. By combining masking-based pre-training with a transformer-based network, we translate the success of masking-based pre-training in other domains to heterogeneous clinical data. We show the benefit of our pre-training method in a self-supervised and a transfer learning setting, utilizing three medical datasets TADPOLE, MIMIC-III, and a Sepsis Prediction Dataset. We find that our proposed pre-training methods help in modeling the data at a patient and population level and improve performance in different fine-tuning tasks on all datasets.

Understanding How Pretraining Regularizes Deep Learning Algorithms.

Deep learning algorithms have led to a series of breakthroughs in computer vision, acoustical signal processing, and others. However, they have only been popularized recently due to the groundbreaking techniques developed for training deep architectures. Understanding the training techniques is important if we want to further improve them. Through extensive experimentation, Erhan et al. (2010) empirically illustrated that unsupervised pretraining has an effect of regularization for deep learning algorithms. However, theoretical justifications for the observation remain elusive. In this article, we provide theoretical supports by analyzing how unsupervised pretraining regularizes deep learning algorithms. Specifically, we interpret deep learning algorithms as the traditional Tikhonov-regularized batch learning algorithms that simultaneously learn predictors in the input feature spaces and the parameters of the neural networks to produce the Tikhonov matrices. We prove that unsupervised pretraining helps in learning meaningful Tikhonov matrices, which will make the deep learning algorithms uniformly stable and the learned predictor will generalize fast w.r.t. the sample size. Unsupervised pretraining, therefore, can be interpreted as to have the function of regularization.