Confidence In Predictions Research Articles

Frequency and associated factors for swallowing impairment in community-dwelling older persons: a systematic review and meta-analysis.

Swallowing impairment (SI) is an underdiagnosed dysfunction frequently seen as an expected condition of aging. However, SI can lead to health complications and considerable social impact. The objective of this systematic review with meta-analysis was to evaluate the frequency and associated factors with SI in community-dwelling older persons. Searches were performed in 13 electronic databases including MEDLINE and EMBASE (from inception to September 18, 2021). Data extraction and methodological quality assessment of included studies were performed by two independent reviewers. Meta-analysis of proportions with 95% confidence interval (CI) and prediction interval (PI) was used to pool estimates. Subgroup analysis by Country and Assessment Method was performed. General meta-analysis was used to pool measures of association between potential risk factors and SI occurrence (odds ratio [OR] or prevalence ratio [PR]). The worldwide estimated frequency of SI in community-dwelling older persons was 20.35% (95%CI 16.61-24.68%, 95%PI 4.79-56.45, I2 99%, n = 33,291). This estimation varied across assessment methods and by country. The main factors associated with SI were a dry mouth (OR 8.1, 95%CI 4.9-13.4), oral diadochokinesis (OR 5.3, 95%CI 1.0-27.3), ≥ 80years old (OR 4.9, 95%CI 2.6-9.2), genetic factor (SNPrs17601696) (OR 4.8, 95%CI 2.7-8.3), and partial dependence (OR 4.3, 95%CI 2.0-9.3). And the main factors associated with SI estimated by PR were dry mouth sensation (PR 4.1, 95%CI 2.6-6.5), oral sensorimotor alteration (PR 2.6, 95%CI 1.4-4.9), osteoporosis (PR 2.51, 95%CI 1.2-5.3), and heart diseases (PR 2.31, 95%CI 1.1-5.0). One in five older adults worldwide are expected to experience SI and factors associated with this underdiagnosed dysfunction included biological and physiological changes related to aging, physical and psychological conditions, and poor oral health. Early assessment is paramount for the prevention of future clinical complications and should be a high priority in health care practices.

Can artificial intelligence prediction of successful atrial fibrillation catheter ablation therapy be interpretable?

Abstract Background Radiofrequency catheter ablation (RFCA) therapy is the first-line treatment for atrial fibrillation (AF), the most common type of cardiac arrhythmia globally. However, the procedure currently has low success rates in dealing with persistent AF, with a reoccurrence rate of ∼50% post-ablation. Therefore, artificial intelligence (AI), particularly deep learning (DL), has increasingly been applied to improve and optimise treatments for AF. However, AI is limited by its black-box nature. Therefore, a factor hindering the broad clinical application of DL in AF is the lack of assurance that the DL model is using physiological relevant features when making its prediction. In order to provide this assurance, its decision process needs to be interpretable and have biomedical relevance. Aim This study aims to explore interpretability in DL prediction of successful ablation therapy for AF and evaluate if pro-arrhythmogenic regions in the left atrium (LA) were used in its decision process. Methods LA models with segmented fibrotic tissue were derived from 122 late gadolinium-enhanced magnetic resonance (LGE MR) images of persistent AF patients. To increase the dataset size, an additional 199 synthetic LA tissue models were generated from the LGE MR images. Two ablation strategies were simulated: fibrosis-based ablation (FIBRO) and a rotor-based ablation (ROTOR). RFCA strategy success was determined by its ability to terminate persistent AF in 2s with less than 40% of the tissue ablated [1]. A convolutional neural network (CNN) was developed to predict the success of each RFCA strategy, and gradient weighted class activation maps (GradCAM) were used to assess if the CNN was using locations of pro-arrhythmogenic regions in its decision process on an independent test set of 50 LA tissue models. Results For predicting the success of the FIBRO strategy, the CNN model had an AUC (area under the receiver operating characteristic curve) of 0.92±0.02, recall of 0.89±0.03 and precision of 0.82±0.02. For the ROTOR strategy, the AUC was 0.77±0.02, the recall was 0.93±0.04 and the precision was 0.76±0.02. Finally, the independent test set's GradCAM saliency maps showed that 62±25% and 71±13% of ablation lesions (known from the LA model simulations, but unseen by the CNN during training) coincided with informative regions in the saliency maps for the FIBRO and ROTOR strategies, respectively (Figure 1). Conclusion The most informative regions of the saliency maps coincided with the successful ablation lesions, suggesting that the DL model was able to identify ablated pro-arrhythmogenic regions by leveraging structural features of LGE MR images. In the future, this technique could provide a clinician with a decision support tool and increase confidence in the AI prediction. Funding Acknowledgement Type of funding sources: Public grant(s) – National budget only. Main funding source(s): UK Medical Research Council

Automatic Evolution of Machine-Learning-Based Quantum Dynamics with Uncertainty Analysis.

The machine learning approaches are applied in the dynamical simulation of open quantum systems. The long short-term memory recurrent neural network (LSTM-RNN) models are used to simulate the long-time quantum dynamics, which are built based on the key information of the short-time evolution. We employ various hyperparameter optimization methods, including simulated annealing, Bayesian optimization with tree-structured parzen estimator, and random search, to achieve the automatic construction and adjustment of the LSTM-RNN models. The implementation details of three hyperparameter optimization methods are examined, and among them, the simulated annealing approach is strongly recommended due to its excellent performance. The uncertainties of the machine learning models are comprehensively analyzed by the combination of bootstrap sampling and Monte Carlo dropout approaches, which give the prediction confidence of the LSTM-RNN models in the simulation of the open quantum dynamics. This work builds an effective machine learning approach to simulate the dynamics evolution of open quantum systems. In addition, the current study provides an efficient protocol to build optimal neural networks and estimate the trustiness of the machine learning models.

The value of confidence: Confidence prediction errors drive value-based learning in the absence of external feedback.

Reinforcement learning algorithms have a long-standing success story in explaining the dynamics of instrumental conditioning in humans and other species. While normative reinforcement learning models are critically dependent on external feedback, recent findings in the field of perceptual learning point to a crucial role of internally generated reinforcement signals based on subjective confidence, when external feedback is not available. Here, we investigated the existence of such confidence-based learning signals in a key domain of reinforcement-based learning: instrumental conditioning. We conducted a value-based decision making experiment which included phases with and without external feedback and in which participants reported their confidence in addition to choices. Behaviorally, we found signatures of self-reinforcement in phases without feedback, reflected in an increase of subjective confidence and choice consistency. To clarify the mechanistic role of confidence in value-based learning, we compared a family of confidence-based learning models with more standard models predicting either no change in value estimates or a devaluation over time when no external reward is provided. We found that confidence-based models indeed outperformed these reference models, whereby the learning signal of the winning model was based on the prediction error between current confidence and a stimulus-unspecific average of previous confidence levels. Interestingly, individuals with more volatile reward-based value updates in the presence of feedback also showed more volatile confidence-based value updates when feedback was not available. Together, our results provide evidence that confidence-based learning signals affect instrumentally learned subjective values in the absence of external feedback.

Development of an Explainable Fault Diagnosis Framework Based on Sensor Data Imagification: A Case Study of the Robotic Spot-Welding Process

In recent years, various advanced fault diagnostic models applying deep learning techniques have been proposed, but the confidence in model prediction in the industrial field is still low. Therefore, a method is required to establish a reliable fault diagnostic model that can provide an understandable rationale for the prediction result. This article develops an explainable fault diagnosis framework that infers the causal relationship of failure by combining domain knowledge. A novel data imagification methodology that generates fuzzy-based energy pattern image (FEPI) data using sensor signal is applied to the framework, and the physical interpretability of the FEPI data plays a key role in inferring the causality of the fault. Furthermore, a case study of the robotic spot-welding process is conducted to validate the proposed framework. Convolutional neural network (CNN)-based fault diagnostic model is trained by the FEPI data, and the result of gradient-weighted class activation mapping that traces the critical region for fault classification is interpreted by the domain knowledge to infer the failure causes. Finally, the accuracy of fault diagnosis and the performance of causal inference for the explainable fault diagnosis framework are verified together.

A Bayesian Deep Learning Framework for RUL Prediction Incorporating Uncertainty Quantification and Calibration

In this article, deep learning (DL) has attracted increasing attention for remaining useful life (RUL) prediction. However, most DL-based prognostics methods only provide deterministic RUL values while ignoring the associated epistemic and aleatoric uncertainties. In practice, it is important to know the exact confidence in model predictions for decision making. In this article, a Bayesian deep learning (BDL) framework for RUL prediction incorporating uncertainty quantification and calibration is proposed. First, the epistemic and aleatoric uncertainties, which account for the ignorance about the model and the noise inherent in the observations, respectively, are characterized by integrating both types of uncertainties into a BDL framework. Second, to avoid under- and over-confident predictions, a novel iterative calibration method is proposed to jointly calibrate epistemic, aleatoric, and predictive uncertainties by combining isotonic regression with standard deviation scaling. The effectiveness of the proposed method is demonstrated by the case study of turbofan engines and lithium-ion batteries datasets.

Understanding Predictability of Daily Southeast U.S. Precipitation Using Explainable Machine Learning

Abstract We investigate the predictability of the sign of daily southeastern U.S. (SEUS) precipitation anomalies associated with simultaneous predictors of large-scale climate variability using machine learning models. Models using index-based climate predictors and gridded fields of large-scale circulation as predictors are utilized. Logistic regression (LR) and fully connected neural networks using indices of climate phenomena as predictors produce neither accurate nor reliable predictions, indicating that the indices themselves are not good predictors. Using gridded fields as predictors, an LR and convolutional neural network (CNN) are more accurate than the index-based models. However, only the CNN can produce reliable predictions that can be used to identify forecasts of opportunity. Using explainable machine learning we identify which variables and grid points of the input fields are most relevant for confident and correct predictions in the CNN. Our results show that the local circulation is most important as represented by maximum relevance of 850-hPa geopotential heights and zonal winds to making skillful, high-probability predictions. Corresponding composite anomalies identify connections with El Niño–Southern Oscillation during winter and the Atlantic multidecadal oscillation and North Atlantic subtropical high during summer.

Multi-label out-of-distribution detection via exploiting sparsity and co-occurrence of labels

Although Out-Of-Distribution (OOD) detection has been extensively studied, the OOD detection under multi-label settings, which is closer to the real world, is still in its infancy. The pioneer work ignores some unique properties of multi-label images, such as the sparsity and co-occurrence of labels. Here, we empirically observe that these properties readily distinguish OOD and in-distribution data. Motivated by this observation, we propose a novel multi-label OOD detection approach named Sparse Label Co-occurrence Scoring (SLCS) to exploit the sparsity and co-occurrence information of labels. SLCS follows conventions and deems the logits outputted by the penultimate layer of the trained multi-label image classification model as the prediction confidences of a sample to categories in the training label set. A logit sparse filtering process is employed to filter out the low-confidence logits for avoiding the interference of low-confidence predictions while preserving the high-confidence logits to obtain the label sparsity. Then, the label co-occurrence pairs are counted for each sample based on its predicted categories and the label co-occurrence matrix constructed on the training set. Finally, the preserved logits are weighted by the label co-occurrence information and accumulated to produce the OOD detection score for each sample. Extensive experimental results on three well-known multi-label image datasets demonstrate the discriminating power of SLCS, which achieves greatly improved performances compared with the only multi-label OOD detection approach — JointEnergry and the state-of-the-art single-label OOD detection approaches. The performance improvements of SLCS over JointEnergy in FPR95 are 12.85%, 12.41%, and 9.50% on MS-COCO, VOC 2012, and NUS-WIDE datasets respectively.

Complementary characteristics fusion network for weakly supervised salient object detection

Salient object detection (SOD) is a challenging and fundamental research in computer vision and image processing. Since the cost of pixel-level annotations is high, scribble annotations are usually used as weak supervisions. However, scribble annotations are too sparse and always located inside the objects with lacking annotations close to the semantic boundaries, which can't make confident predictions. To alleviate these issues, we propose a novel and effective scribble-based weakly supervised approach for SOD, named complementary characteristics fusion network (CCFNet). To be more specific, we design an edge fusion module (EFM) by taking account of local and high-level semantic information to equip our model, which would be beneficial to enhance the power of aggregating edge information. Then to achieve the complementary role of different features, a series of feature correlation modules (FCMs) are employed to strengthen the localization information and details learning. This is based on low-level, high-level global and edge information, which will complement each other to obtain relatively complete salient regions. Alternatively, to encourage the network to learn structural information and further improve the results of saliency maps in foreground and background, we propose a self-supervised salient detection (SSD) loss. Extensive experiments using five benchmark datasets demonstrate that our proposed approach performs favorably against the state-of-the-art weakly supervised algorithms, and even surpasses the performance of those fully supervised.

Uncertainty estimation for margin detection in cancer surgery using mass spectrometry.

Rapid evaporative ionization mass spectrometry (REIMS) is an emerging technology for clinical margin detection. Deployment of REIMS depends on construction of reliable deep learning models that can categorize tissue according to its metabolomic signature. Challenges associated with developing these models include the presence of noise during data acquisition and the variance in tissue signatures between patients. In this study, we propose integration of uncertainty estimation in deep models to factor predictive confidence into margin detection in cancer surgery. iKnife is used to collect 693 spectra of cancer and healthy samples acquired from 91 patients during basal cell carcinoma resection. A Bayesian neural network and two baseline models are trained on these data to perform classification as well as uncertainty estimation. The samples with high estimated uncertainty are then removed, and new models are trained using the clean data. The performance of proposed and baseline models, with different ratios of filtered data, is then compared. The data filtering does not improve the performance of the baseline models as they cannot provide reliable estimations of uncertainty. In comparison, the proposed model demonstrates a statistically significant improvement in average balanced accuracy (75.2%), sensitivity (74.1%) and AUC (82.1%) after removing uncertain training samples. We also demonstrate that if highly uncertain samples are predicted and removed from the test data, sensitivity further improves to 88.2%. This is the first study that applies uncertainty estimation to inform model training and deployment for tissue recognition in cancer surgery. Uncertainty estimation is leveraged in two ways: by factoring a measure of input noise in training the models and by including predictive confidence in reporting the outputs. We empirically show that considering uncertainty for model development can help improve the overall accuracy of a margin detection system using REIMS.

Simultaneous measurements of unstable and stable Alfvén eigenmodes in JET

In this paper, we report the novel experimental observation of both unstable and stable toroidicity-induced Alfvén eigenmodes (TAEs) measured simultaneously in a JET tokamak plasma. The three-ion-heating scheme (D-DNBI-3He) is employed to accelerate deuterons to MeV energies, thereby destabilizing TAEs with toroidal mode numbers n = 3–5, each decreasing in mode amplitude. At the same time, the Alfvén eigenmode active diagnostic resonantly excites a stable n = 6 TAE with total normalized damping rate −γ/ω 0 ≈ 1%–4%. Hybrid kinetic-MHD modeling with codes NOVA-K and MEGA both find eigenmodes with similar frequencies, mode structures, and radial locations as in experiment. NOVA-K demonstrates good agreement with the n = 3, 4, and 6 TAEs, matching the damping rate of the n = 6 mode within uncertainties and identifying radiative damping as the dominant contribution. Improved agreement is found with MEGA for all modes: the unstable n = 3–5 and stable n = 2, 6 modes, with the latter two stabilized by higher intrinsic damping and lower fast ion drive, respectively. While some discrepancies remain to be resolved, this unique validation effort gives us confidence in TAE stability predictions for future fusion devices.

Machine learning models for rat multigeneration reproductive toxicity prediction.

Reproductive toxicity is one of the prominent endpoints in the risk assessment of environmental and industrial chemicals. Due to the complexity of the reproductive system, traditional reproductive toxicity testing in animals, especially guideline multigeneration reproductive toxicity studies, take a long time and are expensive. Therefore, machine learning, as a promising alternative approach, should be considered when evaluating the reproductive toxicity of chemicals. We curated rat multigeneration reproductive toxicity testing data of 275 chemicals from ToxRefDB (Toxicity Reference Database) and developed predictive models using seven machine learning algorithms (decision tree, decision forest, random forest, k-nearest neighbors, support vector machine, linear discriminant analysis, and logistic regression). A consensus model was built based on the seven individual models. An external validation set was curated from the COSMOS database and the literature. The performances of individual and consensus models were evaluated using 500 iterations of 5-fold cross-validations and the external validation data set. The balanced accuracy of the models ranged from 58% to 65% in the 5-fold cross-validations and 45%–61% in the external validations. Prediction confidence analysis was conducted to provide additional information for more appropriate applications of the developed models. The impact of our findings is in increasing confidence in machine learning models. We demonstrate the importance of using consensus models for harnessing the benefits of multiple machine learning models (i.e., using redundant systems to check validity of outcomes). While we continue to build upon the models to better characterize weak toxicants, there is current utility in saving resources by being able to screen out strong reproductive toxicants before investing in vivo testing. The modeling approach (machine learning models) is offered for assessing the rat multigeneration reproductive toxicity of chemicals. Our results suggest that machine learning may be a promising alternative approach to evaluate the potential reproductive toxicity of chemicals.

Deep learning of monocular depth, optical flow and ego-motion with geometric guidance for UAV navigation in dynamic environments

Computer vision-based depth estimation and visual odometry provide perceptual information useful for robot navigation tasks like obstacle avoidance. However, despite the proliferation of state-of-the-art convolutional neural network (CNN) models for monocular depth, ego-motion and optical flow estimation, a relatively low volume of work has been reported on their practical applications in unmanned aerial vehicle (UAV) navigation. This is due to well-known challenges — embedded hardware constraints, viewpoint variations, scarcity of aerial image datasets, and intricacies of dynamic environments. We address these limitations to facilitate real-world deployment of CNN in UAV navigation. First, we devise efficient confidence weighted adaptive network (Cowan) training framework that iteratively leverages intermediate prediction confidences to enforce cross-task consistency over corresponding image regions. This achieves competitive accuracy with a lightweight CNN capable of real-time execution on resource-constrained embedded systems. Second, we devise a test-time refinement method that adapts the network to dynamic environments while simultaneously improving accuracy. To accomplish this, we first update ego-motion using pose information from on-board inertial measurement unit (IMU). Then, we decompose the UAV’s motion into constituent vectors, and for each axis, we formulate geometric relationships between depth and translation. Based on this information, we triangulate corresponding points acquired through optical flow. Finally, we enforce geometric consistency between the initially updated pose and triangulated depth. Cowan with geometric guided refinement (Cowan-GGR) achieves significant accuracy and robustness. Field tests show the proposed model is capable of accurate depth and object-level motion perception in real-world dynamic environments, thus proving its efficacy in facilitating UAV navigation.

PBPK perspective on alternative CYP3A4 inducers for rifampin.

PBPK perspective on alternative CYP3A4 inducers for rifampin.

Multistep Ahead Multiphase Production Prediction of Fractured Wells Using Bidirectional Gated Recurrent Unit and Multitask Learning

Summary Relying on its strong nonlinear mapping ability, machine learning is found to be efficient and accurate for production prediction of fractured wells compared with conventional analytical methods, numerical simulations, and traditional decline curve analysis. However, its application in forecasting future multistep time series production remains challenging, with complications of error accumulation, growing uncertainty, and degraded accuracy. To this end, we propose a novel multistep ahead production prediction framework based on a bidirectional gated recurrent unit (BiGRU) and multitask learning (MTL) combined neural network (BiGRU-MTL), which can improve prediction performance by sharing task-dependent representations among tasks of multiphase production prediction. The forecasting strategies and evaluation setups for multiple timesteps are elaborated to avoid unfair assessment caused by mixing different prediction confidences over several days. In this framework, BiGRU is in charge of capturing nonlinear patterns of production variation by utilizing both forward and backward sequence information. MTL methods including cross-stitch network (CSN) and weighting losses with homoscedastic uncertainty are incorporated to automatically determine the sharing degree of multiple tasks and the weight ratio of the total loss function. By this means, domain knowledge contained in tasks of multiphase production prediction is deeply leveraged, shared, and coupled to enhance multistep ahead prediction accuracy while meeting the need for multiphase production forecasting. The proposed framework is applied to a synthetic well case, a field well case, and a field multiwell case to progressively prove the feasibility, robustness, and generalization of the BiGRU-MTL model. Experiment results show that the proposed framework outperforms conventional single-task models and commonly used recurrent neural networks (RNNs), furnishing a reliable and stable tool for accurate multistep ahead production prediction. This work promises to provide insights into dynamic production optimization and management in oil- and gasfield sites.

Accuracy and uncertainty of geostatistical models versus machine learning for digital mapping of soil calcium and potassium.

Accuracy and uncertainty of models used for digital soil mapping are important for assessing confidence of predictions and reliable land use planning and management. In this study, two approaches of geostatistical (spatial) and machine learning (ML) models were evaluated forpredictivemapping of soil calcium (Ca) and potassium (K). Two spatial models including empirical Bayesian kriging (EBK) and sequential Gaussian simulation (SGS) were compared with machine learning models: Cubist, random forest (RF) and support vector machine (SVM) in terms of their accuracy and uncertainty for mapping soil Ca and K. The study area is in Nowley, New South Wales, Australia, with an area of 2083ha and a variety of soil types and farming systems. For the models training process, 240 soil samples data and for validation 102 independent samples data were used. For accuracy assessment R2, root mean square error (RMSE), concordance and bias and for uncertainty assessment confidence limits were investigated. Also, in order to compare the outcomes for the two soil properties with different measurement units, mean absolute percentage error (MAPE) and relative uncertainty (RU) as accuracy and uncertainty measures, respectively, were evaluated. Results showed that for K map SGS had the highest R2 (0.74) and lowest RMSE (1.96), followed by EBK with R2 = 0.72 and RMSE = 2.02. For Ca map, EBK model showed the highest accuracy (R2 = 0.46; RMSE = 3.21), followed by SVM and SGS with comparable accuracies. Comparing the two soil properties, Ca map showed higher MAPE and RU, compared to K map. The lowest MAPE was obtained for EBK model (for K = 39) and SGS model (for K = 44). Also, the lowest RU values were found for EBK and SGS models. Among the ML models, SVM showed lower sensitivity to higher variance in data input. In general, the spatial models outperformed the ML models with regard to both accuracy and uncertainty. An additional conclusion is that considering the data variance in the two soil properties, geostatistical models with lower RU and MAPE were relatively less susceptible to data variance, compared to ML models.

SelfNet: A semi-supervised local Fisher discriminant network for few-shot learning

Few-shot learning, employing small-scale labeled samples to recognize new objects, has received substantial research interest. The prototypical network (ProtoNet) is a simple yet effective meta-learning method to solve this problem. In the few-shot scenario, however, the scarcity of data usually has a negative impact on the representational ability of prototypes. In this paper, a unique semi-supervised few-shot learning architecture, referred to as Semi-supervised local Fisher discriminant network (SelfNet), which integrates few-shot learning with subspace learning, is proposed. Using the union of the support set and the additional unlabeled set, a feature projection module is constructed to achieve the subspace projection. Additionally, a pseudo-labeling strategy, which adds the unlabeled samples with high prediction confidence to the support set, is employed to refine the original prototypes. Experimental results on two few-shot classification benchmarks demonstrate that SelfNet can achieve superior performance to the state-of-the-arts, indicating the benefits of utilizing unlabeled samples for feature projection.

High-resolution seismic faults interpretation based on adversarial neural networks with a regularization technique

Geologic fault detection at high precision and resolution is the key for fine structure and reservoir modeling. Previous studies using neural networks for fault segmentation mainly focus on the local features of the targets and train the networks using synthetic data sets. To increase the fault segmentation resolution only using a limited amount of seismic field data, we develop an adversarial neural network architecture for high-resolution identification of faults (FaultAdvNet) taking advantage of global feature fusion. The architecture consists of (1) a light-weight segmentation module (approximately 0.49 M parameters), (2) a feature fusion module considering reflectors of faults and surrounding stratums, and (3) a discriminator module acting as a regularization term. Case studies using seismic field data from the Gulf of Mexico show an overwhelming performance improvement of the FaultAdvNet when compared with other fault detection methods. The FaultAdvNet picks all of the faults with sufficiently high confidence and low prediction risk. The predicted faults of the FaultAdvNet have good continuity and show clear boundary with fault probability values mainly ranging from 0.95 to 1. Saliency analysis also suggests that the FaultAdvNet can focus on the target at a sufficiently higher resolution (dozens of meters). Functionality experiments verify the mechanisms of the feature fusion module and the discriminator module in FaultAdvNet. We consider that a neural network (such as the discriminator) can serve as a data-driven regularization term to constrain the target network (the segmentation network) efficiently, especially given a limited amount of seismic data.

Multi-instance semantic similarity transferring for knowledge distillation

Knowledge distillation is a popular paradigm for learning portable neural networks by transferring the knowledge from a large model into a smaller one. Most existing approaches enhance the student model by utilizing the similarity information between the categories of instance level provided by the teacher model. However, these works ignore the similarity correlation between different instances that plays an important role in confidence prediction. To tackle this issue, we propose a novel method in this paper, called multi-instance semantic similarity transferring for knowledge distillation (STKD), which aims to fully utilize the similarities between categories of multiple samples. Furthermore, we propose to better capture the similarity correlation between different instances by the mixup technique, which creates virtual samples by a weighted linear interpolation. Note that, our distillation loss can fully utilize the incorrect classes similarities by the mixed labels. The proposed approach promotes the performance of student model as the virtual sample created by multiple images produces a similar probability distribution in the teacher and student networks. Experiments and ablation studies on several public classification datasets including CIFAR-10, CIFAR-100, CINIC-10 and Tiny-ImageNet verify that this light-weight method can effectively boost the performance of the compact student model. It shows that STKD has substantially outperformed the vanilla knowledge distillation and achieved superior accuracy over the state-of-the-art knowledge distillation methods.

Semisupervised Semantic Segmentation by Improving Prediction Confidence.

Most of the recent image segmentation methods have tried to achieve the utmost segmentation results using large-scale pixel-level annotated data sets. However, obtaining these pixel-level annotated training data is usually tedious and expensive. In this work, we address the task of semisupervised semantic segmentation, which reduces the need for large numbers of pixel-level annotated images. We propose a method for semisupervised semantic segmentation by improving the confidence of the predicted class probability map via two parts. First, we build an adversarial framework that regards the segmentation network as the generator and uses a fully convolutional network as the discriminator. The adversarial learning makes the prediction class probability closer to 1. Second, the information entropy of the predicted class probability map is computed to represent the unpredictability of the segmentation prediction. Then, we infer the label-error map of the segmentation prediction and minimize the uncertainty on misclassified regions for unlabeled images. In contrast to existing semisupervised and weakly supervised semantic segmentation methods, the proposed method results in more confident predictions by focusing on the misclassified regions, especially the boundary regions. Our experimental results on the PASCAL VOC 2012 and PASCAL-CONTEXT data sets show that the proposed method achieves competitive segmentation performance.