Large Training Set Research Articles

A Tool to Assist in the Analysis of Gaze Patterns in Upper Limb Prosthetic Use

Gaze-tracking, where the point of regard of a subject is mapped onto the image of the scene the subject sees, can be employed to study the visual attention of the users of prosthetic hands. It can show whether the user pays greater attention to the actions of their prosthetic hand as they use it to perform manipulation tasks, compared with the general population. Conventional analysis of the video data requires a human operator to identify the key areas of interest in every frame of the video data. Computer vision techniques can assist with this process, but fully automatic systems require large training sets. Prosthetic investigations tend to be limited in numbers. However, if the assessment task is well-controlled, it is possible to make a much simpler system that uses the initial input from an operator to identify the areas of interest and then the computer tracks the objects throughout the task. The tool described here employs colour separation and edge detection on images of the visual field to identify the objects to be tracked. To simplify the computer’s task further, this test uses the Southampton Hand Assessment Procedure (SHAP) to define the activity spatially and temporarily, reducing the search space for the computer. The work reported here concerns the development of a software tool capable of identifying and tracking the points of regard and areas of interest throughout an activity with minimum human operator input. Gaze was successfully tracked for fourteen unimpaired subjects and was compared with the gaze of four users of myoelectric hands. The SHAP cutting task is described and the differences in attention observed with a greater number of shorter fixations by the prosthesis users compared to unimpaired subjects. There was less looking ahead to the next phase of the task by the prosthesis users.

Enhancing sea ice segmentation in Sentinel-1 images with atrous convolutions

ABSTRACT Due to the growing volume of remote sensing data and the low latency required for safe marine navigation, machine learning (ML) algorithms are being developed to accelerate sea ice chart generation, currently a manual interpretation task. However, the low signal-to-noise ratio of the freely available Sentinel-1 Synthetic Aperture Radar (SAR) imagery, the ambiguity of backscatter signals for ice types, and the scarcity of open-source high-resolution labelled data makes automating sea ice mapping challenging. We use Extreme Earth version 2, a high-resolution benchmark dataset generated for ML training and evaluation, to investigate the effectiveness of ML for automated sea ice mapping. Our customized pipeline combines ResNets and Atrous Spatial Pyramid Pooling for SAR image segmentation. We investigate the performance of our model for: i) binary classification of sea ice and open water in a segmentation framework; and ii) a multiclass segmentation of five sea ice types. For binary ice-water classification, models trained with our largest training set have weighted F1 scores all greater than 0.95 for January and July test scenes. Specifically, the median weighted F1 score was 0.98, indicating high performance for both months. By comparison, a competitive baseline U-Net has a weighted average F1 score of ranging from 0.92 to 0.94 (median 0.93) for July, and 0.97 to 0.98 (median 0.97) for January. Multiclass ice type classification is more challenging, and even though our models achieve 2% improvement in weighted F1 average compared to the baseline U-Net, test weighted F1 is generally between 0.6 and 0.80. Our approach can efficiently segment full SAR scenes in one run, is faster than the baseline U-Net, retains spatial resolution and dimension, and is more robust against noise compared to approaches that rely on patch classification.

A new computationally efficient method to tune BERT networks – transfer learning

BERT is a pre-trained language model that achieves state-of-the-art performance on natural language processing (NLP) tasks. Once it was published, it quickly became one of the most popular models in the NLP field. The official recommended method for applying BERT to downstream tasks is fine-tuning. However, we argue that transfer learning is also a very practical approach to applying BERT, and this approach has its own advantages compared to fine-tuning. In this paper, we explore the advantages of the transfer learning approach through the method that uses transfer learning and fine-tuning approaches separately to apply BERT to the same eight GLUE benchmark tasks and compare the training time spent and model performance scores, such as accuracy or F1 score, of the two approaches. Finally, this paper finds that among all eight GLUE tasks and on small training sets, transfer learning saves 30% to 50% of the time compared to a fine-tuning approach to train the models on the data of downstream tasks and can achieve very similar performance. Besides, for the same amount of time, transfer learning can obtain higher performance scores than fine-tuning on larger training sets. In conclusion, on small training sets, transfer learning has a huge advantage in time consumption under the premise of approximate performance, and it also performs better on large training sets under the same length of time, compared to fine-tuning.

Molecular machine learning with conformer ensembles

Virtual screening can accelerate drug discovery by identifying promising candidates for experimental evaluation. Machine learning is a powerful method for screening, as it can learn complex structure–property relationships from experimental data and make rapid predictions over virtual libraries. Molecules inherently exist as a three-dimensional ensemble and their biological action typically occurs through supramolecular recognition. However, most deep learning approaches to molecular property prediction use a 2D graph representation as input, and in some cases a single 3D conformation. Here we investigate how the 3D information of multiple conformers, traditionally known as 4D information in the cheminformatics community, can improve molecular property prediction in deep learning models. We introduce multiple deep learning models that expand upon key architectures such as ChemProp and SchNet, adding elements such as multiple-conformer inputs and conformer attention. We then benchmark the performance trade-offs of these models on 2D, 3D and 4D representations in the prediction of drug activity using a large training set of geometrically resolved molecules. The new architectures perform significantly better than 2D models, but their performance is often just as strong with a single conformer as with many. We also find that 4D deep learning models learn interpretable attention weights for each conformer.

Optimization with Neural Network Feasibility Surrogates: Formulations and Application to Security-Constrained Optimal Power Flow

In many areas of constrained optimization, representing all possible constraints that give rise to an accurate feasible region can be difficult and computationally prohibitive for online use. Satisfying feasibility constraints becomes more challenging in high-dimensional, non-convex regimes which are common in engineering applications. A prominent example that is explored in the manuscript is the security-constrained optimal power flow (SCOPF) problem, which minimizes power generation costs, while enforcing system feasibility under contingency failures in the transmission network. In its full form, this problem has been modeled as a nonlinear two-stage stochastic programming problem. In this work, we propose a hybrid structure that incorporates and takes advantage of both a high-fidelity physical model and fast machine learning surrogates. Neural network (NN) models have been shown to classify highly non-linear functions and can be trained offline but require large training sets. In this work, we present how model-guided sampling can efficiently create datasets that are highly informative to a NN classifier for non-convex functions. We show how the resultant NN surrogates can be integrated into a non-linear program as smooth, continuous functions to simultaneously optimize the objective function and enforce feasibility using existing non-linear solvers. Overall, this allows us to optimize instances of the SCOPF problem with an order of magnitude CPU improvement over existing methods.

MTMVC: Semi-supervised 3D hand pose estimation using multi-task and multi-view consistency

The high performance of state-of-the-art deep learning methods for 3D hand pose estimation heavily depends on a large annotated training set. However, it is difficult and time-consuming to obtain the annotations for 3D hand poses. To leverage unannotated images to reduce the annotation cost, we propose a semi-supervised method based on Multi-Task and Multi-View Consistency (MTMVC) for hand pose estimation. First, we obtain the joints based on heatmap prediction and coordinate regression parallelly and encourage their consistency. Second, we introduce multi-view consistency to encourage the predicted poses to be rotation-invariant. Thirdly, to make the network pay more attention to the hand region, we propose a spatially weighted consistency. Experiments on four public datasets showed that our proposed MTMVC outperformed existing semi-supervised hand pose estimation methods, and by only using half of the annotations, the accuracy of our method was comparable to those of several state-of-the-art fully supervised methods.

How to DP-fy ML: A Practical Guide to Machine Learning with Differential Privacy

Machine Learning (ML) models are ubiquitous in real-world applications and are a constant focus of research. Modern ML models have become more complex, deeper, and harder to reason about. At the same time, the community has started to realize the importance of protecting the privacy of the training data that goes into these models. Differential Privacy (DP) has become a gold standard for making formal statements about data anonymization. However, while some adoption of DP has happened in industry, attempts to apply DP to real world complex ML models are still few and far between. The adoption of DP is hindered by limited practical guidance of what DP protection entails, what privacy guarantees to aim for, and the difficulty of achieving good privacy-utility-computation trade-offs for ML models. Tricks for tuning and maximizing performance are scattered among papers or stored in the heads of practitioners, particularly with respect to the challenging task of hyperparameter tuning. Furthermore, the literature seems to present conflicting evidence on how and whether to apply architectural adjustments and which components are “safe” to use with DP. In this survey paper, we attempt to create a self-contained guide that gives an in-depth overview of the field of DP ML. We aim to assemble information about achieving the best possible DP ML model with rigorous privacy guarantees. Our target audience is both researchers and practitioners. Researchers interested in DP for ML will benefit from a clear overview of current advances and areas for improvement. We also include theory-focused sections that highlight important topics such as privacy accounting and convergence. For a practitioner, this survey provides a background in DP theory and a clear step-by-step guide for choosing an appropriate privacy definition and approach, implementing DP training, potentially updating the model architecture, and tuning hyperparameters. For both researchers and practitioners, consistently and fully reporting privacy guarantees is critical, so we propose a set of specific best practices for stating guarantees. With sufficient computation and a sufficiently large training set or supplemental nonprivate data, both good accuracy (that is, almost as good as a non-private model) and good privacy can often be achievable. And even when computation and dataset size are limited, there are advantages to training with even a weak (but still finite) formal DP guarantee. Hence, we hope this work will facilitate more widespread deployments of DP ML models.

The structural tuning of the convolutional neural network forspeaker identification in mel frequency cepstrumcoefficients space

As a result of the literature analysis, the main methods for speaker identification from speech signals were defined. These are statistical methods based on Gaussian mixture model and a universal background model, as well as neural network methods, in particular, using convolutional or Siamese neural networks. The main characteristics of these methods are the recognition performance, a number of parameters, and the training time. High recognition performance is achieved by using convolutional neural networks, but a number of parameters of these networks are much higher than for statistical methods, although lower than for Siamese neural networks. A large number of parameters require a large training set, which is not always available for the researcher. In addition, despite the effectiveness of convolutional neural networks, model size and inference efficiency remain important fordevices with a limited source of computing power, such as peripheral or mobile devices. Therefore, the aspects of tuning of the structure of existing convolutional neural networks are relevant for research. In this work, we have performed a structural tuning of an existing convolutional neural network based on the VGGNet architecture for speaker identification in the space of mel frequency cepstrum coefficients. The aim of the work was to reduce the number of neural network parameters and, as a result, to reduce the network training time, provided that the recognition performance is sufficient (the correct recognition is above 95 %). The neural network proposed as a result of structural tuning has fewer layers than the architecture of the basic neural network. Insteadof the ReLU activation function, the related Leaky ReLU function with a parameter of 0.1 was used. The number of filters and the size of kernels in convolutional layers are changed. The size of kernels for the max pooling layer has been increased. It is proposedto use the averaging of the results of each convolution to input a two-dimensional convolution results to a fully connected layer with the Softmax activation function. The performed experiment showed that the number of parameters of the proposed neural network is 29 % less than the number of parameters of the basic neural network, provided that the speaker recognition performance is almost the same. In addition, the training time of the proposed and basic neural networks was evaluated on five datasets of audio recordings corresponding to different numbers of speakers. The training time of the proposed network was reduced by 10-39 % compared to the basic neural network. The results of the research show the advisability of the structural tuning of the convolutional neural network for devices with a limited source of computing, namely, peripheral or mobile devices.

Few-shot and meta-learning methods for image understanding: a survey

State-of-the-art deep learning systems (e.g., ImageNet image classification) typically require very large training sets to achieve high accuracies. Therefore, one of the grand challenges is called few-shot learning where only a few training samples are required for good performance. In this survey, we illuminate one of the key paradigms in few-shot learning called meta-learning. These meta-learning methods, by simulating the tasks which will be presented at inference through episodic training, can effectively employ previous prior knowledge to guide the learning of new tasks. In this paper, we provide a comprehensive overview and key insights into the meta-learning approaches and categorize them into three branches according to their technical characteristics, namely metric-based, model-based and optimization-based meta-learning. Due to the major importance of the evaluation process, we also present an overview of current widely used benchmarks, as well as performances of recent meta-learning methods on these datasets. Based on over 200 papers in this survey, we conclude with the major challenges and future directions of few-shot learning and meta-learning.

Recovering Word Forms by Context for Morphologically Rich Languages

In this work, we focus on “sentence-level unlemmatization,” the task of generating a grammatical sentence given a lemmatized one; this task is usually easy to do for humans but may present problems for machine learning models. We treat this setting as a machine translation problem and, as a first try, apply a sequence-to-sequence model to the texts of Russian Wikipedia articles, evaluate the effect of the different training sets sizes quantitatively and achieve the BLUE score of 67, 3 using the largest training set available. We discuss preliminary results and flaws of traditional machine translation evaluation methods for this task and suggest directions for future research.

O-243 Personalized pregnancy odds estimation can be obtained using AI embryo evaluation and individual patient characteristics

Abstract Study question Can EMATM (AIVF, Israel) artificial intelligence (AI) platform provide personalized success estimates based on the patient’s metadata and embryonic development? Summary answer Individual patients can be given an accurate estimation of their chances for a clinical pregnancy using AI-based embryo evaluation and patient metadata. What is known already: AI models for embryo evaluation are trained on diverse datasets using data from multiple clinics and from patients with varying ages and clinical history. Precision medicine can be attained by AI models that provide individual patients with personalized success rates per embryo transfer based on their characteristics. Study design, size, duration A large dataset (9,812 embryos) from 3 geographically diverse IVF Clinics (Israel, Spain, USA) with known clinical pregnancy (fetal heartbeat) and patient characteristics obtained from Electronic Medical Records (EMR) were used to evaluate the importance of individual features contributing to pregnancy. Participants/materials, setting, methods Machine learning models were trained to predict clinical pregnancy on a large training and feature set which combines AI scores (EMATM, AIVF) with additional information obtained from EMR systems and intrinsic features that can be derived from the embryo cohort. The importance of each one of the features for the prediction ability of the models was evaluated. Main results and the role of chance AI embryo score, maternal age and cohort features were found to be the major contributing factors for the prediction models, thus improving the accuracy of the estimates for pregnancy probability. Significant contributing factors were: cohort size, number of viable embryos, patient age, number of past treatments, BMI and EMA score. The AUC of the AI embryo score model was 0.7, when factoring in the patient metadata and the cohort features, the AUC of the combined model was 0.75. Limitations, reasons for caution This study is limited by its retrospective design. A prospective study is needed to validate the results. Wider implications of the findings This study validates the use of AI-based scores for embryo evaluation, for relative grading of the embryos within the cohort, and also to estimate the true pregnancy odds of each embryo based on individual patient features. This can provide a superior decision support tool for doctors, embryologists, and patients. Trial registration number not applicable

Task-specific self-supervision for CT image denoising

ABSTRACT CT image quality is largely reliant on radiation dose, which causes a trade-off between image quality and dose, affecting the subsequent image-based diagnostic and treatment performances. Deep learning approaches for low-dose CT image denoising require access to large training sets and specifically the reference full-dose images, which can be difficult to obtain. Self-supervised learning enables learning with reduced reference data burden. Currently available self-supervised CT denoising works are either dependent on foreign domains or pretexts that are not very task-relevant, requiring lots of adjustments (architectural, loss function, training parameters, etc.) and additional skills to perform the downstream denoising task. To tackle the aforementioned challenges, we propose a novel self-supervised pretraining approach, namely Self-Supervised Window-Leveling for Image DeNoising (SSWL-IDN), leveraging an innovative, task-relevant, simple yet effective surrogate – prediction of the window-leveled equivalent. SSWL-IDN leverages residual learning and a hybrid loss combining perceptual loss and MSE, all incorporated in a VAE framework. Our extensive (in- and cross-domain) experimentation demonstrates the effectiveness of SSWL-IDN in aggressive denoising of CT (abdomen and chest) images acquired at two different dose levels (5% and 25%), without requiring any kind of architectural or downstream training adjustments. We have made our code publicly available at https://github.com/ayaanzhaque/SSWL-IDN

Learning system parameters from turing patterns.

The Turing mechanism describes the emergence of spatial patterns due to spontaneous symmetry breaking in reaction-diffusion processes and underlies many developmental processes. Identifying Turing mechanisms in biological systems defines a challenging problem. This paper introduces an approach to the prediction of Turing parameter values from observed Turing patterns. The parameter values correspond to a parametrized system of reaction-diffusion equations that generate Turing patterns as steady state. The Gierer-Meinhardt model with four parameters is chosen as a case study. A novel invariant pattern representation based on resistance distance histograms is employed, along with Wasserstein kernels, in order to cope with the highly variable arrangement of local pattern structure that depends on the initial conditions which are assumed to be unknown. This enables us to compute physically plausible distances between patterns, to compute clusters of patterns and, above all, model parameter prediction based on training data that can be generated by numerical model evaluation with random initial data: for small training sets, classical state-of-the-art methods including operator-valued kernels outperform neural networks that are applied to raw pattern data, whereas for large training sets the latter are more accurate. A prominent property of our approach is that only a single pattern is required as input data for model parameter predicion. Excellent predictions are obtained for single parameter values and reasonably accurate results for jointly predicting all four parameter values.

3D data-augmentation methods for semantic segmentation of tomato plant parts.

3D semantic segmentation of plant point clouds is an important step towards automatic plant phenotyping and crop modeling. Since traditional hand-designed methods for point-cloud processing face challenges in generalisation, current methods are based on deep neural network that learn to perform the 3D segmentation based on training data. However, these methods require a large annotated training set to perform well. Especially for 3D semantic segmentation, the collection of training data is highly labour intensitive and time consuming. Data augmentation has been shown to improve training on small training sets. However, it is unclear which data-augmentation methods are effective for 3D plant-part segmentation. In the proposed work, five novel data-augmentation methods (global cropping, brightness adjustment, leaf translation, leaf rotation, and leaf crossover) were proposed and compared to five existing methods (online down sampling, global jittering, global scaling, global rotation, and global translation). The methods were applied to PointNet++ for 3D semantic segmentation of the point clouds of three cultivars of tomato plants (Merlice, Brioso, and Gardener Delight). The point clouds were segmented into soil base, stick, stemwork, and other bio-structures. Among the data augmentation methods being proposed in this paper, leaf crossover indicated the most promising result which outperformed the existing ones. Leaf rotation (around Z axis), leaf translation, and cropping also performed well on the 3D tomato plant point clouds, which outperformed most of the existing work apart from global jittering. The proposed 3D data augmentation approaches significantly improve the overfitting caused by the limited training data. The improved plant-part segmentation further enables a more accurate reconstruction of the plant architecture.

Real-Time structure topology optimization using CNN driven Moving Morphable component method

Classical optimization methods require finite element analysis in iterations, which increase the computing time and decrease the algorithmic efficiency. The deep learning model can potentially realize real-time topology optimization design, but it normally requires large training set. This paper presents a real-time topology optimization algorithm based on the Moving Morphable Component (MMC) method using a Convolutional Neural Network (CNN). The optimization algorithm uses a new data pre-processing method, which can preserve the numerical characteristics and smoothness of the structure boundary, hence it can help CNN to capture data features with a limited sample set. The topology optimization boundary information of the optimized result is used as the sample set label to avoid the components dislocation phenomenon. The new algorithm effectiveness has been verified with several examples. The trained model can significantly improve the optimization efficiency of the MMC method and offer accurate results with a clear structure boundary.

Improved automatic kernel construction for Gaussian process regression in small sample learning for predicting lift body aerodynamic performance

A Gaussian process regression (GPR) model based on an improved automatic kernel construction (AKC) algorithm using beam search is proposed to establish a surrogate model between lift body shape parameters and aerodynamic coefficients with various training sets sizes. The precision of our proposed surrogate model is assessed through tenfold cross-validation. The improved AKC-GPR algorithm, polynomial regression, and support vector regression (SVR) are employed to construct the regression model. The interpolation and extrapolation capabilities of the model, as generated by the improved AKC-GPR algorithm, are examined using six shapes beyond the sample set. The results show that the three models perform similarly with a large training set. However, when the training set size is less than 40% sample dataset, the model constructed by the improved AKC-GPR algorithm has better fitting and prediction capabilities than the other models. Specifically, the max relative error of the improved model is one-fourth of that of SVR and one-half of that of polynomial regression with the training set size of 8% of the sample dataset. Furthermore, the lift-to-drag ratio relative error of interpolation is only 3%, and extrapolation error is 6%. In terms of the fitting and prediction abilities for small samples, the lift-to-drag ratio model outperforms the drag coefficient model, while the lift coefficient model performs the poorest. These findings suggest that the proposed AKC-GPR algorithm can be an effective approach for building a surrogate model in the field of aerodynamics.

Calibrating Data Mismatches in Deep Learning-Based Quantitative Ultrasound Using Setting Transfer Functions.

Deep learning (DL) can fail when there are data mismatches between training and testing data distributions. Due to its operator-dependent nature, acquisition-related data mismatches, caused by different scanner settings, can occur in ultrasound imaging. As a result, it is crucial to mitigate the effects of these mismatches to enable wider clinical adoption of DL-powered ultrasound imaging and tissue characterization. To address this challenge, we propose an inexpensive and generalizable method that involves collecting a large training set at a single setting and a small calibration set at each scanner setting. Then, the calibration set will be used to calibrate data mismatches by using a signals and systems perspective. We tested the proposed solution to classify two phantoms using an L9-4 array connected to a SonixOne scanner. To investigate generalizability of the proposed solution, we calibrated three types of data mismatches: pulse frequency mismatch, focus mismatch, and output power mismatch. Two well-known convolutional neural networks (CNNs), i.e., ResNet-50 and DenseNet-201, were trained using the ultrasound radio frequency (RF) data. To calibrate the setting mismatches, we calculated the setting transfer functions. The CNNs trained without calibration resulted in mean classification accuracies of around 52%, 84%, and 85% for pulse frequency, focus, and output power mismatches, respectively. By using the setting transfer functions, which allowed a matching of the training and testing domains, we obtained the mean accuracies of 96%, 96%, and 98%, respectively. Therefore, the incorporation of the setting transfer functions between scanner settings can provide an economical means of generalizing a DL model for specific classification tasks where scanner settings are not fixed by the operator.

Unsupervised deep learning techniques for automatic detection of plant diseases: reducing the need of manual labelling of plant images

Crop protection from diseases through applications of plant protection products is crucial to secure worldwide food production. Nevertheless, sustainable management of plant diseases is an open challenge with a major role in the economic and environmental impact of agricultural activities. A primary contribution is expected to come from precision crop protection approaches, with treatments tailored to spatial and time-specific needs of the crop, in contrast to the current practice of applying treatments uniformly to fields. In view of this, image-based automatic detection of early disease symptoms is considered a key enabling technology for high throughput scouting of the crop, in order to timely target the treatments on emerging infection spots. Thanks to the unprecedented performance in image-recognition problems, Deep Learning (DL) methods based on Convolutional Neural Networks (CNNs) have recently entered the domain of plant disease detection. This work develops two DL approaches for automatic recognition of powdery mildew disease on cucumber leaves, with a specific focus on exploring unsupervised techniques to overcome the need of large training set of manually labelled images. To this aim, autoencoder networks were implemented for unsupervised detection of disease symptoms through: i) clusterization of features in a compressed space; ii) anomaly detection. The two proposed approaches were applied to multispectral images acquired during in-vivo experiments, and the obtained results were assessed by quantitative indices. The clusterization approach showed only partially capability to provide accurate disease detection, even if it gathered some relevant information. Anomaly detection showed instead to possess a significant potential of discrimination which could be further exploited as a prior step to train more powerful supervised architectures with a very limited number of labelled samples.

USING SNN NEURAL NETWORKS TRAINED WITH HIGH RESOLUTION DATA 2 AND APPLIED TO COPERNICUS SENTINEL-2 DATA

Abstract. Data from the Copernicus project are proving to be fundamental in so many areas of spatial governance, and in particular in the context of safeguarding the cultural and natural heritage. This is mainly due to their dissemination mode, which allows them to be used by a countless number of institutions and organizations. With regard to their use in order to create effective supervised-training AI (Artificial Intelligence) classifiers (useful for continuous land-use monitoring), their low resolution could be a problem in terms of accuracy or in any case during the training process. This process which could be very costly (large training sets and high-end hardware required). In this paper, a methodology is tested that creates an AI classifier from high-resolution training data, Remote Sensing Image Classification Benchmark (RSI-CB128) and using Self-Normalising Neural Networks (SNNs). The efficiency of the Neural Network is, however, measured with a test dataset constructed with Raster Sentinel-2. In other words, the aim is to assess the goodness of an AI classifier on Sentinel-2 images built from Rasters generated by high-resolution sensors. Therefore, there seem to be clear advantages over the classification methodologies in use today.

Noise-resilient deep learning for integrated circuit tomography.

X-ray tomography is a non-destructive imaging technique that reveals the interior of an object from its projections at different angles. Under sparse-view and low-photon sampling, regularization priors are required to retrieve a high-fidelity reconstruction. Recently, deep learning has been used in X-ray tomography. The prior learned from training data replaces the general-purpose priors in iterative algorithms, achieving high-quality reconstructions with a neural network. Previous studies typically assume the noise statistics of test data are acquired a priori from training data, leaving the network susceptible to a change in the noise characteristics under practical imaging conditions. In this work, we propose a noise-resilient deep-reconstruction algorithm and apply it to integrated circuit tomography. By training the network with regularized reconstructions from a conventional algorithm, the learned prior shows strong noise resilience without the need for additional training with noisy examples, and allows us to obtain acceptable reconstructions with fewer photons in test data. The advantages of our framework may further enable low-photon tomographic imaging where long acquisition times limit the ability to acquire a large training set.