Fluorescence lifetime imaging microscopy (FLIM) systems are limited by their slow processing speed, low signal-to-noise ratio (SNR), and expensive and challenging hardware setups. In this work, we demonstrate applying a denoising convolutional network to improve FLIM SNR. The network will be integrated with an instant FLIM system with fast data acquisition based on analog signal processing, high SNR using high-efficiency pulse-modulation, and cost-effective implementation utilizing off-the-shelf radio-frequency components. Our instant FLIM system simultaneously provides the intensity, lifetime, and phasor plots \textit{in vivo} and \textit{ex vivo}. By integrating image denoising using the trained deep learning model on the FLIM data, provide accurate FLIM phasor measurements are obtained. The enhanced phasor is then passed through the K-means clustering segmentation method, an unbiased and unsupervised machine learning technique to separate different fluorophores accurately. Our experimental \textit{in vivo} mouse kidney results indicate that introducing the deep learning image denoising model before the segmentation effectively removes the noise in the phasor compared to existing methods and provides clearer segments. Hence, the proposed deep learning-based workflow provides fast and accurate automatic segmentation of fluorescence images using instant FLIM. The denoising operation is effective for the segmentation if the FLIM measurements are noisy. The clustering can effectively enhance the detection of biological structures of interest in biomedical imaging applications.
Sign language is the primary language for people with a hearing loss. Sign language recognition (SLR) is the automatic recognition of sign language, which represents a challenging problem for computers, though some progress has been made recently using deep learning. Huge amounts of data are generally required to train deep learning models. However, corresponding datasets are missing for the majority of sign languages. Transfer learning is a technique to utilize a related task with an abundance of data available to help solve a target task lacking sufficient data. Transfer learning has been applied highly successfully in computer vision and natural language processing. However, much less research has been conducted in the field of SLR. This paper investigates how effectively transfer learning can be applied to isolated SLR using an inflated 3D convolutional neural network as the deep learning architecture. Transfer learning is implemented by pre-training a network on the American Sign Language dataset MS-ASL and subsequently fine-tuning it separately on three different sizes of the German Sign Language dataset SIGNUM. The results of the experiments give clear empirical evidence that transfer learning can be effectively applied to isolated SLR. The accuracy performances of the networks applying transfer learning increased substantially by up to 21% as compared to the baseline models that were not pre-trained on the MS-ASL dataset.
Fluorescence microscopy has enabled a dramatic development in modern biology by visualizing biological organisms with micrometer scale resolution. However, due to the diffraction limit, sub-micron/nanometer features are difficult to resolve. While various super-resolution techniques are developed to achieve nanometer-scale resolution, they often either require expensive optical setup or specialized fluorophores. In recent years, deep learning has shown the potentials to reduce the technical barrier and obtain super-resolution from diffraction-limited images. For accurate results, conventional deep learning techniques require thousands of images as a training dataset. Obtaining large datasets from biological samples is not often feasible due to the photobleaching of fluorophores, phototoxicity, and dynamic processes occurring within the organism. Therefore, achieving deep learning-based super-resolution using small datasets is challenging. We address this limitation with a new convolutional neural network-based approach that is successfully trained with small datasets and achieves super-resolution images. We captured 750 images in total from 15 different field-of-views as the training dataset to demonstrate the technique. In each FOV, a single target image is generated using the super-resolution radial fluctuation method. As expected, this small dataset failed to produce a usable model using traditional super-resolution architecture. However, using the new approach, a network can be trained to achieve super-resolution images from this small dataset. This deep learning model can be applied to other biomedical imaging modalities such as MRI and X-ray imaging, where obtaining large training datasets is challenging.
Over the past decade, unprecedented progress in the development of neural networks influenced dozens of different industries, including weed recognition in the agro-industrial sector. The use of neural networks in agro-industrial activity in the task of recognizing cultivated crops is a new direction. The absence of any standards significantly complicates the understanding of the real situation of the use of the neural network in the agricultural sector. The manuscript presents the complete analysis of researches over the past 10 years on the use of neural networks for the classification and tracking of weeds due to neural networks. In particular, the analysis of the results of using various neural network algorithms for the task of classification and tracking was presented. As a result, we presented the recommendation for the use of neural networks in the tasks of recognizing a cultivated object and weeds. Using this standard can significantly improve the quality of research on this topic and simplify the analysis and understanding of any paper.
The boom of DL technology leads to massive DL models built and shared, which facilitates the acquisition and reuse of DL models. For a given task, we encounter multiple DL models available with the same functionality, which are considered as candidates to achieve this task. Testers are expected to compare multiple DL models and select the more suitable ones w.r.t. the whole testing context. Due to the limitation of labeling effort, testers aim to select an efficient subset of samples to make an as precise rank estimation as possible for these models. To tackle this problem, we propose Sample Discrimination based Selection (SDS) to select efficient samples that could discriminate multiple models, i.e., the prediction behaviors (right/wrong) of these samples would be helpful to indicate the trend of model performance. To evaluate SDS, we conduct an extensive empirical study with three widely-used image datasets and 80 real world DL models. The experimental results show that, compared with state-of-the-art baseline methods, SDS is an effective and efficient sample selection method to rank multiple DL models.
We study low-rank parameterizations of weight matrices with embedded spectral properties in the Deep Learning context. The low-rank property leads to parameter efficiency and permits taking computational shortcuts when computing mappings. Spectral properties are often subject to constraints in optimization problems, leading to better models and stability of optimization. We start by looking at the compact SVD parameterization of weight matrices and identifying redundancy sources in the parameterization. We further apply the Tensor Train (TT) decomposition to the compact SVD components, and propose a non-redundant differentiable parameterization of fixed TT-rank tensor manifolds, termed the Spectral Tensor Train Parameterization (STTP). We demonstrate the effects of neural network compression in the image classification setting and both compression and improved training stability in the generative adversarial training setting.
We present the first edition of "VIPriors: Visual Inductive Priors for Data-Efficient Deep Learning" challenges. We offer four data-impaired challenges, where models are trained from scratch, and we reduce the number of training samples to a fraction of the full set. Furthermore, to encourage data efficient solutions, we prohibited the use of pre-trained models and other transfer learning techniques. The majority of top ranking solutions make heavy use of data augmentation, model ensembling, and novel and efficient network architectures to achieve significant performance increases compared to the provided baselines.
The generalization performance of deep learning models for medical image analysis often decreases on images collected with different devices for data acquisition, device settings, or patient population. A better understanding of the generalization capacity on new images is crucial for clinicians' trustworthiness in deep learning. Although significant research efforts have been recently directed toward establishing generalization bounds and complexity measures, still, there is often a significant discrepancy between the predicted and actual generalization performance. As well, related large empirical studies have been primarily based on validation with general-purpose image datasets. This paper presents an empirical study that investigates the correlation between 25 complexity measures and the generalization abilities of supervised deep learning classifiers for breast ultrasound images. The results indicate that PAC-Bayes flatness-based and path norm-based measures produce the most consistent explanation for the combination of models and data. We also investigate the use of multi-task classification and segmentation approach for breast images, and report that such learning approach acts as an implicit regularizer and is conducive toward improved generalization.
This paper adds to the fundamental body of work on benchmarking the robustness of deep learning (DL) classifiers. We innovate a new benchmarking methodology to evaluate robustness of DL classifiers. Also, we introduce a new four-quadrant statistical visualization tool, including minimum accuracy, maximum accuracy, mean accuracy, and coefficient of variation, for benchmarking robustness of DL classifiers. To measure robust DL classifiers, we created a comprehensive 69 benchmarking image set, including a clean set, sets with single factor perturbations, and sets with two-factor perturbation conditions. After collecting experimental results, we first report that using two-factor perturbed images improves both robustness and accuracy of DL classifiers. The two-factor perturbation includes (1) two digital perturbations (salt & pepper noise and Gaussian noise) applied in both sequences, and (2) one digital perturbation (salt & pepper noise) and a geometric perturbation (rotation) applied in both sequences. All source codes, related image sets, and preliminary data, figures are shared on a GitHub website to support future academic research and industry projects. The web resources locate at https://github.com/caperock/robustai
Deep learning has successfully been leveraged for medical image segmentation. It employs convolutional neural networks (CNN) to learn distinctive image features from a defined pixel-wise objective function. However, this approach can lead to less output pixel interdependence producing incomplete and unrealistic segmentation results. In this paper, we present a fully automatic deep learning method for robust medical image segmentation by formulating the segmentation problem as a recurrent framework using two systems. The first one is a forward system of an encoder-decoder CNN that predicts the segmentation result from the input image. The predicted probabilistic output of the forward system is then encoded by a fully convolutional network (FCN)-based context feedback system. The encoded feature space of the FCN is then integrated back into the forward system's feed-forward learning process. Using the FCN-based context feedback loop allows the forward system to learn and extract more high-level image features and fix previous mistakes, thereby improving prediction accuracy over time. Experimental results, performed on four different clinical datasets, demonstrate our method's potential application for single and multi-structure medical image segmentation by outperforming the state of the art methods. With the feedback loop, deep learning methods can now produce results that are both anatomically plausible and robust to low contrast images. Therefore, formulating image segmentation as a recurrent framework of two interconnected networks via context feedback loop can be a potential method for robust and efficient medical image analysis.
Accelerated multi-coil magnetic resonance imaging reconstruction has seen a substantial recent improvement combining compressed sensing with deep learning. However, most of these methods rely on estimates of the coil sensitivity profiles, or on calibration data for estimating model parameters. Prior work has shown that these methods degrade in performance when the quality of these estimators are poor or when the scan parameters differ from the training conditions. Here we introduce Deep J-Sense as a deep learning approach that builds on unrolled alternating minimization and increases robustness: our algorithm refines both the magnetization (image) kernel and the coil sensitivity maps. Experimental results on a subset of the knee fastMRI dataset show that this increases reconstruction performance and provides a significant degree of robustness to varying acceleration factors and calibration region sizes.
Understanding how the brain responds to sensory inputs is challenging: brain recordings are partial, noisy, and high dimensional; they vary across sessions and subjects and they capture highly nonlinear dynamics. These challenges have led the community to develop a variety of preprocessing and analytical (almost exclusively linear) methods, each designed to tackle one of these issues. Instead, we propose to address these challenges through a specific end-to-end deep learning architecture, trained to predict the brain responses of multiple subjects at once. We successfully test this approach on a large cohort of magnetoencephalography (MEG) recordings acquired during a one-hour reading task. Our Deep Recurrent Encoding (DRE) architecture reliably predicts MEG responses to words with a three-fold improvement over classic linear methods. To overcome the notorious issue of interpretability of deep learning, we describe a simple variable importance analysis. When applied to DRE, this method recovers the expected evoked responses to word length and word frequency. The quantitative improvement of the present deep learning approach paves the way to better understand the nonlinear dynamics of brain activity from large datasets.
In this work, we investigate different machine learning-based strategies for denoising raw simulation data from the ProtoDUNE experiment. The ProtoDUNE detector is hosted by CERN and it aims to test and calibrate the technologies for DUNE, a forthcoming experiment in neutrino physics. The reconstruction workchain consists of converting digital detector signals into physical high-level quantities. We address the first step in reconstruction, namely raw data denoising, leveraging deep learning algorithms. We design two architectures based on graph neural networks, aiming to enhance the receptive field of basic convolutional neural networks. We benchmark this approach against traditional algorithms implemented by the DUNE collaboration. We test the capabilities of graph neural network hardware accelerator setups to speed up training and inference processes.
Deep learning-based coastline detection algorithms have begun to outshine traditional statistical methods in recent years. However, they are usually trained only as single-purpose models to either segment land and water or delineate the coastline. In contrast to this, a human annotator will usually keep a mental map of both segmentation and delineation when performing manual coastline detection. To take into account this task duality, we therefore devise a new model to unite these two approaches in a deep learning model. By taking inspiration from the main building blocks of a semantic segmentation framework (UNet) and an edge detection framework (HED), both tasks are combined in a natural way. Training is made efficient by employing deep supervision on side predictions at multiple resolutions. Finally, a hierarchical attention mechanism is introduced to adaptively merge these multiscale predictions into the final model output. The advantages of this approach over other traditional and deep learning-based methods for coastline detection are demonstrated on a dataset of Sentinel-1 imagery covering parts of the Antarctic coast, where coastline detection is notoriously difficult. An implementation of our method is available at \url{https://github.com/khdlr/HED-UNet}.
Counting cells in fluorescent microscopy is a tedious, time-consuming task that researchers have to accomplish to assess the effects of different experimental conditions on biological structures of interest. Although such objects are generally easy to identify, the process of manually annotating cells is sometimes subject to arbitrariness due to the operator's interpretation of the borderline cases. We propose a Machine Learning approach that exploits a fully-convolutional network in a binary segmentation fashion to localize the objects of interest. Counts are then retrieved as the number of detected items. Specifically, we adopt a UNet-like architecture leveraging residual units and an extended bottleneck for enlarging the field-of-view. In addition, we make use of weighted maps that penalize the errors on cells boundaries increasingly with overcrowding. These changes provide more context and force the model to focus on relevant features during pixel-wise classification. As a result, the model performance is enhanced, especially in presence of clumping cells, artifacts and confounding biological structures. Posterior assessment of the results with domain experts confirms that the model detects cells of interest correctly. The model demonstrates a human-level ability inasmuch even erroneous predictions seem to fall within the limits of operator interpretation. This qualitative assessment is also corroborated by quantitative metrics as an ${F_1}$ score of 0.87. Despite some difficulties in interpretation, results are also satisfactory with respect to the counting task, as testified by mean and median absolute error of, respectively, 0.8 and 1.
We consider training models with differential privacy (DP) using mini-batch gradients. The existing state-of-the-art, Differentially Private Stochastic Gradient Descent (DP-SGD), requires privacy amplification by sampling or shuffling to obtain the best privacy/accuracy/computation trade-offs. Unfortunately, the precise requirements on exact sampling and shuffling can be hard to obtain in important practical scenarios, particularly federated learning (FL). We design and analyze a DP variant of Follow-The-Regularized-Leader (DP-FTRL) that compares favorably (both theoretically and empirically) to amplified DP-SGD, while allowing for much more flexible data access patterns. DP-FTRL does not use any form of privacy amplification. The code is available at https://github.com/google-research/federated/tree/master/dp_ftrl and https://github.com/google-research/DP-FTRL .
This article presents the first results from using a learning classifier system capable of performing adaptive computation with deep neural networks. Individual classifiers within the population are composed of two neural networks. The first acts as a gating or guarding component, which enables the conditional computation of an associated deep neural network on a per instance basis. Self-adaptive mutation is applied upon reproduction and prediction networks are refined with stochastic gradient descent during lifetime learning. The use of fully-connected and convolutional layers are evaluated on handwritten digit recognition tasks where evolution adapts (i) the gradient descent learning rate applied to each layer (ii) the number of units within each layer, i.e., the number of fully-connected neurons and the number of convolutional kernel filters (iii) the connectivity of each layer, i.e., whether each weight is active (iv) the weight magnitudes, enabling escape from local optima. The system automatically reduces the number of weights and units while maintaining performance after achieving a maximum prediction error.
Graph neural networks (GNNs) have attracted tremendous attention from the graph learning community in recent years. It has been widely adopted in various real-world applications from diverse domains, such as social networks and biological graphs. The research and applications of graph deep learning present new challenges, including the sparse nature of graph data, complicated training of GNNs, and non-standard evaluation of graph tasks. To tackle the issues, we present CogDL, a comprehensive library for graph deep learning that allows researchers and practitioners to conduct experiments, compare methods, and build applications with ease and efficiency. In CogDL, we propose a unified design for the training and evaluation of GNN models for various graph tasks, making it unique among existing graph learning libraries. By utilizing this unified trainer, CogDL can optimize the GNN training loop with several training techniques, such as mixed precision training. Moreover, we develop efficient sparse operators for CogDL, enabling it to become the most competitive graph library for efficiency. Another important CogDL feature is its focus on ease of use with the aim of facilitating open and reproducible research of graph learning. We leverage CogDL to report and maintain benchmark results on fundamental graph tasks, which can be reproduced and directly used by the community.
We present Tails, an open-source deep-learning framework for the identification and localization of comets in the image data of the Zwicky Transient Facility (ZTF), a robotic optical time-domain survey currently in operation at the Palomar Observatory in California, USA. Tails employs a custom EfficientDet-based architecture and is capable of finding comets in single images in near real time, rather than requiring multiple epochs as with traditional methods. The system achieves state-of-the-art performance with 99% recall, 0.01% false positive rate, and 1-2 pixel root mean square error in the predicted position. We report the initial results of the Tails efficiency evaluation in a production setting on the data of the ZTF Twilight survey, including the first AI-assisted discovery of a comet (C/2020 T2) and the recovery of a comet (P/2016 J3 = P/2021 A3).
Deep learning has been the engine powering many successes of data science. However, the deep neural network (DNN), as the basic model of deep learning, is often excessively over-parameterized, causing many difficulties in training, prediction and interpretation. We propose a frequentist-like method for learning sparse DNNs and justify its consistency under the Bayesian framework: the proposed method could learn a sparse DNN with at most $O(n/\log(n))$ connections and nice theoretical guarantees such as posterior consistency, variable selection consistency and asymptotically optimal generalization bounds. In particular, we establish posterior consistency for the sparse DNN with a mixture Gaussian prior, show that the structure of the sparse DNN can be consistently determined using a Laplace approximation-based marginal posterior inclusion probability approach, and use Bayesian evidence to elicit sparse DNNs learned by an optimization method such as stochastic gradient descent in multiple runs with different initializations. The proposed method is computationally more efficient than standard Bayesian methods for large-scale sparse DNNs. The numerical results indicate that the proposed method can perform very well for large-scale network compression and high-dimensional nonlinear variable selection, both advancing interpretable machine learning.