This paper aims to clarify the representational status of Deep Learning Models (DLMs). While commonly referred to as 'representations', what this entails is ambiguous due to a conflation of functional and relational conceptions of representation. This paper argues that while DLMs represent their targets in a relational sense, in general, we have no good reason to believe that DLMs encode locally semantically decomposable representations of their targets. That is, the representational capacity these models have is largely global, rather than decomposable into stable, local subrepresentations. This result has immediate implications for explainable AI (XAI) and directs attention toward exploring the global relational nature of deep learning representations and their relationship both to models more generally to understand their potential role in future scientific inquiry.
As human-machine voice interfaces provide easy access to increasingly intelligent machines, many state-of-the-art automatic speech recognition (ASR) systems are proposed. However, commercial ASR systems usually have poor performance on domain-specific speech especially under low-resource settings. The author works with pre-trained DeepSpeech2 and Wav2Vec2 acoustic models to develop benefit-specific ASR systems. The domain-specific data are collected using proposed semi-supervised learning annotation with little human intervention. The best performance comes from a fine-tuned Wav2Vec2-Large-LV60 acoustic model with an external KenLM, which surpasses the Google and AWS ASR systems on benefit-specific speech. The viability of using error prone ASR transcriptions as part of spoken language understanding (SLU) is also investigated. Results of a benefit-specific natural language understanding (NLU) task show that the domain-specific fine-tuned ASR system can outperform the commercial ASR systems even when its transcriptions have higher word error rate (WER), and the results between fine-tuned ASR and human transcriptions are similar.
Deep learning has emerged as a strong alternative for classical iterative methods for deformable medical image registration, where the goal is to find a mapping between the coordinate systems of two images. Popular classical image registration methods enforce the useful inductive biases of symmetricity, inverse consistency, and topology preservation by construction. However, while many deep learning registration methods encourage these properties via loss functions, no earlier methods enforce all of them by construction. Here, we propose a novel registration architecture based on extracting multi-resolution feature representations which is by construction symmetric, inverse consistent, and topology preserving. We also develop an implicit layer for memory efficient inversion of the deformation fields. Our method achieves state-of-the-art registration accuracy on three datasets. The code is available at https://github.com/honkamj/SITReg.
Undersampled MRI reconstruction is crucial for accelerating clinical scanning. Dual-domain reconstruction network is performant among SoTA deep learning methods. In this paper, we rethink dual-domain model design from the perspective of the receptive field, which is needed for image recovery and K-space interpolation problems. Further, we introduce domain-specific modules for dual-domain reconstruction, namely k-space global initialization and image-domain parallel local detail enhancement. We evaluate our modules by translating a SoTA method DuDoRNet under different conventions of MRI reconstruction including image-domain, dual-domain, and reference-guided reconstruction on the public IXI dataset. Our model DuDoRNet+ achieves significant improvements over competing deep learning methods.
Camera calibration involves estimating camera parameters to infer geometric features from captured sequences, which is crucial for computer vision and robotics. However, conventional calibration is laborious and requires dedicated collection. Recent efforts show that learning-based solutions have the potential to be used in place of the repeatability works of manual calibrations. Among these solutions, various learning strategies, networks, geometric priors, and datasets have been investigated. In this paper, we provide a comprehensive survey of learning-based camera calibration techniques, by analyzing their strengths and limitations. Our main calibration categories include the standard pinhole camera model, distortion camera model, cross-view model, and cross-sensor model, following the research trend and extended applications. As there is no benchmark in this community, we collect a holistic calibration dataset that can serve as a public platform to evaluate the generalization of existing methods. It comprises both synthetic and real-world data, with images and videos captured by different cameras in diverse scenes. Toward the end of this paper, we discuss the challenges and provide further research directions. To our knowledge, this is the first survey for the learning-based camera calibration (spanned 8 years). The summarized methods, datasets, and benchmarks are available and will be regularly updated at https://github.com/KangLiao929/Awesome-Deep-Camera-Calibration.
The accelerated development of machine learning methods, primarily deep learning, are causal to the recent breakthroughs in medical image analysis and computer aided intervention. The resource consumption of deep learning models in terms of amount of training data, compute and energy costs are known to be massive. These large resource costs can be barriers in deploying these models in clinics, globally. To address this, there are cogent efforts within the machine learning community to introduce notions of resource efficiency. For instance, using quantisation to alleviate memory consumption. While most of these methods are shown to reduce the resource utilisation, they could come at a cost in performance. In this work, we probe into the trade-off between resource consumption and performance, specifically, when dealing with models that are used in critical settings such as in clinics.
The question of what makes a data distribution suitable for deep learning is a fundamental open problem. Focusing on locally connected neural networks (a prevalent family of architectures that includes convolutional and recurrent neural networks as well as local self-attention models), we address this problem by adopting theoretical tools from quantum physics. Our main theoretical result states that a certain locally connected neural network is capable of accurate prediction over a data distribution if and only if the data distribution admits low quantum entanglement under certain canonical partitions of features. As a practical application of this result, we derive a preprocessing method for enhancing the suitability of a data distribution to locally connected neural networks. Experiments with widespread models over various datasets demonstrate our findings. We hope that our use of quantum entanglement will encourage further adoption of tools from physics for formally reasoning about the relation between deep learning and real-world data.
Background: In medical imaging, images are usually treated as deterministic, while their uncertainties are largely underexplored. Purpose: This work aims at using deep learning to efficiently estimate posterior distributions of imaging parameters, which in turn can be used to derive the most probable parameters as well as their uncertainties. Methods: Our deep learning-based approaches are based on a variational Bayesian inference framework, which is implemented using two different deep neural networks based on conditional variational auto-encoder (CVAE), CVAE-dual-encoder and CVAE-dual-decoder. The conventional CVAE framework, i.e., CVAE-vanilla, can be regarded as a simplified case of these two neural networks. We applied these approaches to a simulation study of dynamic brain PET imaging using a reference region-based kinetic model. Results: In the simulation study, we estimated posterior distributions of PET kinetic parameters given a measurement of time-activity curve. Our proposed CVAE-dual-encoder and CVAE-dual-decoder yield results that are in good agreement with the asymptotically unbiased posterior distributions sampled by Markov Chain Monte Carlo (MCMC). The CVAE-vanilla can also be used for estimating posterior distributions, although it has an inferior performance to both CVAE-dual-encoder and CVAE-dual-decoder. Conclusions: We have evaluated the performance of our deep learning approaches for estimating posterior distributions in dynamic brain PET. Our deep learning approaches yield posterior distributions, which are in good agreement with unbiased distributions estimated by MCMC. All these neural networks have different characteristics and can be chosen by the user for specific applications. The proposed methods are general and can be adapted to other problems.
Many natural and man-made systems are prone to critical transitions -- abrupt and potentially devastating changes in dynamics. Deep learning classifiers can provide an early warning signal (EWS) for critical transitions by learning generic features of bifurcations (dynamical instabilities) from large simulated training data sets. So far, classifiers have only been trained to predict continuous-time bifurcations, ignoring rich dynamics unique to discrete-time bifurcations. Here, we train a deep learning classifier to provide an EWS for the five local discrete-time bifurcations of codimension-1. We test the classifier on simulation data from discrete-time models used in physiology, economics and ecology, as well as experimental data of spontaneously beating chick-heart aggregates that undergo a period-doubling bifurcation. The classifier outperforms commonly used EWS under a wide range of noise intensities and rates of approach to the bifurcation. It also predicts the correct bifurcation in most cases, with particularly high accuracy for the period-doubling, Neimark-Sacker and fold bifurcations. Deep learning as a tool for bifurcation prediction is still in its nascence and has the potential to transform the way we monitor systems for critical transitions.
We develop and test new machine learning strategies for accelerating molecular crystal structure ranking and crystal property prediction using tools from geometric deep learning on molecular graphs. Leveraging developments in graph-based learning and the availability of large molecular crystal datasets, we train models for density prediction and stability ranking which are accurate, fast to evaluate, and applicable to molecules of widely varying size and composition. Our density prediction model, MolXtalNet-D, achieves state of the art performance, with lower than 2% mean absolute error on a large and diverse test dataset. Our crystal ranking tool, MolXtalNet-S, correctly discriminates experimental samples from synthetically generated fakes and is further validated through analysis of the submissions to the Cambridge Structural Database Blind Tests 5 and 6. Our new tools are computationally cheap and flexible enough to be deployed within an existing crystal structure prediction pipeline both to reduce the search space and score/filter crystal candidates.
The morphology of retinal blood vessels can indicate various diseases in the human body, and researchers have been working on automatic scanning and segmentation of retinal images to aid diagnosis. This project compares the performance of four neural network architectures in segmenting retinal images, using a combined dataset from different databases, namely the UNet, DR-VNet, UNet-ResNet and UNet-VGG.
Improving the interpretability of geospatial artificial intelligence (GeoAI) models has become critically important to open the "black box" of complex AI models, such as deep learning. This paper compares popular saliency map generation techniques and their strengths and weaknesses in interpreting GeoAI and deep learning models' reasoning behaviors, particularly when applied to geospatial analysis and image processing tasks. We surveyed two broad classes of model explanation methods: perturbation-based and gradient-based methods. The former identifies important image areas, which help machines make predictions by modifying a localized area of the input image. The latter evaluates the contribution of every single pixel of the input image to the model's prediction results through gradient backpropagation. In this study, three algorithms-the occlusion method, the integrated gradients method, and the class activation map method-are examined for a natural feature detection task using deep learning. The algorithms' strengths and weaknesses are discussed, and the consistency between model-learned and human-understandable concepts for object recognition is also compared. The experiments used two GeoAI-ready datasets to demonstrate the generalizability of the research findings.
Deep learning has become a prominent computational modeling tool in the areas of computer vision and image processing in recent years. This research comprehensively analyzes the different deep-learning methods used for image-to-image translation and reconstruction in the natural and medical imaging domains. We examine the famous deep learning frameworks, such as convolutional neural networks and generative adversarial networks, and their variants, delving into the fundamental principles and difficulties of each. In the field of natural computer vision, we investigate the development and extension of various deep-learning generative models. In comparison, we investigate the possible applications of deep learning to generative medical imaging problems, including medical image translation, MRI reconstruction, and multi-contrast MRI synthesis. This thorough review provides scholars and practitioners in the areas of generative computer vision and medical imaging with useful insights for summarizing past works and getting insight into future research paths.
This article focuses on proposing a new framework for self-organizing multichannel deep learning system (SMDLS) to solve the problem of river turbidity monitoring, which is one of the most critical indicators of water contamination. First, the basic architecture of our proposed SMDLS is established based on newly designed fractal modules, each of which is composed of one fixed block and one nested block. The nested block in the higher level fractal module is essentially the lower level fractal module. Second, we develop a novel riverway-like generalized regression loss function to improve the robustness of the proposed SMDLS to distorted labels, which is caused by small noise (e.g., the temperature drift noise and wet drift noise of turbidity sensors) in most cases and big noise (e.g., the outlier noise due to solar radiation shielded by mountains) on rare occasions. Finally, we further introduce an ensemble-induced multichannel (EMC) fusion to modify the proposed SMDLS for strengthening its performance and generalization ability. Experiment conducted on large-size river turbidity monitoring database demonstrates that our system considerably outperforms the existing state-of-the-art learning systems.
Until recently, conventional biochemical staining had the undisputed status as well-established benchmark for most biomedical problems related to clinical diagnostics, fundamental research and biotechnology. Despite this role as gold-standard, staining protocols face several challenges, such as a need for extensive, manual processing of samples, substantial time delays, altered tissue homeostasis, limited choice of contrast agents for a given sample, 2D imaging instead of 3D tomography and many more. Label-free optical technologies, on the other hand, do not rely on exogenous and artificial markers, by exploiting intrinsic optical contrast mechanisms, where the specificity is typically less obvious to the human observer. Over the past few years, digital staining has emerged as a promising concept to use modern deep learning for the translation from optical contrast to established biochemical contrast of actual stainings. In this review article, we provide an in-depth analysis of the current state-of-the-art in this field, suggest methods of good practice, identify pitfalls and challenges and postulate promising advances towards potential future implementations and applications.
In recent years, the training requirements of many state-of-the-art Deep Learning (DL) models have scaled beyond the compute and memory capabilities of a single processor, and necessitated distribution among processors. Training such massive models necessitates advanced parallelism strategies to maintain efficiency. However, such distributed DL parallelism strategies require a varied mixture of collective and point-to-point communication operations across a broad range of message sizes and scales. Examples of models using advanced parallelism strategies include Deep Learning Recommendation Models (DLRM) and Mixture-of-Experts (MoE). Communication libraries' performance varies wildly across different communication operations, scales, and message sizes. We propose MCR-DL: an extensible DL communication framework that supports all point-to-point and collective operations while enabling users to dynamically mix-and-match communication backends for a given operation without deadlocks. MCR-DL also comes packaged with a tuning suite for dynamically selecting the best communication backend for a given input tensor. We select DeepSpeed-MoE and DLRM as candidate DL models and demonstrate a 31% improvement in DS-MoE throughput on 256 V100 GPUs on the Lassen HPC system. Further, we achieve a 20% throughput improvement in a dense Megatron-DeepSpeed model and a 25% throughput improvement in DLRM on 32 A100 GPUs with the Theta-GPU HPC system.
Iris recognition is a secure biometric technology known for its stability and privacy. With no two irises being identical and little change throughout a person's lifetime, iris recognition is considered more reliable and less susceptible to external factors than other biometric recognition methods. Unlike traditional machine learning-based iris recognition methods, deep learning technology does not rely on feature engineering and boasts excellent performance. This paper collects 120 relevant papers to summarize the development of iris recognition based on deep learning. We first introduce the background of iris recognition and the motivation and contribution of this survey. Then, we present the common datasets widely used in iris recognition. After that, we summarize the key tasks involved in the process of iris recognition based on deep learning technology, including identification, segmentation, presentation attack detection, and localization. Finally, we discuss the challenges and potential development of iris recognition. This review provides a comprehensive sight of the research of iris recognition based on deep learning.
Objective: To develop a natural language processing (NLP) system to extract medications and contextual information that help understand drug changes. This project is part of the 2022 n2c2 challenge. Materials and methods: We developed NLP systems for medication mention extraction, event classification (indicating medication changes discussed or not), and context classification to classify medication changes context into 5 orthogonal dimensions related to drug changes. We explored 6 state-of-the-art pretrained transformer models for the three subtasks, including GatorTron, a large language model pretrained using >90 billion words of text (including >80 billion words from >290 million clinical notes identified at the University of Florida Health). We evaluated our NLP systems using annotated data and evaluation scripts provided by the 2022 n2c2 organizers. Results:Our GatorTron models achieved the best F1-scores of 0.9828 for medication extraction (ranked 3rd), 0.9379 for event classification (ranked 2nd), and the best micro-average accuracy of 0.9126 for context classification. GatorTron outperformed existing transformer models pretrained using smaller general English text and clinical text corpora, indicating the advantage of large language models. Conclusion: This study demonstrated the advantage of using large transformer models for contextual medication information extraction from clinical narratives.
Radio frequency (RF) fingerprinting is a tool which allows for authentication by utilizing distinct and random distortions in a received signal based on characteristics of the transmitter. We introduce a deep learning-based authentication method for a novel RF fingerprinting system called Physically Unclonable Wireless Systems (PUWS). An element of PUWS is based on the concept of Chaotic Antenna Arrays (CAAs) that can be cost effectively manufactured by utilizing mask-free laser-enhanced direct print additive manufacturing (LE-DPAM). In our experiments, using simulation data of 300 CAAs each exhibiting 4 antenna elements, we test 3 different convolutional neural network (CNN) architectures under different channel conditions and compare their authentication performance to the current state-of-the-art RF fingerprinting authentication methods.
Probabilistic diffusion models enjoy increasing popularity in the deep learning community. They generate convincing samples from a learned distribution of input images with a wide field of practical applications. Originally, these approaches were motivated from drift-diffusion processes, but these origins find less attention in recent, practice-oriented publications. We investigate probabilistic diffusion models from the viewpoint of scale-space research and show that they fulfil generalised scale-space properties on evolving probability distributions. Moreover, we discuss similarities and differences between interpretations of the physical core concept of drift-diffusion in the deep learning and model-based world. To this end, we examine relations of probabilistic diffusion to osmosis filters.