Fine-tuned Deep Neural Networks Research Articles

Reg-Tune: A Regression-Focused Fine-Tuning Approach for Profiling Low Energy Consumption and Latency

Fine-tuning deep neural networks is pivotal for creating inference modules that can be suitably imported to edge or field-programmable gate array (FPGA) platforms. Traditionally, exploration of different parameters throughout the layers of deep neural networks has been done using grid search and other brute force techniques. Although these methods lead to the optimal choice of network parameters, the search process can be very time consuming and may not consider deployment constraints across different target platforms. This work addresses this problem by proposing Reg-Tune, a regression-based profiling approach to quickly determine the trend of different metrics in relation to hardware deployment of neural networks on tinyML platforms like FPGAs and edge devices. We start by training a handful of configurations belonging to different combinations of \(\mathcal {NN}\scriptstyle \langle q (quantization),\,s (scaling)\rangle \displaystyle\) or \(\mathcal {NN}\scriptstyle \langle r (resolution),\,s\rangle \displaystyle\) workloads to generate the accuracy values respectively for their corresponding application. Next, we deploy these configurations on the target device to generate energy/latency values. According to our hypothesis, the most energy-efficient configuration suitable for deployment on the target device is a function of the variables q , r , and s . Finally, these trained and deployed configurations and their related results are used as data points for polynomial regression with the variables q , r , and s to realize the trend for accuracy/energy/latency on the target device. Our setup allows us to choose the near-optimal energy-consuming or latency-driven configuration for the desired accuracy from the contour profiles of energy/latency across different tinyML device platforms. To this extent, we demonstrate the profiling process for three different case studies and across two platforms for energy and latency fine-tuning. Our approach results in at least 5.7 \(\times\) better energy efficiency when compared to recent implementations for human activity recognition on FPGA and 74.6% reduction in latency for semantic segmentation of aerial imagery on edge devices compared to baseline deployments.

Classification of Brain Tumors: A Comparative Approach of Shallow and Deep Neural Networks

Brain tumors can be generated anywhere in the brain, with an extensive size range and morphology that makes it challenging to identify and classify. Classifying brain tumors is essential for developing personalized treatment plans. Different types of brain tumors have different responses to treatment, and an accurate classification can help medical professionals develop treatment plans tailored to each patient’s needs. Therefore, this case study aimed to classify T1-weighted contrast-enhanced images of three types of tumors through various approaches, from shallow neural networks to fine-tuning deep neural networks trained. Comparing shallow and deep neural network approaches could help to understand the trade-offs between their performance, interoperability, interpretability, benefits, limitations, scopes, and overall, choosing the best method for a given problem.

Enhancement of Multiclass Dermoscopic Image Classification with a Fusion Model and Fine-Tuned Deep Neural Networks

Since malignancy of the skin is a frequent and sometimes fatal illness, prompt and precise detection is essential to successful treatment. Dermoscopic imaging is now emerging as a helpful tool for dermatologists in aiding with the identification of various types of skin lesions and offering timely treatments. Due to varied image conditions, data unbalancing, and fine-grained pattern classification task are is challenging in skin cancer detection. Dermoscopic images vary greatly in quality, illumination, and orientation, making it challenging for models to efficiently generalize across varied image conditions. To obtain accurate and reliable skin cancer detection, such challenges require the utilization of sophisticated deep learning techniques and rigorous model development. In multiple classes skin malignancy classification, convolutional neural networks outperform dermatology professionals, according to recent research. In the proposed system, firstly a preprocessing technique has been developed that includes hair and noise removal, data augmentation, and image resizing with model parameters. Then training has been performed using transfer learning on the fusion of DenseNet121 and ResNet50V2 model on the HAM10000 dataset. The proposed technique's performance is assessed using measures such as sensitivity, f-1 score, specificity, and accuracy. For multi-class skin lesion classification, the suggested technique achieves a precision of 97.88%, F-Score of 98%, sensitivity of 99%, accuracy of 97.08%, and Mathew's correlation coefficient (MCC) of 96.4%, it executes superior to cutting-edge techniques.

Towards accurate knowledge transfer via target-awareness representation disentanglement

Fine-tuning deep neural networks pre-trained on large scale datasets is one of the most practical transfer learning paradigm given limited quantity of training samples. To obtain better generalization, using the starting point as the reference (SPAR), either through weights or features, has been successfully applied to transfer learning as a regularizer. However, due to the domain discrepancy between the source and target task, there exists obvious risk of negative transfer in a straightforward manner of knowledge preserving. In this paper, we propose a novel transfer learning algorithm, introducing the idea of Target-awareness REpresentation Disentanglement (TRED\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ extrm{TRED}$$\\end{document}), where the relevant knowledge with respect to the target task is disentangled from the original source model and used as a regularizer during fine-tuning the target model. Two alternative approaches, maximizing Maximum Mean Discrepancy (Max-MMD) and minimizing mutual information (Min-MI) are introduced to achieve the desired disentanglement. Experiments on various real world datasets show that our method stably improves the standard fine-tuning by more than 2% in average. TRED\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ extrm{TRED}$$\\end{document} also outperforms related state-of-the-art transfer learning regularizers such as L2-SP\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\mathrm {L^2\ ext {-}SP}$$\\end{document}, AT\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ extrm{AT}$$\\end{document}, DELTA\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ extrm{DELTA}$$\\end{document}, and BSS\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\ extrm{BSS}$$\\end{document}. Moreover, our solution is compatible with different choices of disentangling strategies. While the combination of Max-MMD and Min-MI typically achieves higher accuracy, only using Max-MMD can be a preferred choice in applications with low resource budgets.

Leveraging Transfer Learning for Spatio-Temporal Human Activity Recognition from Video Sequences

Human Activity Recognition (HAR) is an active research area due to its applications in pervasive computing, human-computer interaction, artificial intelligence, health care, and social sciences. Moreover, dynamic environments and anthropometric differences between individuals make it harder to recognize actions. This study focused on human activity in video sequences acquired with an RGB camera because of its vast range of real-world applications. It uses two-stream ConvNet to extract spatial and temporal information and proposes a fine-tuned deep neural network. Moreover, the transfer learning paradigm is adopted to extract varied and fixed frames while reusing object identification information. Six state-of-the-art pre-trained models are exploited to find the best model for spatial feature extraction. For temporal sequence, this study uses dense optical flow following the two-stream ConvNet and Bidirectional Long Short Term Memory (BiLSTM) to capture long-term dependencies. Two state-of-the-art datasets, UCF101 and HMDB51, are used for evaluation purposes. In addition, seven state-of-the-art optimizers are used to fine-tune the proposed network parameters. Furthermore, this study utilizes an ensemble mechanism to aggregate spatial-temporal features using a four-stream Convolutional Neural Network (CNN), where two streams use RGB data. In contrast, the other uses optical flow images. Finally, the proposed ensemble approach using max hard voting outperforms state-of-the-art methods with 96.30% and 90.07% accuracies on the UCF101 and HMDB51 datasets.

Defect classification of radius shaping in the tire curing process using Fine-Tuned Deep Neural Network

The curing process or vulcanization process is the final stage of the tire manufacturing process, where the properties of the tire compound change from rubber-plastic material to become elastic by forming cross-links in its molecular structure. The green tire is formed in the curing process, which is placed on the bottom mould. The inside of the green tire surrounds the bladder. The top mould will close to carry out the next curing process. In closing the mould, there is a shaping process of forming a green tire placed on the bladder and given a proportional pressure. Improper or abnormal radius shaping results cause seventy percent of product defects. This paper proposed abnormal detection of radius shaping in the curing process using Fine-tuned Deep Neural Network (DNN). Several DNN models have been examined to analyze an optimized DNN model for abnormal detection of radius shaping in the curing process. The fine-tuned DNN architecture has been exported for the curing system. The DNN was trained with a training accuracy of 97.88%, a validation accuracy of 95%, a testing accuracy of 100%, and a loss of 4.93%.

CD-SDN: Unsupervised Sensitivity Disparity Networks for Hyper-Spectral Image Change Detection

Deep neural networks (DNNs) could be affected by the regression level of learning frameworks and challenging changes caused by external factors; their deep expressiveness is greatly restricted. Inspired by the fine-tuned DNNs with sensitivity disparity to the pixels of two states, in this paper, we propose a novel change detection scheme served by sensitivity disparity networks (CD-SDN). The CD-SDN is proposed for detecting changes in bi-temporal hyper-spectral images captured by the AVIRIS sensor and HYPERION sensor over time. In the CD-SDN, two deep learning frameworks, unchanged sensitivity network (USNet) and changed sensitivity network (CSNet), are utilized as the dominant part for the generation of binary argument map (BAM) and high assurance map (HAM). Then two approaches, arithmetic mean and argument learning, are employed to re-estimate the changes of BAM. Finally, the detected results are merged with HAM and obtain the final detected binary change maps (BCMs). Experiments are performed on three real-world hyperspectral image datasets, and the results indicate the good universality and adaptability of the proposed scheme, as well as its superiority over other existing state-of-the-art algorithms.

Fine-tuned deep neural networks for polyp detection in colonoscopy images

Fine-tuned deep neural networks for polyp detection in colonoscopy images

Detection of Malicious Software by Analyzing Distinct Artifacts Using Machine Learning and Deep Learning Algorithms

Malware is one of the most significant threats in today’s computing world since the number of websites distributing malware is increasing at a rapid rate. Malware analysis and prevention methods are increasingly becoming necessary for computer systems connected to the Internet. This software exploits the system’s vulnerabilities to steal valuable information without the user’s knowledge, and stealthily send it to remote servers controlled by attackers. Traditionally, anti-malware products use signatures for detecting known malware. However, the signature-based method does not scale in detecting obfuscated and packed malware. Considering that the cause of a problem is often best understood by studying the structural aspects of a program like the mnemonics, instruction opcode, API Call, etc. In this paper, we investigate the relevance of the features of unpacked malicious and benign executables like mnemonics, instruction opcodes, and API to identify a feature that classifies the executable. Prominent features are extracted using Minimum Redundancy and Maximum Relevance (mRMR) and Analysis of Variance (ANOVA). Experiments were conducted on four datasets using machine learning and deep learning approaches such as Support Vector Machine (SVM), Naïve Bayes, J48, Random Forest (RF), and XGBoost. In addition, we also evaluate the performance of the collection of deep neural networks like Deep Dense network, One-Dimensional Convolutional Neural Network (1D-CNN), and CNN-LSTM in classifying unknown samples, and we observed promising results using APIs and system calls. On combining APIs/system calls with static features, a marginal performance improvement was attained comparing models trained only on dynamic features. Moreover, to improve accuracy, we implemented our solution using distinct deep learning methods and demonstrated a fine-tuned deep neural network that resulted in an F1-score of 99.1% and 98.48% on Dataset-2 and Dataset-3, respectively.

Conventional Machine Learning versus Deep Learning for Magnification Dependent Histopathological Breast Cancer Image Classification: A Comparative Study with Visual Explanation.

Breast cancer is a serious threat to women. Many machine learning-based computer-aided diagnosis (CAD) methods have been proposed for the early diagnosis of breast cancer based on histopathological images. Even though many such classification methods achieved high accuracy, many of them lack the explanation of the classification process. In this paper, we compare the performance of conventional machine learning (CML) against deep learning (DL)-based methods. We also provide a visual interpretation for the task of classifying breast cancer in histopathological images. For CML-based methods, we extract a set of handcrafted features using three feature extractors and fuse them to get image representation that would act as an input to train five classical classifiers. For DL-based methods, we adopt the transfer learning approach to the well-known VGG-19 deep learning architecture, where its pre-trained version on the large scale ImageNet, is block-wise fine-tuned on histopathological images. The evaluation of the proposed methods is carried out on the publicly available BreaKHis dataset for the magnification dependent classification of benign and malignant breast cancer and their eight sub-classes, and a further validation on KIMIA Path960, a magnification-free histopathological dataset with 20 image classes, is also performed. After providing the classification results of CML and DL methods, and to better explain the difference in the classification performance, we visualize the learned features. For the DL-based method, we intuitively visualize the areas of interest of the best fine-tuned deep neural networks using attention maps to explain the decision-making process and improve the clinical interpretability of the proposed models. The visual explanation can inherently improve the pathologist’s trust in automated DL methods as a credible and trustworthy support tool for breast cancer diagnosis. The achieved results show that DL methods outperform CML approaches where we reached an accuracy between 94.05% and 98.13% for the binary classification and between 76.77% and 88.95% for the eight-class classification, while for DL approaches, the accuracies range from 85.65% to 89.32% for the binary classification and from 63.55% to 69.69% for the eight-class classification.

Vision-assisted recognition of stereotype behaviors for early diagnosis of Autism Spectrum Disorders

Medical diagnosis supported by computer-assisted technologies is getting more popularity and acceptance among medical society. In this paper, we propose a non-intrusive vision-assisted method based on human action recognition to facilitate the diagnosis of Autism Spectrum Disorder (ASD). We collected a novel and comprehensive video dataset f the most distinctive Stereotype actions of this disorder with the assistance of professional clinicians. Several frameworks as a function of different input modalities were developed and used to produce extensive baseline results. Various local descriptors, which are commonly used within the Bag-of-Visual-Words approach, were tested with Multi-layer Perceptron (MLP), Gaussian Naive Bayes (GNB), and Support Vector Machines (SVM) classifiers for recognizing ASD associated behaviors. Additionally, we developed a framework that first receives articulated pose-based skeleton sequences as input and follows an LSTM network to learn the temporal evolution of the poses. Finally, obtained results were compared with two fine-tuned deep neural networks: ConvLSTM and 3DCNN. The results revealed that the Histogram of Optical Flow (HOF) descriptor achieves the best results when used with MLP classifier. The promising baseline results also confirmed that an action-recognition-based system can be potentially used to assist clinicians to provide a reliable, accurate, and timely diagnosis of ASD disorder.

Automated diagnosis of COVID-19 with limited posteroanterior chest X-ray images using fine-tuned deep neural networks.

The novel coronavirus 2019 (COVID-19) is a respiratory syndrome that resembles pneumonia. The current diagnostic procedure of COVID-19 follows reverse-transcriptase polymerase chain reaction (RT-PCR) based approach which however is less sensitive to identify the virus at the initial stage. Hence, a more robust and alternate diagnosis technique is desirable. Recently, with the release of publicly available datasets of corona positive patients comprising of computed tomography (CT) and chest X-ray (CXR) imaging; scientists, researchers and healthcare experts are contributing for faster and automated diagnosis of COVID-19 by identifying pulmonary infections using deep learning approaches to achieve better cure and treatment. These datasets have limited samples concerned with the positive COVID-19 cases, which raise the challenge for unbiased learning. Following from this context, this article presents the random oversampling and weighted class loss function approach for unbiased fine-tuned learning (transfer learning) in various state-of-the-art deep learning approaches such as baseline ResNet, Inception-v3, Inception ResNet-v2, DenseNet169, and NASNetLarge to perform binary classification (as normal and COVID-19 cases) and also multi-class classification (as COVID-19, pneumonia, and normal case) of posteroanterior CXR images. Accuracy, precision, recall, loss, and area under the curve (AUC) are utilized to evaluate the performance of the models. Considering the experimental results, the performance of each model is scenario dependent; however, NASNetLarge displayed better scores in contrast to other architectures, which is further compared with other recently proposed approaches. This article also added the visual explanation to illustrate the basis of model classification and perception of COVID-19 in CXR images.

Twice fine‐tuning deep neural networks for paraphrase identification

In this Letter, the authors introduce a novel approach to learn representations for sentence-level paraphrase identification (PI) using BERT and ten natural language processing tasks. Their method trains an unsupervised model called BERT with two different tasks to detect whether two sentences are in paraphrase relation or not. Unlike conventional BERT, which fine tunes the target task such as PI to pre-trained BERT, twice fine-tuning deep neural networks first fine tune each task (e.g. general language understanding evaluation tasks, question answering, and paraphrase adversaries from word scrambling task) and second fine tune target PI task. As a result, the multi-fine-tuned BERT model outperformed the fine-tuned model only with Microsoft Research Paraphrase Corpus (MRPC), which is paraphrase data, except for one case of Stanford Sentiment Treebank - 2 (SST-2). Multi-task fine-tuning is a simple idea but experimentally powerful. Experiments show that fine-tuning just PI tasks to the BERT already gives enough performance, but additionally, fine-tuning similar tasks can affect performance (3.4% point absolute improvement) and be competitive with the state-of-the-art systems.

Adaptive Medical Image Deep Color Perception Algorithm

The existing medical imaging technologies have little consideration on color information, thus most of medical images are gray. Classical hand-craft features-based methods have obtained unsatisfactory results in colorizing medical images. Moreover, these methods ignore the deep feature of medical images that represent pathology and color information. In this paper, we propose a novel method that iteratively colorizes grayscale medical images under preserving content in fine-tuned deep neural network. To the best of our knowledge, there is no currently work that attempts to colorize the medical image by using deep neural network. Specifically, we propose Y-loss which is defined as nonlinear combination of $\ell _{1}$ and $\ell _{2}$ norm to preserve content invariance between target and colorized medical image. Then, adaptive reference image search algorithm is introduced to code reference and target medical image with D-hash and search reference image in hash code automatically, which free the manual selection of the reference image. Extensive experiment results show that the proposed method can generate higher quality colored medical image than recent state-of-the-art methods, and can be approved by the doctor. The objective evaluation (PSNR and SSIM) outperform an average increment 24% and 47% than baseline method, respectively. Our code is available at: https://github.com/Tongshiyue/Adaptive-medical-image-deep-color-perception-algorithm .

Deep learning with cinematic rendering: fine-tuning deep neural networks using photorealistic medical images

Deep learning has emerged as a powerful artificial intelligence tool to interpret medical images for a growing variety of applications. However, the paucity of medical imaging data with high-quality annotations that is necessary for training such methods ultimately limits their performance. Medical data is challenging to acquire due to privacy issues, shortage of experts available for annotation, limited representation of rare conditions and cost. This problem has previously been addressed by using synthetically generated data. However, networks trained on synthetic data often fail to generalize to real data. Cinematic rendering simulates the propagation and interaction of light passing through tissue models reconstructed from CT data, enabling the generation of photorealistic images. In this paper, we present one of the first applications of cinematic rendering in deep learning, in which we propose to fine-tune synthetic data-driven networks using cinematically rendered CT data for the task of monocular depth estimation in endoscopy. Our experiments demonstrate that: (a) convolutional neural networks (CNNs) trained on synthetic data and fine-tuned on photorealistic cinematically rendered data adapt better to real medical images and demonstrate more robust performance when compared to networks with no fine-tuning, (b) these fine-tuned networks require less training data to converge to an optimal solution, and (c) fine-tuning with data from a variety of photorealistic rendering conditions of the same scene prevents the network from learning patient-specific information and aids in generalizability of the model. Our empirical evaluation demonstrates that networks fine-tuned with cinematically rendered data predict depth with 56.87% less error for rendered endoscopy images and 27.49% less error for real porcine colon endoscopy images.

Deep peak-neutral difference feature for facial expression recognition

Facial expression recognition (FER) is important in vision-related applications. Deep neural networks demonstrate impressive performance for face recognition; however, it should be noted that this method relies heavily on a great deal of manually labeled training data, which is not available for facial expressions in real-world applications. Hence, we propose a powerful facial feature called deep peak–neutral difference (DPND) for FER. DPND is defined as the difference between two deep representations of the fully expressive (peak) and neutral facial expression frames. The difference tends to emphasize the facial parts that are changed in the transition from the neutral to the expressive face and to eliminate the face identity information retained in the fine-tuned deep neural network for facial expression, the network has been trained on large-scale face recognition dataset. Furthermore, unsupervised clustering and semi-supervised classification methods are presented to automatically acquire the neutral and peak frames from the expression sequence. The proposed facial expression feature achieved encouraging results on public databases, which suggests that it has strong potential to recognize facial expressions in real-world applications.

Age and gender classification from speech and face images by jointly fine-tuned deep neural networks

The classification of human's age and gender from speech and face images is a challenging task that has important applications in real-life and its applications are expected to grow more in the future. Deep neural networks (DNNs) and Convolutional neural networks (CNNs) are considered as one of the state-of-art systems as feature extractors and classifiers and are proven to be very efficient in analyzing problems with complex feature space. In this work, we propose a new cost function for fine-tuning two DNNs jointly. The proposed cost function is evaluated by using speech utterances and unconstrained face images for age and gender classification task. The proposed classifier design consists of two DNNs trained on different feature sets, which are extracted from the same input data. Mel-frequency cepstral coefficients (MFCCs) and fundamental frequency (F0) and the shifted delta cepstral coefficients (SDC) are extracted from speech as the first and second feature sets, respectively. Facial appearance and the depth information are extracted from face images as the first and second feature sets, respectively. Jointly training of two DNNs with the proposed cost function improved the classification accuracies and minimized the over-fitting effect for both speech-based and image-based systems. Extensive experiments have been conducted to evaluate the performance and the accuracy of the proposed work. Two publicly available databases, the Age-Annotated Database of the German Telephone Speech database (aGender) and the Adience database, are used to evaluate the proposed system. The overall accuracy of the proposed system is calculated as 56.06% for seven speaker classes and overall exact accuracy is calculated as 63.78% for Adience database.