In recent years, larger and deeper models are springing up and continuously pushing state-of-the-art (SOTA) results across various fields like natural language processing (NLP) and computer vision (CV). However, despite promising results, it needs to be noted that the computations required by SOTA models have been increased at an exponential rate. Massive computations not only have a surprisingly large carbon footprint but also have negative effects on research inclusiveness and deployment on real-world applications. Green deep learning is an increasingly hot research field that appeals to researchers to pay attention to energy usage and carbon emission during model training and inference. The target is to yield novel results with lightweight and efficient technologies. Many technologies can be used to achieve this goal, like model compression and knowledge distillation. This paper focuses on presenting a systematic review of the development of Green deep learning technologies. We classify these approaches into four categories: (1) compact networks, (2) energy-efficient training strategies, (3) energy-efficient inference approaches, and (4) efficient data usage. For each category, we discuss the progress that has been achieved and the unresolved challenges.
Analogical proportions are statements of the form "A is to B as C is to D". They constitute an inference tool that provides a logical framework to address learning, transfer, and explainability concerns and that finds useful applications in artificial intelligence and natural language processing. In this paper, we address two problems, namely, analogy detection and resolution in morphology. Multiple symbolic approaches tackle the problem of analogies in morphology and achieve competitive performance. We show that it is possible to use a data-driven strategy to outperform those models. We propose an approach using deep learning to detect and solve morphological analogies. It encodes structural properties of analogical proportions and relies on a specifically designed embedding model capturing morphological characteristics of words. We demonstrate our model's competitive performance on analogy detection and resolution over multiple languages. We provide an empirical study to analyze the impact of balancing training data and evaluate the robustness of our approach to input perturbation.
Polarization of the mammalian embryo at the right developmental time is critical for its development to term and would be valuable in assessing the potential of human embryos. However, tracking polarization requires invasive fluorescence staining, impermissible in the in vitro fertilization clinic. Here, we report the use of artificial intelligence to detect polarization from unstained time-lapse movies of mouse embryos. We assembled a dataset of bright-field movie frames from 8-cell-stage embryos, side-by-side with corresponding images of fluorescent markers of cell polarization. We then used an ensemble learning model to detect whether any bright-field frame showed an embryo before or after onset of polarization. Our resulting model has an accuracy of 85% for detecting polarization, significantly outperforming human volunteers trained on the same data (61% accuracy). We discovered that our self-learning model focuses upon the angle between cells as one known cue for compaction, which precedes polarization, but it outperforms the use of this cue alone. By compressing three-dimensional time-lapsed image data into two-dimensions, we are able to reduce data to an easily manageable size for deep learning processing. In conclusion, we describe a method for detecting a key developmental feature of embryo development that avoids clinically impermissible fluorescence staining.
Photoacoustic imaging (PAI) is a non-invasive imaging modality that detects the ultrasound signal generated from tissue with light excitation. Photoacoustic computed tomography (PACT) uses unfocused large-area light to illuminate the target with ultrasound transducer array for PA signal detection. Limited-view issue could cause a low-quality image in PACT due to the limitation of geometric condition. The model-based method is used to resolve this problem, which contains different regularization. To adapt fast and high-quality reconstruction of limited-view PA data, in this paper, a model-based method that combines the mathematical variational model with deep learning is proposed to speed up and regularize the unrolled procedure of reconstruction. A deep neural network is designed to adapt the step of the gradient updated term of data consistency in the gradient descent procedure, which can obtain a high-quality PA image only with a few iterations. Note that all parameters and priors are automatically learned during the offline training stage. In experiments, we show that this method outperforms the other methods with half-view (180 degrees) simulation and real data. The comparison of different model-based methods show that our proposed scheme has superior performances (over 0.05 for SSIM) with same iteration (3 times) steps. Furthermore, an unseen data is used to validate the generalization of different methods. Finally, we find that our method obtains superior results (0.94 value of SSIM for in vivo) with a high robustness and accelerated reconstruction.
Facial beauty prediction (FBP) aims to develop a machine that automatically makes facial attractiveness assessment. In the past those results were highly correlated with human ratings, therefore also with their bias in annotating. As artificial intelligence can have racist and discriminatory tendencies, the cause of skews in the data must be identified. Development of training data and AI algorithms that are robust against biased information is a new challenge for scientists. As aesthetic judgement usually is biased, we want to take it one step further and propose an Unbiased Convolutional Neural Network for FBP. While it is possible to create network models that can rate attractiveness of faces on a high level, from an ethical point of view, it is equally important to make sure the model is unbiased. In this work, we introduce AestheticNet, a state-of-the-art attractiveness prediction network, which significantly outperforms competitors with a Pearson Correlation of 0.9601. Additionally, we propose a new approach for generating a bias-free CNN to improve fairness in machine learning.
Computer vision has been thriving since AI development was gaining thrust. Using deep learning techniques has been the most popular way which computer scientists thought the solution of. However, deep learning techniques tend to show lower performance than manual processing. Using deep learning is not always the answer to a problem related to computer vision.
The field of deep learning has witnessed a remarkable shift towards extremely compute- and memory-intensive neural networks. These newer larger models have enabled researchers to advance state-of-the-art tools across a variety of fields. This phenomenon has spurred the development of algorithms for distributed training of neural networks over a larger number of hardware accelerators. In this paper, we discuss and compare current state-of-the-art frameworks for large scale distributed deep learning. First, we survey current practices in distributed learning and identify the different types of parallelism used. Then, we present empirical results comparing their performance on large image and language training tasks. Additionally, we address their statistical efficiency and memory consumption behavior. Based on our results, we discuss algorithmic and implementation portions of each framework which hinder performance.
Driving behaviour is one of the primary causes of road crashes and accidents, and these can be decreased by identifying and minimizing aggressive driving behaviour. This study identifies the timesteps when a driver in different circumstances (rush, mental conflicts, reprisal) begins to drive aggressively. An observer (real or virtual) is needed to examine driving behaviour to discover aggressive driving occasions; we overcome this problem by using a smartphone's GPS sensor to detect locations and classify drivers' driving behaviour every three minutes. To detect timeseries patterns in our dataset, we employ RNN (GRU, LSTM) algorithms to identify patterns during the driving course. The algorithm is independent of road, vehicle, position, or driver characteristics. We conclude that three minutes (or more) of driving (120 seconds of GPS data) is sufficient to identify driver behaviour. The results show high accuracy and a high F1 score.
Accurate cardiac computing, analysis and modeling from multi-modality images are important for the diagnosis and treatment of cardiac disease. Late gadolinium enhancement magnetic resonance imaging (LGE MRI) is a promising technique to visualize and quantify myocardial infarction (MI) and atrial scars. Automating quantification of MI and atrial scars can be challenging due to the low image quality and complex enhancement patterns of LGE MRI. Moreover, compared with the other sequences LGE MRIs with gold standard labels are particularly limited, which represents another obstacle for developing novel algorithms for automatic segmentation and quantification of LGE MRIs. This chapter aims to summarize the state-of-the-art and our recent advanced contributions on deep learning based multi-modality cardiac image analysis. Firstly, we introduce two benchmark works for multi-sequence cardiac MRI based myocardial and pathology segmentation. Secondly, two novel frameworks for left atrial scar segmentation and quantification from LGE MRI were presented. Thirdly, we present three unsupervised domain adaptation techniques for cross-modality cardiac image segmentation.
In this paper we investigate a variety of deep learning strategies for solving inverse problems. We classify existing deep learning solutions for inverse problems into three categories of Direct Mapping, Data Consistency Optimizer, and Deep Regularizer. We choose a sample of each inverse problem type, so as to compare the robustness of the three categories, and report a statistical analysis of their differences. We perform extensive experiments on the classic problem of linear regression and three well-known inverse problems in computer vision, namely image denoising, 3D human face inverse rendering, and object tracking, selected as representative prototypes for each class of inverse problems. The overall results and the statistical analyses show that the solution categories have a robustness behaviour dependent on the type of inverse problem domain, and specifically dependent on whether or not the problem includes measurement outliers. Based on our experimental results, we conclude by proposing the most robust solution category for each inverse problem class.
The classification of galaxy morphology plays a crucial role in understanding galaxy formation and evolution. Traditionally, this process is done manually. The emergence of deep learning techniques has given room for the automation of this process. As such, this paper offers a comparison of deep learning architectures to determine which is best suited for optical galaxy morphology classification. Adapting the model training method proposed by Walmsley et al in 2021, the Zoobot Python library is used to train models to predict Galaxy Zoo DECaLS decision tree responses, made by volunteers, using EfficientNet B0, DenseNet121 and ResNet50 as core model architectures. The predicted results are then used to generate accuracy metrics per decision tree question to determine architecture performance. DenseNet121 was found to produce the best results, in terms of accuracy, with a reasonable training time. In future, further testing with more deep learning architectures could prove beneficial.
Convolutional Neural Networks (CNNs) have achieved impressive performance on many computer vision related tasks, such as object detection, image recognition, image retrieval, etc. These achievements benefit from the CNNs' outstanding capability to learn discriminative features with deep layers of neuron structures and iterative training process. This has inspired the EEG research community to adopt CNN in performing EEG classification tasks. However, CNNs learned features are not immediately interpretable, causing a lack of understanding of the CNNs' internal working mechanism. To improve CNN interpretability, CNN visualization methods are applied to translate the internal features into visually perceptible patterns for qualitative analysis of CNN layers. Many CNN visualization methods have been proposed in the Computer Vision literature to interpret the CNN network structure, operation, and semantic concept, yet applications to EEG data analysis have been limited. In this work we use 3 different methods to extract EEG-relevant features from a CNN trained on raw EEG data: optimal samples for each classification category, activation maximization, and reverse convolution. We applied these methods to a high-performing Deep Learning model with state-of-the-art performance for an EEG sex classification task, and show that the model features a difference in the theta frequency band. We show that visualization of a CNN model can reveal interesting EEG results. Using these tools, EEG researchers using Deep Learning can better identify the learned EEG features, possibly identifying new class relevant biomarkers.
Uncertainty quantification (UQ) helps to make trustworthy predictions based on collected observations and uncertain domain knowledge. With increased usage of deep learning in various applications, the need for efficient UQ methods that can make deep models more reliable has increased as well. Among applications that can benefit from effective handling of uncertainty are the deep learning based differential equation (DE) solvers. We adapt several state-of-the-art UQ methods to get the predictive uncertainty for DE solutions and show the results on four different DE types.
Breast cancer has long been a prominent cause of mortality among women. Diagnosis, therapy, and prognosis are now possible, thanks to the availability of RNA sequencing tools capable of recording gene expression data. Molecular subtyping being closely related to devising clinical strategy and prognosis, this paper focuses on the use of gene expression data for the classification of breast cancer into four subtypes, namely, Basal, Her2, LumA, and LumB. In stage 1, we suggested a deep learning-based model that uses an autoencoder to reduce dimensionality. The size of the feature set is reduced from 20,530 gene expression values to 500 by using an autoencoder. This encoded representation is passed to the deep neural network of the second stage for the classification of patients into four molecular subtypes of breast cancer. By deploying the combined network of stages 1 and 2, we have been able to attain a mean 10-fold test accuracy of 0.907 on the TCGA breast cancer dataset. The proposed framework is fairly robust throughout 10 different runs, as shown by the boxplot for classification accuracy. Compared to related work reported in the literature, we have achieved a competitive outcome. In conclusion, the proposed two-stage deep learning-based model is able to accurately classify four breast cancer subtypes, highlighting the autoencoder's capacity to deduce the compact representation and the neural network classifier's ability to correctly label breast cancer patients.
Predicting future stock prices and their movement patterns is a complex problem. Hence, building a portfolio of capital assets using the predicted prices to achieve the optimization between its return and risk is an even more difficult task. This work has carried out an analysis of the time series of the historical prices of the top five stocks from the nine different sectors of the Indian stock market from January 1, 2016, to December 31, 2020. Optimum portfolios are built for each of these sectors. For predicting future stock prices, a long-and-short-term memory (LSTM) model is also designed and fine-tuned. After five months of the portfolio construction, the actual and the predicted returns and risks of each portfolio are computed. The predicted and the actual returns of each portfolio are found to be high, indicating the high precision of the LSTM model.
We introduce an effective strategy to generate an annotated synthetic dataset of microbiological images of Petri dishes that can be used to train deep learning models in a fully supervised fashion. The developed generator employs traditional computer vision algorithms together with a neural style transfer method for data augmentation. We show that the method is able to synthesize a dataset of realistic looking images that can be used to train a neural network model capable of localising, segmenting, and classifying five different microbial species. Our method requires significantly fewer resources to obtain a useful dataset than collecting and labeling a whole large set of real images with annotations. We show that starting with only 100 real images, we can generate data to train a detector that achieves comparable results (detection mAP = 0.416, and counting MAE = 4.49) to the same detector but trained on a real, several dozen times bigger dataset (mAP = 0.520, MAE = 4.31), containing over 7k images. We prove the usefulness of the method in microbe detection and segmentation, but we expect that it is general and flexible and can also be applicable in other domains of science and industry to detect various objects.
Deep learning researchers have a keen interest in proposing two new novel activation functions which can boost network performance. A good choice of activation function can have significant consequences in improving network performance. A handcrafted activation is the most common choice in neural network models. ReLU is the most common choice in the deep learning community due to its simplicity though ReLU has some serious drawbacks. In this paper, we have proposed a new novel activation function based on approximation of known activation functions like Leaky ReLU, and we call this function Smooth Maximum Unit (SMU). Replacing ReLU by SMU, we have got 6.22% improvement in the CIFAR100 dataset with the ShuffleNet V2 model.
Decentralized algorithm is a form of computation that achieves a global goal through local dynamics that relies on low-cost communication between directly-connected agents. On large-scale optimization tasks involving distributed datasets, decentralized algorithms have shown strong, sometimes superior, performance over distributed algorithms with a central node. Recently, developing decentralized algorithms for deep learning has attracted great attention. They are considered as low-communication-overhead alternatives to those using a parameter server or the Ring-Allreduce protocol. However, the lack of an easy-to-use and efficient software package has kept most decentralized algorithms merely on paper. To fill the gap, we introduce BlueFog, a python library for straightforward, high-performance implementations of diverse decentralized algorithms. Based on a unified abstraction of various communication operations, BlueFog offers intuitive interfaces to implement a spectrum of decentralized algorithms, from those using a static, undirected graph for synchronous operations to those using dynamic and directed graphs for asynchronous operations. BlueFog also adopts several system-level acceleration techniques to further optimize the performance on the deep learning tasks. On mainstream DNN training tasks, BlueFog reaches a much higher throughput and achieves an overall $1.2\times \sim 1.8\times$ speedup over Horovod, a state-of-the-art distributed deep learning package based on Ring-Allreduce. BlueFog is open source at https://github.com/Bluefog-Lib/bluefog.
To the best of our knowledge, this is the first paper to introduce Convolutions to Gated MultiLayer Perceptron and contributes an implementation of this novel Deep Learning architecture. Google Brain introduced the gMLP in May 2021. Microsoft introduced Convolutions in Vision Transformer in Mar 2021. Inspired by both gMLP and CvT, we introduce convolutional layers in gMLP. CvT combined the power of Convolutions and Attention. Our implementation combines the best of Convolutional learning along with spatial gated MLP. Further, the paper visualizes how CgMLP learns. Visualizations show how CgMLP learns from features such as outline of a car. While Attention was the basis of much of recent progress in Deep Learning, gMLP proposed an approach that doesn't use Attention computation. In Transformer based approaches, a whole lot of Attention matrixes need to be learnt using vast amount of training data. In gMLP, the fine tunning for new tasks can be challenging by transfer learning with smaller datasets. We implement CgMLP and compares it with gMLP on CIFAR dataset. Experimental results explore the power of generaliza-tion of CgMLP, while gMLP tend to drastically overfit the training data. To summarize, the paper contributes a novel Deep Learning architecture and demonstrates the learning mechanism of CgMLP through visualizations, for the first time in literature.
We develop deep learning models to learn the hedge ratio for S&P500 index options directly from options data. We compare different combinations of features and show that a feedforward neural network model with time to maturity, Black-Scholes delta and a sentiment variable (VIX for calls and index return for puts) as input features performs the best in the out-of-sample test. This model significantly outperforms the standard hedging practice that uses the Black-Scholes delta and a recent data-driven model. Our results demonstrate the importance of market sentiment for hedging efficiency, a factor previously ignored in developing hedging strategies.