Optimal Neural Architecture Research Articles

Advances in neural architecture search.

Automated machine learning (AutoML) has achieved remarkable success in automating the non-trivial process of designing machine learning models. Among the focal areas of AutoML, neural architecture search (NAS) stands out, aiming to systematically explore the complex architecture space to discover the optimal neural architecture configurations without intensive manual interventions. NAS has demonstrated its capability of dramatic performance improvement across a large number of real-world tasks. The core components in NAS methodologies normally include (i) defining the appropriate search space, (ii) designing the right search strategy and (iii) developing the effective evaluation mechanism. Although early NAS endeavors are characterized via groundbreaking architecture designs, the imposed exorbitant computational demands prompt a shift towards more efficient paradigms such as weight sharing and evaluation estimation, etc. Concurrently, the introduction of specialized benchmarks has paved the way for standardized comparisons of NAS techniques. Notably, the adaptability of NAS is evidenced by its capability of extending to diverse datasets, including graphs, tabular data and videos, etc., each of which requires a tailored configuration. This paper delves into the multifaceted aspects of NAS, elaborating on its recent advances, applications, tools, benchmarks and prospective research directions.

Exploring Physics-Informed Neural Networks for the Generalized Nonlinear Sine-Gordon Equation

The nonlinear sine-Gordon equation is a prevalent feature in numerous scientific and engineering problems. In this paper, we propose a machine learning-based approach, physics-informed neural networks (PINNs), to investigate and explore the solution of the generalized non-linear sine-Gordon equation, encompassing Dirichlet and Neumann boundary conditions. To incorporate physical information for the sine-Gordon equation, a multiobjective loss function has been defined consisting of the residual of governing partial differential equation (PDE), initial conditions, and various boundary conditions. Using multiple densely connected independent artificial neural networks (ANNs), called feedforward deep neural networks designed to handle partial differential equations, PINNs have been trained through automatic differentiation to minimize a loss function that incorporates the given PDE that governs the physical laws of phenomena. To illustrate the effectiveness, validity, and practical implications of our proposed approach, two computational examples from the nonlinear sine-Gordon are presented. We have developed a PINN algorithm and implemented it using Python software. Various experiments were conducted to determine an optimal neural architecture. The network training was employed by using the current state-of-the-art optimization methods in machine learning known as Adam and L-BFGS-B minimization techniques. Additionally, the solutions from the proposed method are compared with the established analytical solutions found in the literature. The findings show that the proposed method is a computational machine learning approach that is accurate and efficient for solving nonlinear sine-Gordon equations with a variety of boundary conditions as well as any complex nonlinear physical problems across multiple disciplines.

Prediction of QCE using ANN and ANFIS for milling Alloy 2017A

Population growth and economic development, particularly in developing market nations, are driving up global energy consumption at an alarming rate. Despite increased wealth, growing demand presents new obstacles. Computer Numerical Control (CNC) machine tools are widely used in most metal machining processes due to their efficiency and repeatability in achieving high-precision machining. It has been shown that figuring out the best cutting parameters can improve the results of machining, leading to high efficiency and low costs. This study identifies and examines thoroughly the scientific contributions of the influence of strategies, machining sequences, and cutting parameters on surface quality, machining cost, and energy consumption (QCE) using artificial intelligence (ANN and ANFIS). The results show that the 3.10−3 architecture with the Bayesian Regularization (BR) algorithm is the optimal neural architecture that yields an overall mean square error (MSE) of 2.74 10−3. The correlation coefficients ( R2) for Etot, Ctot, and Ra are 0.9992, 1, and 0.9117 respectively. Similarly, for the adaptive neuro-fuzzy inference system (ANFIS), the optimal structure which gives a better error and better correlation is the {2, 2, 2} structure, and this for the three output variables (Etot, Ctot, and Ra). The correlation coefficient ( R2) for the variables Etot, Ctot, and Ra are respectively 0.95, 0.965, and 0.968. The results show that the use of the Bayesian Regularization algorithm with a multi-criteria output response can give good results when compared with the adaptive neuro-fuzzy inference system.

Advanced Misinformation Detection: A Bi-LSTM Model Optimized by Genetic Algorithms

The proliferation of misinformation, as insidious and pervasive as water, presents an unprecedented challenge to public discourse and comprehension. Often propagated to further specific ideologies or political objectives, misinformation not only misleads the populace but also fuels online advertising revenue generation. As such, the urgent need to pinpoint and eliminate misinformation from digital platforms has never been more critical. In response to this dilemma, this paper proposes a solution built on the backbone of massive data generation in today’s digital landscape. By leveraging advanced technologies, such as AI-driven systems with deep learning models and natural language processing capabilities, we can monitor and analyze an extensive scope of social media data. This, in turn, facilitates the identification of misinformation across multiple platforms and alerts users to potential propaganda. Central to our study is the development of misinformation classifiers based on a deep bi-directional long short-term memory (Bi-LSTM) model. This model is further enhanced by employing a genetic algorithm (GA), which automates the search for an optimal neural architecture, thereby significantly impacting the training behavior of the deep learning algorithm and the performance of the model being trained. To validate our approach, we compared the efficacy of our proposed model with nine traditional machine learning algorithms and a deep learning model rooted in long short-term memory (LSTM). The results affirmed the superiority of our GA-tuned Bi-LSTM model, which outperformed all other models in detecting misinformation with remarkable accuracy. Our intention with this paper is not to present our model as a comprehensive solution to misinformation but rather as a technological tool that can aid in the process, supplementing and bolstering the existing methodologies in the field of misinformation detection.

Neural Architecture Search via Proxy Validation.

This paper searches for the optimal neural architecture by minimizing a proxy of validation loss. Existing neural architecture search (NAS) methods used to discover the optimal neural architecture that best fits the validation examples given the up-to-date network weights. These intermediate validation results are invaluable but have not been fully explored. We propose to approximate the validation loss landscape by learning a mapping from neural architectures to their corresponding validate losses. The optimal neural architecture thus can be easily identified as the minimum of this proxy validation loss landscape. To improve the efficiency, a novel architecture sampling strategy is developed for the approximation of the proxy validation loss landscape. We also propose an operation importance weight (OIW) to balance the randomness and certainty of architecture sampling. The representation of neural architecture is learned through a graph autoencoder (GAE) over both architectures sampled during search and randomly generated architectures. We provide theoretical analyses on the validation loss estimator learned with our sampling strategy. Experimental results demonstrate that the proposed proxy validation loss landscape can be effective in both the differentiable NAS and the evolutionary-algorithm-based (EA-based) NAS.

Differentiable channel pruning guided via attention mechanism: a novel neural network pruning approach

Neural network pruning offers great prospects for facilitating the deployment of deep neural networks on computational resource limited devices. Neural architecture search (NAS) provides an efficient way to automatically seek appropriate neural architecture design for compressed model. It is observed that, for existing NAS-based pruning methods, there is usually a lack of layer information when searching the optimal neural architecture. In this paper, we propose a new NAS approach, namely, differentiable channel pruning method guided via attention mechanism (DCP-A), where the adopted attention mechanism is able to provide layer information to guide the optimization of the pruning policy. The training process is differentiable with Gumbel-softmax sampling, while parameters are optimized under a two-stage training procedure. The neural network block with the shortcut is dedicatedly designed, which is of help to prune the network not only on its width but also on its depth. Extensive experiments are performed to verify the applicability and superiority of the proposed method. Detailed analysis with visualization of the pruned model architecture shows that our proposed DCP-A learns explainable pruning policies.

Multi-View Graph Neural Architecture Search for Biomedical Entity and Relation Extraction.

Recently, graph neural architecture search (GNAS) frameworks have been successfully used to automatically design the optimal neural architectures for many problems such as node classification and graph classification. In the existing GNAS frameworks, the designed graph neural network (GNN) architectures learn the representation of homogenous graphs with one type of relationship connecting two nodes. However, multi-view graphs, where each view represents a type of relationship among nodes, are ubiquitous in the real world. The traditional GNAS frameworks learn the graph representation without considering the interactions between nodes and multiple relationships, so they fail to solve multi-view graph-based problems, such as multi-view graphs modelling the biomedical entity and relation extraction tasks. In this paper, we propose MVGNAS, a multi-view graph neural network automatic modelling framework for biomedical entity and relation extraction, to resolve this challenge. In MVGNAS, we propose an automatic multi-view representation learning to learn low-dimensional representations of nodes that capture multiple relationships in a multi-view graph, representing the first research work in literature to solve the problem of multi-view graph representation learning architecture search for biomedical entity and relation extraction tasks. The experimental results demonstrate that MVGNAS can achieve the best performance in biomedical entity and relation extraction tasks against the state-of-the-art baseline methods.

DDPNAS: Efficient Neural Architecture Search via Dynamic Distribution Pruning

Neural Architecture Search (NAS) has demonstrated state-of-the-art performance on various computer vision tasks. Despite the superior performance achieved, the efficiency and generality of existing methods are highly valued due to their high computational complexity and low generality. In this paper, we propose an efficient and unified NAS framework termed DDPNAS via dynamic distribution pruning, facilitating a theoretical bound on accuracy and efficiency. In particular, we first sample architectures from a joint categorical distribution. Then the search space is dynamically pruned and its distribution is updated every few epochs. With the proposed efficient network generation method, we directly obtain the optimal neural architectures on given constraints, which is practical for on-device models across diverse search spaces and constraints. The architectures searched by our method achieve remarkable top-1 accuracies, 97.56 and 77.2 on CIFAR-10 and ImageNet (mobile settings), respectively, with the fastest search process, i.e., only 1.8 GPU hours on a Tesla V100. Codes for searching and network generation are available at: https://openi.pcl.ac.cn/PCL_AutoML/XNAS.

An Insight into NeuroEvolution and Genetic Algorithms for Text Classification

Natural Language Processing (NLP) systems have, over the past decade, shifted from using rule-based techniques to using machine learning-based algorithms. This has led to the development of different architectures and models for different tasks. Some of these architectures include models like the transformers, the CNN and the RNN, which have now become ubiquitous in NLP. However, designing these neural network architectures usually requires in-depth analysis and knowledge of multiple domain areas involved with the problem at hand. In our work, we evaluate an alternative solution to this problem in the domain of text classification. Here, we suggest using the Genetic Algorithm with gradient descent (GAGD) and NeuroEvolution of Augmenting Topologies (NEAT) to search for an optimal neural architecture for the Reuters-21578 and 20 Newsgroups datasets. We evaluate and compare the results of the two algorithms against the current state-of-the-art architectures and provide insight into their performance.

AAtt-CNN: Automatic Attention-Based Convolutional Neural Networks for Hyperspectral Image Classification

Convolutional models have provided outstanding performance in the analysis of hyperspectral images (HSIs). These architectures are carefully designed to extract intricate information from non-linear features for classification tasks. Notwithstanding their results, model architectures are manually engineered and further optimized for generalized feature extraction. In general terms, deep architectures are time consuming for complex scenarios since they require fine tuning. Neural architecture search (NAS) has emerged as a suitable approach to tackle this shortcoming. In parallel, modern attention-based methods have boosted the recognition of sophisticated features. The search for optimal neural architectures combined with attention procedures motivates the development of this work. This paper develops a new method to automatically design and optimize convolutional neural networks (CNNs) for HSI classification using channel-based attention mechanisms. Specifically, one-dimensional (1D) and spectral-spatial (3D) classifiers are considered to handle the large amount of information contained in HSIs from different perspectives. Furthermore, the proposed AAtt-CNN method meets the requirement to lower the large computational overheads associated with architectural search. It is compared with current state-of-the-art (SOTA) classifiers. Our experiments, conducted using a wide range of HSI images, demonstrate that AAtt-CNN succeeds in finding optimal architectures for classification, leading to SOTA results.

MIND: Maximum Mutual Information Based Neural Decoder

We are assisting at a growing interest in the development of learning architectures with application to digital communication systems. Herein, we consider the detection/decoding problem. We aim at developing an optimal neural architecture for such a task. The definition of the optimal criterion is a fundamental step. We propose to use the mutual information (MI) of the channel input-output signal pair, which yields to the minimization of the a-posteriori information of the transmitted codeword given the communication channel output observation. The computation of the a-posteriori information is a formidable task, and for the majority of channels it is unknown. Therefore, it has to be learned. For such an objective, we propose a novel neural estimator based on a discriminative formulation. This leads to the derivation of the mutual information neural decoder (MIND). The developed neural architecture is capable not only to solve the decoding problem in unknown channels, but also to return an estimate of the average MI achieved with the coding scheme, as well as the decoding error probability. Several numerical results are reported and compared with maximum a-posteriori and maximum likelihood decoding strategies.

Toward Tailored Models on Private AIoT Devices: Federated Direct Neural Architecture Search

Neural networks often encounter various stringent resource constraints while deploying on edge devices. To tackle these problems with less human efforts, automated machine learning becomes popular in finding various neural architectures that fit diverse Artificial Intelligence of Things (AIoT) scenarios. Recently, to prevent the leakage of private information while enable automated machine intelligence, there is an emerging trend to integrate federated learning and neural architecture search (NAS). Although promising as it may seem, the coupling of difficulties from both tenets makes the algorithm development quite challenging. In particular, how to efficiently search the optimal neural architecture directly from massive nonindependent and identically distributed (non-IID) data among AIoT devices in a federated manner is a hard nut to crack. In this article, to tackle this challenge, by leveraging the advances in ProxylessNAS, we propose a federated direct neural architecture search (FDNAS) framework that allows for hardware-friendly NAS from non-IID data across devices. To further adapt to both various data distributions and different type of devices with heterogeneous embedded hardware platforms, inspired by meta-learning, a cluster federated direct neural architecture search (CFDNAS) framework is proposed to achieve device-aware NAS, in the sense that each device can learn a tailored deep learning model for its particular data distribution and hardware constraint. Extensive experiments on non-IID data sets have shown the state-of-the-art accuracy–efficiency tradeoffs achieved by the proposed solution in the presence of both data and device heterogeneity.

Federated Neural Architecture Search for Medical Data Security

Medical data widely exist in the hospital and personal life, usually across institutions and regions. They have essential diagnostic value and therapeutic significance. The disclosure of patient information causes people’s panic, therefore, medical data security solution is very crucial for intelligent health care. The emergence of federated learning (FL) provides an effective solution, which only transmits model parameters, breaking through the bottleneck of medical data sharing, protecting data security, and avoiding economic losses. Meanwhile, the neural architecture search (NAS) has become a popular method to automatically search the optimal neural architecture for solving complex practical problems. However, few papers have combined the FL and NAS for simultaneous privacy protection and model architecture selection. Convolutional neural network (CNN) has outstanding performance in the image recognition field. Combining CNN and fuzzy rough sets can effectively improve the interpretability of deep neural networks. This article aims to develop a multiobjective convolutional interval type-2 fuzzy rough FL model based on NAS (CIT2FR-FL-NAS) for medical data security with an improved multiobjective evolutionary algorithm. We test the proposed framework on the LC25000 lung and colon histopathological image dataset. Experimental verification demonstrates that the designed multiobjective CIT2FR-FL-NAS framework can achieve high accuracy superior to state-of-the-art models and reduce network complexity under the condition of protecting medical data security.

Differentiable neural architecture learning for efficient neural networks

Efficient neural networks has received ever-increasing attention with the evolution of convolutional neural networks (CNNs), especially involving their deployment on embedded and mobile platforms. One of the biggest problems to obtaining such efficient neural networks is efficiency, even recent differentiable neural architecture search (DNAS) requires to sample a small number of candidate neural architectures for the selection of the optimal neural architecture. To address this computational efficiency issue, we introduce a novel architecture parameterization based on scaled sigmoid function, and propose a general Differentiable Neural Architecture Learning (DNAL) method to obtain efficient neural networks without the need to evaluate candidate neural networks. Specifically, for stochastic supernets as well as conventional CNNs, we build a new channel-wise module layer with the architecture components controlled by a scaled sigmoid function. We train these neural network models from scratch. The network optimization is decoupled into the weight optimization and the architecture optimization, which avoids the interaction between the two types of parameters and alleviates the vanishing gradient problem. We address the non-convex optimization problem of efficient neural networks by the continuous scaled sigmoid method instead of the common softmax method. Extensive experiments demonstrate our DNAL method delivers superior performance in terms of efficiency, and adapts to conventional CNNs (e.g., VGG16 and ResNet50), lightweight CNNs (e.g., MobileNetV2) and stochastic supernets (e.g., ProxylessNAS). The optimal neural networks learned by DNAL surpass those produced by the state-of-the-art methods on the benchmark CIFAR-10 and ImageNet-1K dataset in accuracy, model size and computational complexity. Our source code is available at https://github.com/QingbeiGuo/DNAL.git.

On streaming disaster damage assessment in social sensing: A crowd-driven dynamic neural architecture searching approach

Motivated by the recent advances in Internet and communication techniques and the proliferation of online social media, social sensing has emerged as a new sensing paradigm to obtain timely observations of the physical world from “human sensors”. In this study, we focus on an emerging application in social sensing – streaming disaster damage assessment (DDA), which aims to automatically assess the damage severity of affected areas in a disaster event on the fly by leveraging the streaming imagery data about the disaster on social media. In particular, we study a dynamic optimal neural architecture searching (NAS) problem in streaming DDA applications. Our goal is to dynamically determine the optimal neural network architecture that accurately estimates the damage severity for each newly arrived image in the stream by leveraging human intelligence from the crowdsourcing systems. The present study is motivated by the observation that the neural network architectures in current DDA solutions are mainly designed by artificial intelligence (AI) experts, which often leads to non-negligible costs and errors given the dynamic nature of the streaming DDA applications and the lack of real-time annotations of the massive social media data inputs. Two critical technical challenges exist in solving our problem: (i) it is non-trivial to dynamically identify the optimal neural network architecture for each image on the fly without knowing its ground-truth label a priori; (ii) it is challenging to effectively leverage the imperfect crowd intelligence to correctly identify the optimal neural network architecture for each image. To address the above challenges, we developed CD-NAS, a dynamic crowd-AI collaborative NAS framework that carefully explores the human intelligence from crowdsourcing systems to solve the dynamic optimal NAS problem and optimize the performance of streaming DDA applications. The evaluation results from a real-world streaming DDA application show that CD-NAS consistently outperforms the state-of-the-art AI and NAS baselines by achieving the highest disaster damage assessment accuracy while maintaining the lowest computational cost.

Hierarchical Spatio-Temporal Modeling of Naturalistic Functional Magnetic Resonance Imaging Signals via Two-Stage Deep Belief Network With Neural Architecture Search.

Naturalistic functional magnetic resonance imaging (NfMRI) has become an effective tool to study brain functional activities in real-life context, which reduces the anxiety or boredom due to difficult or repetitive tasks and avoids the problem of unreliable collection of brain activity caused by the subjects’ microsleeps during resting state. Recent studies have made efforts on characterizing the brain’s hierarchical organizations from fMRI data by various deep learning models. However, most of those models have ignored the properties of group-wise consistency and inter-subject difference in brain function under naturalistic paradigm. Another critical issue is how to determine the optimal neural architecture of deep learning models, as manual design of neural architecture is time-consuming and less reliable. To tackle these problems, we proposed a two-stage deep belief network (DBN) with neural architecture search (NAS) combined framework (two-stage NAS-DBN) to model both the group-consistent and individual-specific naturalistic functional brain networks (FBNs), which reflected the hierarchical organization of brain function and the nature of brain functional activities under naturalistic paradigm. Moreover, the test-retest reliability and spatial overlap rate of the FBNs identified by our model reveal better performance than that of widely used traditional methods. In general, our model provides a promising method for characterizing hierarchical spatiotemporal features under the natural paradigm.

An automated deep learning pipeline based on advanced optimisations for leveraging spectral classification modelling

In deep learning (DL) modelling for spectral data, a major challenge is related to the choice of DL network architecture and the selection of the best hyperparameters. Often, slight changes to the neural architecture or its hyperparameter can have a direct influence on the model's performance, making its robustness questionable. To deal with it, this study presents an automated deep learning modelling based on advanced optimisation techniques involving Hyperband and Bayesian optimisation, to automatically find optimal neural architecture and its hyperparameters to reach robust DL models. The optimisation requires a base neural architecture to be initialized, however, later it automatically adjusts the neural architecture and the hyperparameters to reach the optimal model. Furthermore, to support the interpretation of the DL models, a wavelength weighing schema based on gradient-weighted class activation mapping (Grad-CAM) was implemented. The potential of the approach was showed on a real case of wheat variety classification with near-infrared spectral data. The performance of the classification was compared with that previously reported on the same dataset with different DL and chemometric approaches. The results showed that with the proposed approach a classification accuracy of 94.9% was reached, which was better than the best reported accuracy on the same data set i.e., 93%. Furthermore, the better performance was obtained with a simpler neural architecture compared to what was used in earlier studies. The automated deep learning based on advanced optimisation can support DL modelling of spectral data.

Toward a theory of machine learning

We define a neural network as a septuple consisting of (1) a state vector, (2) an input projection, (3) an output projection, (4) a weight matrix, (5) a bias vector, (6) an activation map and (7) a loss function. We argue that the loss function can be imposed either on the boundary (i.e. input and/or output neurons) or in the bulk (i.e. hidden neurons) for both supervised and unsupervised systems. We apply the principle of maximum entropy to derive a canonical ensemble of the state vectors subject to a constraint imposed on the bulk loss function by a Lagrange multiplier (or an inverse temperature parameter). We show that in an equilibrium the canonical partition function must be a product of two factors: a function of the temperature, and a function of the bias vector and weight matrix. Consequently, the total Shannon entropy consists of two terms which represent, respectively, a thermodynamic entropy and a complexity of the neural network. We derive the first and second laws of learning: during learning the total entropy must decrease until the system reaches an equilibrium (i.e. the second law), and the increment in the loss function must be proportional to the increment in the thermodynamic entropy plus the increment in the complexity (i.e. the first law). We calculate the entropy destruction to show that the efficiency of learning is given by the Laplacian of the total free energy, which is to be maximized in an optimal neural architecture, and explain why the optimization condition is better satisfied in a deep network with a large number of hidden layers. The key properties of the model are verified numerically by training a supervised feedforward neural network using the stochastic gradient descent method. We also discuss a possibility that the entire Universe at its most fundamental level is a neural network.

Use of Endoscopic Images in the Prediction of Submucosal Invasion of Gastric Neoplasms: Automated Deep Learning Model Development and Usability Study.

BackgroundIn a previous study, we examined the use of deep learning models to classify the invasion depth (mucosa-confined versus submucosa-invaded) of gastric neoplasms using endoscopic images. The external test accuracy reached 77.3%. However, model establishment is labor intense, requiring high performance. Automated deep learning (AutoDL) models, which enable fast searching of optimal neural architectures and hyperparameters without complex coding, have been developed.ObjectiveThe objective of this study was to establish AutoDL models to classify the invasion depth of gastric neoplasms. Additionally, endoscopist–artificial intelligence interactions were explored.MethodsThe same 2899 endoscopic images that were employed to establish the previous model were used. A prospective multicenter validation using 206 and 1597 novel images was conducted. The primary outcome was external test accuracy. Neuro-T, Create ML Image Classifier, and AutoML Vision were used in establishing the models. Three doctors with different levels of endoscopy expertise were asked to classify the invasion depth of gastric neoplasms for each image without AutoDL support, with faulty AutoDL support, and with best performance AutoDL support in sequence.ResultsThe Neuro-T–based model reached 89.3% (95% CI 85.1%-93.5%) external test accuracy. For the model establishment time, Create ML Image Classifier showed the fastest time of 13 minutes while reaching 82.0% (95% CI 76.8%-87.2%) external test accuracy. While the expert endoscopist's decisions were not influenced by AutoDL, the faulty AutoDL misled the endoscopy trainee and the general physician. However, this was corrected by the support of the best performance AutoDL model. The trainee gained the most benefit from the AutoDL support.ConclusionsAutoDL is deemed useful for the on-site establishment of customized deep learning models. An inexperienced endoscopist with at least a certain level of expertise can benefit from AutoDL support.

Multiclass machine learning classification of functional brain images for Parkinson's disease stage prediction

AbstractWe analyzed a data set containing functional brain images from 6 healthy controls and 196 individuals with Parkinson's disease (PD), who were divided into five stages according to illness severity. The goal was to predict patients' PD illness stages by using their functional brain images. We employed the following prediction approaches: multivariate statistical methods (linear discriminant analysis, support vector machine, decision tree, and multilayer perceptron [MLP]), ensemble learning models (random forest [RF] and adaptive boosting), and deep convolutional neural network (CNN). For statistical and ensemble models, various feature extraction approaches (principal component analysis [PCA], multilinear PCA, intensity summary statistics [IStat], and Laws' texture energy measure) were employed to extract features, the synthetic minority over‐sampling technique was used to address imbalanced data, and the optimal combination of hyperparameters was found using a grid search. For CNN modeling, we applied an image augmentation technique to increase and balance data sizes over different disease stages. We adopted transfer learning to incorporate pretrained VGG16 weights and architecture into the model fitting, and we also tested a state‐of‐the‐art machine learning model that could automatically generate an optimal neural architecture. We found that IStat consistently outperformed other feature extraction approaches. MLP and RF were the analytic approaches with the highest prediction accuracy rate for multivariate statistical and ensemble learning models, respectively. Overall, the deep CNN model with pretrained VGG16 weights and architecture outperformed other approaches; it captured critical features from imaging, effectively distinguished between normal controls and patients with PD, and achieved the highest classification accuracy.