Search results for: ReLU

Authors: Apoorva Vinod, Archana Mathur, Snehanshu Saha

Abstract:

Though there exists a plethora of Activation Functions (AFs) used in single and multiple hidden layer Neural Networks (NN), their behavior always raised curiosity, whether used in combination or singly. The popular AFs –Sigmoid, ReLU, and Tanh–have performed prominently well for shallow and deep architectures. Most of the time, AFs are used singly in multi-layered NN, and, to the best of our knowledge, their performance is never studied and analyzed deeply when used in combination. In this manuscript, we experiment with multi-layered NN architecture (both on shallow and deep architectures; Convolutional NN and VGG16) and investigate how well the network responds to using two different AFs (Sigmoid-Tanh, Tanh-ReLU, ReLU-Sigmoid) used alternately against a traditional, single (Sigmoid-Sigmoid, Tanh-Tanh, ReLUReLU) combination. Our results show that using two different AFs, the network achieves better accuracy, substantially lower loss, and faster convergence on 4 computer vision (CV) and 15 Non-CV (NCV) datasets. When using different AFs, not only was the accuracy greater by 6-7%, but we also accomplished convergence twice as fast. We present a case study to investigate the probability of networks suffering vanishing and exploding gradients when using two different AFs. Additionally, we theoretically showed that a composition of two or more AFs satisfies Universal Approximation Theorem (UAT).

Keywords: activation function, universal approximation function, neural networks, convergence

Procedia PDF Downloads 125

Authors: Zhifeng Kong

Abstract:

Over-parameterized neural networks have attracted a great deal of attention in recent deep learning theory research, as they challenge the classic perspective of over-fitting when the model has excessive parameters and have gained empirical success in various settings. While a number of theoretical works have been presented to demystify properties of such models, the convergence properties of such models are still far from being thoroughly understood. In this work, we study the convergence properties of training two-hidden-layer partially over-parameterized fully connected networks with the Rectified Linear Unit activation via gradient descent. To our knowledge, this is the first theoretical work to understand convergence properties of deep over-parameterized networks without the equally-wide-hidden-layer assumption and other unrealistic assumptions. We provide a probabilistic lower bound of the widths of hidden layers and proved linear convergence rate of gradient descent. We also conducted experiments on synthetic and real-world datasets to validate our theory.

Keywords: over-parameterization, rectified linear units ReLU, convergence, gradient descent, neural networks

Procedia PDF Downloads 105

Authors: Gulfam Haider, sana danish

Abstract:

Convolutional Neural Networks (CNNs) have emerged as powerful tools for image classification, and the choice of optimizers profoundly affects their performance. The study of optimizers and their adaptations remains a topic of significant importance in machine learning research. While numerous studies have explored and advocated for various optimizers, the efficacy of these optimization techniques is still subject to scrutiny. This work aims to address the challenges surrounding the effectiveness of optimizers by conducting a comprehensive analysis and evaluation. The primary focus of this investigation lies in examining the performance of different optimizers when employed in conjunction with the popular activation function, Rectified Linear Unit (ReLU). By incorporating ReLU, known for its favorable properties in prior research, the aim is to bolster the effectiveness of the optimizers under scrutiny. Specifically, we evaluate the adjustment of these optimizers with both the original Softmax activation function and the modified ReLU activation function, carefully assessing their impact on overall performance. To achieve this, a series of experiments are conducted using a well-established benchmark dataset for image classification tasks, namely the Canadian Institute for Advanced Research dataset (CIFAR-10). The selected optimizers for investigation encompass a range of prominent algorithms, including Adam, Root Mean Squared Propagation (RMSprop), Adaptive Learning Rate Method (Adadelta), Adaptive Gradient Algorithm (Adagrad), and Stochastic Gradient Descent (SGD). The performance analysis encompasses a comprehensive evaluation of the classification accuracy, convergence speed, and robustness of the CNN models trained with each optimizer. Through rigorous experimentation and meticulous assessment, we discern the strengths and weaknesses of the different optimization techniques, providing valuable insights into their suitability for image classification tasks. By conducting this in-depth study, we contribute to the existing body of knowledge surrounding optimizers in CNNs, shedding light on their performance characteristics for image classification. The findings gleaned from this research serve to guide researchers and practitioners in making informed decisions when selecting optimizers and activation functions, thus advancing the state-of-the-art in the field of image classification with convolutional neural networks.

Keywords: deep neural network, optimizers, RMsprop, ReLU, stochastic gradient descent

Procedia PDF Downloads 59

Authors: Andrew N. Saylor, James R. Peters

Abstract:

Scoliosis is a complex 3D deformity of the thoracic and lumbar spines, clinically diagnosed by measurement of a Cobb angle of 10 degrees or more on a coronal X-ray. The Cobb angle is the angle made by the lines drawn along the proximal and distal endplates of the respective proximal and distal vertebrae comprising the curve. Traditionally, Cobb angles are measured manually using either a marker, straight edge, and protractor or image measurement software. The task of measuring the Cobb angle can also be represented by a function taking the spine geometry rendered using X-ray imaging as input and returning the approximate angle. Although the form of such a function may be unknown, it can be approximated using artificial neural networks (ANNs). The performance of ANNs is affected by many factors, including the choice of activation function and network architecture; however, the effects of these parameters on the accuracy of scoliotic deformity measurements are poorly understood. Therefore, the objective of this study was to systematically investigate the effect of ANN architecture and activation function on Cobb angle measurement from the coronal X-rays of scoliotic subjects. The data set for this study consisted of 609 coronal chest X-rays of scoliotic subjects divided into 481 training images and 128 test images. These data, which included labeled Cobb angle measurements, were obtained from the SpineWeb online database. In order to normalize the input data, each image was resized using bi-linear interpolation to a size of 500 × 187 pixels, and the pixel intensities were scaled to be between 0 and 1. A fully connected (dense) ANN with a fixed cost function (mean squared error), batch size (10), and learning rate (0.01) was developed using Python Version 3.7.3 and TensorFlow 1.13.1. The activation functions (sigmoid, hyperbolic tangent [tanh], or rectified linear units [ReLU]), number of hidden layers (1, 3, 5, or 10), and number of neurons per layer (10, 100, or 1000) were varied systematically to generate a total of 36 network conditions. Stochastic gradient descent with early stopping was used to train each network. Three trials were run per condition, and the final mean squared errors and mean absolute errors were averaged to quantify the network response for each condition. The network that performed the best used ReLU neurons had three hidden layers, and 100 neurons per layer. The average mean squared error of this network was 222.28 ± 30 degrees2, and the average mean absolute error was 11.96 ± 0.64 degrees. It is also notable that while most of the networks performed similarly, the networks using ReLU neurons, 10 hidden layers, and 1000 neurons per layer, and those using Tanh neurons, one hidden layer, and 10 neurons per layer performed markedly worse with average mean squared errors greater than 400 degrees2 and average mean absolute errors greater than 16 degrees. From the results of this study, it can be seen that the choice of ANN architecture and activation function has a clear impact on Cobb angle inference from coronal X-rays of scoliotic subjects.

Keywords: scoliosis, artificial neural networks, cobb angle, medical imaging

Procedia PDF Downloads 95

Authors: Diego Aguirre, Olac Fuentes

Abstract:

In this paper, we present a taxonomy of weight initialization schemes used in deep learning. We survey the most representative techniques in each class and compare them in terms of overhead cost, convergence rate, and applicability. We also introduce a new weight initialization scheme. In this technique, we perform an initial feedforward pass through the network using an initialization mini-batch. Using statistics obtained from this pass, we initialize the weights of the network, so the following properties are met: 1) weight matrices are orthogonal; 2) ReLU layers produce a predetermined number of non-zero activations; 3) the output produced by each internal layer has a unit variance; 4) weights in the last layer are chosen to minimize the error in the initial mini-batch. We evaluate our method on three popular architectures, and a faster converge rates are achieved on the MNIST, CIFAR-10/100, and ImageNet datasets when compared to state-of-the-art initialization techniques.

Keywords: deep learning, image classification, supervised learning, weight initialization

Procedia PDF Downloads 105

Authors: Aisulu Tileukulova, Zhexebay Dauren

Abstract:

Radar plays an important role because it is widely used in civil and military fields. Target detection is one of the most important radar applications. The accuracy of detecting inconspicuous aerial objects in radar facilities is lower against the background of noise. Convolutional neural networks can be used to improve the recognition of this type of aerial object. The purpose of this work is to develop an algorithm for recognizing aerial objects using convolutional neural networks, as well as training a neural network. In this paper, the structure of a convolutional neural network (CNN) consists of different types of layers: 8 convolutional layers and 3 layers of a fully connected perceptron. ReLU is used as an activation function in convolutional layers, while the last layer uses softmax. It is necessary to form a data set for training a neural network in order to detect a target. We built a Confusion Matrix of the CNN model to measure the effectiveness of our model. The results showed that the accuracy when testing the model was 95.7%. Classification of echo signals using CNN shows high accuracy and significantly speeds up the process of predicting the target.

Keywords: radar, neural network, convolutional neural network, echo signals

Procedia PDF Downloads 311

Authors: Arya Prakash Padhi, Anupam Chakrabarti, Rajib Chowdhury

Abstract:

Topology or layout optimization often gives better performing economic structures and is very helpful in the conceptual design phase. But traditionally it is being done in finite element-based optimization schemes which, although gives a good result, is very time-consuming especially in 3D structures. Among other alternatives machine learning, especially deep learning-based methods, have a very good potential in resolving this computational issue. Here convolutional neural network (3D-CNN) based variational auto encoder (VAE) is trained using a dataset generated from commercially available topology optimization code ABAQUS Tosca using solid isotropic material with penalization (SIMP) method for compliance minimization. The encoded data in latent space is then fed to a 3D generative adversarial network (3D-GAN) to generate the outcome in 64x64x64 size. Here the network consists of 3D volumetric CNN with rectified linear unit (ReLU) activation in between and sigmoid activation in the end. The proposed network is seen to provide almost optimal results with significantly reduced computational time, as there is no iteration involved.

Keywords: 3D generative adversarial network, deep learning, structural topology optimization, variational auto encoder

Procedia PDF Downloads 136

Authors: Nonlapun Meenil, Pisitpong Intarapong, Thitima Wongsheree, Pranchalee Samanpiboon

Abstract:

In Thailand, once pineapples are harvested, they must be classified into two classes based on their sweetness: sweet and unsweet. This paper has studied and developed the assessment of internal quality of pineapples using a low-cost compact spectroscopy sensor according to the Spectroscopy approach and Neural Network (NN). During the experiments, Batavia pineapples were utilized, generating 100 samples. The extracted pineapple juice of each sample was used to determine the Soluble Solid Content (SSC) labeling into sweet and unsweet classes. In terms of experimental equipment, the sensor cover was specifically designed to install the sensor and light source to read the reflectance at a five mm depth from pineapple flesh. By using a spectroscopy sensor, data on visible and near-infrared reflectance (Vis-NIR) were collected. The NN was used to classify the pineapple classes. Before the classification step, the preprocessing methods, which are Class balancing, Data shuffling, and Standardization were applied. The 510 nm and 900 nm reflectance values of the middle parts of pineapples were used as features of the NN. With the Sequential model and Relu activation function, 100% accuracy of the training set and 76.67% accuracy of the test set were achieved. According to the abovementioned information, using a low-cost compact spectroscopy sensor has achieved favorable results in classifying the sweetness of the two classes of pineapples.

Keywords: neural network, pineapple, soluble solid content, spectroscopy

Procedia PDF Downloads 41

Authors: Luis C. Parra

Abstract:

The significant wave height prediction is an issue of great interest in the field of coastal activities because of the non-linear behavior of the wave height and its complexity of prediction. This study aims to present a machine learning model to forecast the significant wave height of the oceanographic wave measuring buoys anchored at Mooloolaba of the Queensland Government Data. Modeling was performed by a multilayer perceptron neural network-genetic algorithm (GA-MLP), considering Relu(x) as the activation function of the MLPNN. The GA is in charge of optimized the MLPNN hyperparameters (learning rate, hidden layers, neurons, and activation functions) and wrapper feature selection for the window width size. Results are assessed using Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The GAMLPNN algorithm was performed with a population size of thirty individuals for eight generations for the prediction optimization of 5 steps forward, obtaining a performance evaluation of 0.00104 MSE, 0.03222 RMSE, 0.02338 MAE, and 0.71163% of MAPE. The results of the analysis suggest that the MLPNNGA model is effective in predicting significant wave height in a one-step forecast with distant time windows, presenting 0.00014 MSE, 0.01180 RMSE, 0.00912 MAE, and 0.52500% of MAPE with 0.99940 of correlation factor. The GA-MLP algorithm was compared with the ARIMA forecasting model, presenting better performance criteria in all performance criteria, validating the potential of this algorithm.

Keywords: significant wave height, machine learning optimization, multilayer perceptron neural networks, evolutionary algorithms

Procedia PDF Downloads 67