Deep Learning

Deep Learning for Image Classification using CIFAR-10 Dataset Name: Student Number: Student Email: Abstract This project uses deep learning techniques to classify images from the CIFAR 10 dataset. The aim is to construct and train a convolutional neural network (CNN) model to categorise the images into ten different classes. It is compared with baseline models, then hyperparameter tuned, and the model architecture is refined to determine its effectiveness. It also includes the implementation of an actual applied AI problem for image classification, utilising the trained model in practice. Finally, results indicate that the proposed deep learning model achieves better accuracy and robustness than baseline performance. Quantitative results demonstrate significant improvements, and qualitative results offer visual cues as to how the model was making decisions. Table of Contents TOC \o "1-3" \h \z \u Abstract PAGEREF _Toc197167156 \h 1I. Introduction PAGEREF _Toc197167157 \h 21.1 Background PAGEREF _Toc197167158 \h 21.2 Aims and Objectives PAGEREF _Toc197167159 \h 21.3 Input and Output PAGEREF _Toc197167160 \h 31.4 Literature Review PAGEREF _Toc197167161 \h 3II. Dataset PAGEREF _Toc197167162 \h 32.1 Dataset Description PAGEREF _Toc197167163 \h 32.2 Data Source PAGEREF _Toc197167164 \h 42.3 Data Preprocessing PAGEREF _Toc197167165 \h 43.1 Deep Learning Model PAGEREF _Toc197167166 \h 43.2 Loss Function PAGEREF _Toc197167167 \h 53.3 Optimisation Process PAGEREF _Toc197167168 \h 53.4 Mathematical Formulation PAGEREF _Toc197167169 \h 6IV. Experiments PAGEREF _Toc197167170 \h 64.1 Evaluation Metrics PAGEREF _Toc197167171 \h 64.2 Quantitative Results PAGEREF _Toc197167172 \h 74.3 Qualitative Results PAGEREF _Toc197167173 \h 74.4 Ablation Studies PAGEREF _Toc197167174 \h 74.5 Application PAGEREF _Toc197167175 \h 74.6 Comparison PAGEREF _Toc197167176 \h 84.7 Hyperparameter Tuning PAGEREF _Toc197167177 \h 8V. Conclusion PAGEREF _Toc197167178 \h 8VI. References PAGEREF _Toc197167179 \h 9 I. Introduction 1.1 Background Deep learning, which can solve complicated problems in all disciplines, has excelled at image recognition. CNNs excel at handling spatial picture data, which has a clear hierarchy. CNNs automatically learn feature representations by abstracting features at several levels, making them powerful for image classification, object identification, and segmentation [4]. Their achievement does not solve the problem of boosting model accuracy without overfitting and keeping training time efficient on sparse label data. The CIFA-10 dataset is a solid deep learning image classification benchmark. Second, we categorized 60,000 32x32 color photos into ten categories to evaluate image classification systems. This project investigates how Deep Learning models (CNNs) can improve CIFAR-10 classification accuracy qualitatively and quantitatively [3]. 1.2 Aims and Objectives The specific objectives of the project are: * To design and train a CNN model to classify images into one of the ten categories of CIFAR-10. * To compare the performance of the proposed model with existing methods. * To fine-tune the model architecture and hyperparameters to achieve the highest possible accuracy. * To create a functional AI application that uses the trained model to classify images in real time. 1.3 Input and Output Deep learning model takes in a 32x32 pixel RGB colour image. In CIFAR-10 dataset, the content is a CIFAR-10 dataset. The image is classified to one of ten classes (animals, vehicles as aeroplanes, automobiles, birds, cats, dogs, frogs, horses, ships or trucks, deer). At last, it gives a predicted label for each image among the ten described [5]. Even if you give the model an input image of a dog, you will get the class label "dog" as the output. 1.4 Literature Review CNNS excels at picture classification, according to research. Krzheksky et al. proposed AlexNet, a CNN architecture that outperforms existing ImageNet algorithms, in 2012. CNNS have become the standard image-related task. Other architectures like VGG, ResNet, Inception, etc., would have included deeper layers, residual connections, and parallel convolutions [1]. In smaller datasets like CIFAR-10, these models always overfit. II. Datas...