A Comparative Study of Deep Learning Models on Benchmark Datasets

Abstract

This paper presents a systematic experimental study spanning the major paradigms of modern deep learning as implemented in a structured laboratory curriculum. Beginning with single-perceptron logic gates and multi-layer perceptrons (MLPs), the work progressively advances through Convolutional Neural Networks (CNNs) evaluated on CIFAR-10 and MNIST, convolutional feature-map visualisation, guided backpropagation, and the analysis of classical CNN architectures (LeNet, AlexNet, ZFNet, VGGNet, GoogLeNet, ResNet). The study continues with autoencoders-undercomplete, overcomplete, denoising, sparse, and contractive variants-trained on CIFAR-10, followed by a comparison of autoencoder-based and PCA-based dimensionality reduction. Generative Adversarial Networks (GANs and DCGANs) are implemented on MNIST and CIFAR-10 to demonstrate adversarial image synthesis. Sequence modelling is addressed through RNNs, LSTMs, GRUs, encoderdecoder architectures, and attention mechanisms applied to character- and word-level prediction tasks. Finally, pre-trained BERT models are fine-tuned for masked-language modelling and sentiment classification. Across all experiments, CNNs substantially outperform ANNs on spatial data, denoising autoencoders achieve robust image reconstruction under additive Gaussian noise, DCGANs generate plausible CIFAR-10-like images after sufficient training, and LSTM/GRU networks capture long-range dependencies that defeat vanilla RNNs. The integrated study provides a practitioner-oriented reference connecting theoretical foundations to reproducible implementation.