Deep Learning Vision Architectures: From LeNet to Vision Transformers Explained

The Evolution of Vision Architectures

• 💡 The course traces the evolution of deep learning vision models, starting from foundational architectures like LeNet and AlexNet and progressing to modern models such as ResNet, EfficientNet, and Vision Transformers.
• 🧠 Each architecture is explained through its design philosophy, historical context, and the problems it solved, emphasizing the "why" behind their structures.

Foundational Architectures: LeNet and AlexNet

• 🧱 LeNet-5 (1998) introduced the convolution-pooling pattern and learnable kernels, demonstrating hierarchical feature learning from pixels to characters and achieving real-world impact in ATM check reading.
• 🚀 AlexNet (2012) proved depth matters by leveraging GPUs, ReLU activations, dropout, and data augmentation, overcoming computational barriers and sparking the modern AI renaissance.

Advancements in Depth and Simplicity: VGG and GoogLeNet

• 🧱 VGG (2014) embraced simplicity and depth by using repeated 3x3 convolution layers, showing that uniform depth could outperform complex, manually engineered structures and achieving strong performance on ImageNet.
• 🧩 GoogLeNet (Inception) (2014) introduced parallel filters (1x1, 3x3, 5x5) within inception modules to capture multi-scale features efficiently, using bottleneck 1x1 convolutions to control computational cost.

Enabling Deeper Networks: Highway Networks and ResNet

• 🛣️ Highway Networks (2015) introduced trainable gates to skip connections, allowing information to bypass transformations and enabling the training of networks with hundreds of layers by preserving identity and mitigating vanishing gradients.
• 💡 ResNet (2015) revolutionized depth by learning only the residual changes to an identity mapping via skip connections, ensuring signal flow and enabling the training of truly deep models without performance degradation.

Refinements and New Paradigms: DenseNet, Exception, MobileNets, and EfficientNet

• 🔗 DenseNet (2016) connected each layer to all subsequent ones via concatenation, maximizing feature reuse, reducing redundancy, and alleviating vanishing gradients by valuing every intermediate feature map.
• 🔀 Exception (2016) distilled Inception into depthwise separable convolutions, decoupling spatial and channel learning for improved parameter efficiency and performance.
• 📱 MobileNets (2017) focused on efficiency for mobile and embedded devices using depthwise separable convolutions and width/resolution multipliers to create lightweight, adaptable models.
• ⚖️ EfficientNet (2019) introduced compound scaling, jointly tuning depth, width, and resolution based on principled search, achieving state-of-the-art accuracy with significantly fewer parameters and faster inference.

Beyond Convolution: Vision Transformers

• 🚀 Vision Transformers (ViT) (2020) extend the transformer architecture to vision by treating images as sequences of patches, using self-attention to capture global context while retaining identity preservation through residual connections, outperforming CNNs on large-scale datasets.