GA vs PSO & Deep Learning

GA is inspired by genetics and uses selection, crossover, and mutation. PSO is inspired by social behavior and uses velocity and position updates. GAs are more explorative; PSO tends to converge faster.

Transfer learning is a method where a model trained on one task (e.g., ImageNet classification) is reused and adapted for another task. This reduces the need for large datasets and allows faster development of models for new applications.

Frozen layers are layers whose weights are fixed during training. They retain the pre-learned features from the source dataset and help maintain general knowledge, especially in the early convolutional layers.

Trainable layers are updated during fine-tuning. They are typically the last few layers in the network, and they adapt the model to the new task by learning task-specific patterns from the new dataset.

Fine-tuning is the process of continuing training a pre-trained model on a new task or dataset. It usually involves unfreezing some layers and updating them to specialize the model without losing its general capabilities.

Additive fine-tuning involves adding new layers or adapter modules to the pre-trained model while keeping the original weights unchanged. These new components are trained for the target task, preserving the base model's stability.

VGG16 uses 3x3 convolutional filters, consistent architecture, and depth (16 layers) to learn complex image features. Its simplicity and uniformity make it widely used, despite its large number of parameters (138 million).

GoogleNet uses Inception modules to process features at multiple scales using 1x1, 3x3, and 5x5 convolutions in parallel. It significantly reduces the number of parameters while maintaining high accuracy on image classification tasks.

An Inception module is a network block that applies multiple convolutions (1x1, 3x3, 5x5) and pooling operations in parallel, then concatenates the outputs. This allows the model to learn both local and global features simultaneously.

1x1 convolutions are used for dimensionality reduction. They reduce the number of input channels before applying larger filters, making the model more efficient and reducing computational cost.