lesson_6_flashcards
What is a convolutional neural network (CNN)?
A neural network architecture that uses convolutional layers to extract spatial features, alternating with pooling layers for downsampling.
What is a receptive field in CNNs?
The region of the input image that influences a particular activation in deeper layers, growing with network depth.
What is transfer learning in deep learning?
A method to reuse features learned from large datasets like ImageNet for new tasks, reducing the need for large labeled datasets.
What are advanced convolutional network architectures?
Architectures like AlexNet, VGG, Inception, and ResNet that introduce innovations like small filters, residual connections, and modular designs for scalability.
What is the role of skip connections in ResNet?
They allow gradients to bypass layers, improving gradient flow and enabling the training of very deep networks.
What are the benefits of small convolution filters (e.g., 3x3) over large ones?
They reduce parameters, improve efficiency, and achieve the same receptive field depth with stacked layers.
What is the backwards pass in a convolution layer?
A process to compute gradients for weights and inputs during backpropagation using the chain rule and cross-correlation operations.
What is the role of data augmentation in CNN training?
Enhances generalization by artificially expanding datasets through transformations like rotations, flips, and noise addition.
What is the main limitation of transfer learning?
It performs poorly when the target task is significantly different from the source task, such as natural images to sketches.
What is the difference between cross-correlation and convolution in forward passes?
Cross-correlation does not flip the kernel, while convolution does; deep learning often uses cross-correlation for simplicity.
What are modular designs in advanced CNN architectures?
Repeated patterns of layers, such as 3x3 convolutions in VGG or parallel filters in Inception, to increase depth and scalability.
What is the importance of initialization in CNNs?
Proper initialization ensures stable gradient flow and faster convergence during training.
How does transfer learning reduce computation?
By freezing pre-trained convolutional layers and updating only the final layers, it reduces the parameters to train on small datasets.
What is semi-supervised learning in the context of CNNs?
A learning paradigm where models are trained on a small labeled dataset alongside a larger unlabeled dataset.
How do residual connections improve optimization?
By enabling identity mappings, they prevent gradient degradation and allow efficient training of very deep networks.
