Deep Learning - Topics
Semantic vs instance vs panoptic segmentation
Backpropagation through time: why RNNs forget long sequences
Skip connections: why ResNets can train hundreds of layers
Optical flow: Lucas-Kanade vs learned dense flow
Catastrophic forgetting when networks are fine-tuned sequentially
CNN receptive fields: how stacked 3x3 convolutions beat one 7x7
Attention is not just for text: how transformers displaced recurrence
Vanishing gradients and why ReLU replaced sigmoid in deep stacks
Weight initialization: Xavier vs He and why starting from zeros fails
Data augmentation for vision: cutout and mixup do more than flipping
Batch normalization vs layer normalization in sequence models
3D reconstruction with structure from motion and neural radiance fields
Why do two runs of the same training script give different results?
Overfitting explained: how do you tell a model just memorized the answers?
Feature engineering vs end-to-end deep learning: which still wins in 2025?
Is gradient descent still the backbone of every model people build today?