Deep Learning

Deep learning is neural networks taken to scale: many layers, large datasets, and hardware that makes gradient descent on millions of parameters routine. This map sits one zoom level above the Neural Networks guide. It assumes you can picture a layer and a loss, then organizes the field into model families, representation learning, and the engineering that makes training stable.

The first branch is still Neural Networks — depth, width, and the idea that hierarchical features beat hand-designed ones for perception. From there, split vision and sequence. CNNs vs RNNs is the classic contrast: weight sharing over space versus recurrence over time. Convolutions remain strong for images and other grids. Recurrent nets (and their gated variants) were the default for language and time series until attention made parallel sequence modeling practical. Knowing both is useful even if you mostly use transformers now; many production systems still mix a conv stem or a recurrent head with a larger backbone.

Transformers are the current default architecture for language and, increasingly, vision. Self-attention lets every token look at every other token, positional encodings restore order, and stacked blocks with residual connections and layer norm train more reliably than deep vanilla RNNs. Encoder-only models shine at understanding; decoder-only models at generation; encoder-decoder models at transduction. Pretraining on unlabeled text or images, then fine-tuning, is the playbook. This branch is where “deep learning” stopped meaning “a deeper CNN” and started meaning “foundation models.”

Training at this scale is its own subject: mixed precision, distributed data parallel, checkpointing, and the stubborn details of learning-rate warmup, batch size, and data quality. Representation learning — embeddings, contrastive losses, autoencoders — explains why unlabeled data helps. Evaluation is no longer a single accuracy number; you care about calibration, robustness, and whether the model generalizes beyond the benchmark.

Walk the outline in the order that matches your gap. If CNNs and backprop are rusty, open the Neural Networks map first. If you want to implement things, jump to Python for Data Science for the library layer (NumPy, PyTorch-adjacent workflows) that these papers assume. The canvas and the prose describe the same tree so you can search, skim, or zoom without losing the hierarchy.