Yapay Sinir Ağları

Neuron

1. The Artificial NeuronWeighted sum, then nonlinearity: the single unit everything stacks from.1–2 sa
2. Affine Transformationsw dot x plus b, as geometry: the linear stage before the bend.1–2 sa
3. Activation FunctionsThe bend that makes depth matter: why linearity stacks flat, and sigmoid/ReLU/GELU roles.1–2 sa

Architecture

1. Layers and Multilayer CompositionStacking as function composition: depth builds features from features.1–2 sa
2. The Multilayer PerceptronThe full map: layers, widths, parameter counts, and what an MLP can represent.1–2 sa
3. Forward PropagationOne honest pass: six weights, three biases, every value worked.1–2 sa
4. Tensors and Batched ComputationShapes, batch dimensions, and broadcasting: how a layer processes many examples at once.1–2 sa
5. Universal Approximation and the Role of DepthWhy one wide hidden layer can approximate any continuous function, and why depth is still more efficient.1–2 sa

Gradients

1. Computational GraphsNodes compute, edges carry: the shared structure behind forward and backward passes.1–2 sa
2. The Chain Rule in Neural NetsLocal derivatives along paths: how one edge's nudge propagates to the loss.1–2 sa
3. BackpropagationCredit flows backward: exact gradients for every edge via the chain rule, once.1–2 sa
4. Gradient Flow and DepthSignals fade or explode across layers: multiplicative paths and their perils.1–2 sa
5. Automatic DifferentiationHow frameworks compute gradients: forward-mode vs reverse-mode accumulation over a recorded graph.1–2 sa

Training

1. InitializationWhere weights start decides whether signals survive: scale intuition for init schemes.1–2 sa
2. Normalization LayersRe-centering activations mid-net: BatchNorm and LayerNorm roles without the folklore.1–2 sa
3. Training DynamicsReading the trace: what loss curves confess about rate, data, and architecture.1–2 sa
4. Learning-Rate SchedulesRates that adapt over time: decay, warmup, and what each phase buys.1–2 sa
5. DropoutNoise as regularization: training thinned nets, testing the whole one.1–2 sa
6. Residual ConnectionsSkip paths that let depth grow: identity shortcuts as gradient highways.1–2 sa
7. Training InstabilityWhen training breaks: divergence, NaNs, and reading the wreckage.1–2 sa