Pembelajaran Mendalam

Each layer transforms its input, and training adjusts the network's parameters so its outputs better match a defined objective. Depth describes the model's structure; it does not prove human-like understanding.

Key takeaways

• Multiple layers and nonlinear transformations let a network learn complex representations.
• Training updates parameters; inference uses the model to process new inputs.
• Choose models using held-out task performance and practical constraints, not depth alone.

A network turns an input into numbers that later layers can use. For an image classifier, the input might be pixel values and the output might be a score for each category. Hidden layers sit between input and output. They combine learned weights with nonlinear activation functions; simply stacking linear transformations would still give a linear transformation. Training and using the model are different operations. During training, a forward pass produces predictions, a loss function measures error, and backpropagation calculates gradients. An optimizer uses those gradients to update parameters. During inference, the trained model processes a new input without necessarily updating its weights. A complete experiment includes data preparation, a model, a loss, an optimizer, and evaluation on examples excluded from training. PyTorch's beginner tutorial demonstrates this workflow with clothing-image classification. Start with a small reproducible task, record the data split and settings, and inspect mistakes rather than looking only at the final accuracy number. Lower training loss is not proof that a model will work on new data. A network can fit patterns that are specific to its training examples. Keep evaluation data separate, investigate duplicates across splits, and test the conditions the application will encounter. The useful question is whether the model generalizes to the intended task, not whether it has the most layers.

1. Construct an illustrative fully connected network with two input values, a first hidden layer of three units, a second hidden layer of two units, and one output unit. Give every hidden and output unit a bias.
2. The first hidden layer has 2 × 3 weights and 3 biases: 9 parameters. The second has 3 × 2 weights and 2 biases: 8 parameters.
3. The output has 2 × 1 weights and 1 bias: 3 parameters. The network therefore has 9 + 8 + 3 = 20 trainable parameters. Apply nonlinear activations between the hidden layers.

This constructed example shows what parameters and layers mean. It does not demonstrate a trained model or useful accuracy. To test usefulness, choose a task, train the network, and evaluate it against a simpler baseline on unseen examples.