cchakri.aiLearn All courses
// COURSE · AI & DATA

Deep Learning

Learn tensor computation, neural network training, convolutional and sequence models, and deployment practice.

01

Neural network foundations

Syllabus

  1. 01Tensors and computational graphs
  2. 02Perceptrons and activation functions
  3. 03Loss, gradients, and backpropagation

Detailed content

Tensors and computational graphs

Tensors generalize scalars, vectors, and matrices to higher dimensions. Neural network frameworks record operations in a computational graph so gradients can be calculated. Track tensor shape, device, and dtype at every layer. Shape mismatches and unintended broadcasting are frequent sources of silent bugs.

Perceptrons and activation functions

A neuron combines weighted inputs and a bias, then applies a nonlinear activation. Stacking linear layers without nonlinearities still produces a linear function, so activations give a network expressive power. ReLU is common for hidden layers; output activations depend on whether the task is regression, binary, or multiclass classification.

Loss, gradients, and backpropagation

A loss function measures prediction error. Backpropagation applies the chain rule through the computational graph to compute how each parameter affects that loss. Gradient descent updates parameters in the direction that reduces the objective. Check gradients and loss behavior on a tiny example before scaling to large data.

02

Training and generalization

Syllabus

  1. 01Optimizers and learning rates
  2. 02Regularization and normalization
  3. 03Training loop and experiment tracking

Detailed content

Optimizers and learning rates

Stochastic gradient descent updates parameters from batches of examples. Momentum uses prior updates to smooth progress; adaptive optimizers adjust step sizes per parameter. The learning rate strongly affects stability and speed. Track training and validation curves, and use a schedule when fixed-size updates stop making progress.

Regularization and normalization

Weight decay discourages overly large parameters; dropout randomly removes activations during training; data augmentation creates varied examples that preserve labels. Batch or layer normalization stabilizes activation scales. These techniques can improve generalization, but they should be chosen based on observed overfitting and training behavior.

Training loop and experiment tracking

A training loop loads batches, computes predictions and loss, calculates gradients, updates parameters, and records metrics. Separate training mode from inference mode where layers behave differently. Save checkpoints and log configuration, data version, and validation performance so interrupted runs and comparisons remain understandable.

03

Vision and sequence models

Syllabus

  1. 01Convolutional neural networks
  2. 02Sequence models and embeddings
  3. 03Attention and transformer foundations

Detailed content

Convolutional neural networks

Convolutions apply learned filters across local image regions, reusing weights and preserving spatial structure. Pooling or strided convolutions reduce resolution. CNNs are effective for visual patterns, but the receptive field, channel count, and input resolution affect compute and detail. Inspect feature maps and evaluate on representative images.

Sequence models and embeddings

Sequence tasks need representations that retain order. Embeddings map tokens or categories into learned vectors. Recurrent networks process steps sequentially and maintain a hidden state, which can make long-range learning difficult. Padding, masking, and sequence length must be handled carefully in batches.

Attention and transformer foundations

Attention computes how strongly each token should use information from other tokens. Self-attention creates contextual representations; multiple heads learn different relationships. Transformers process tokens in parallel during training but attention cost grows with sequence length. Positional information helps the model represent order.

04

Evaluation and deployment

Syllabus

  1. 01Data quality and model evaluation
  2. 02Efficient inference and deployment
  3. 03Monitoring and responsible deep learning

Detailed content

Data quality and model evaluation

Deep models can memorize accidental patterns in a dataset. Build splits that prevent near-duplicate leakage, inspect labels, and report task-appropriate metrics. Evaluate robustness across input conditions and meaningful subgroups. Compare against a simpler baseline so improvements justify added compute and complexity.

Efficient inference and deployment

Inference can be optimized with batching, smaller architectures, quantization, or compilation. Measure end-to-end latency and memory on the actual target hardware. Validate that precision changes preserve acceptable quality. Package preprocessing, weights, and runtime requirements together with an explicit model version.

Monitoring and responsible deep learning

Track input drift, output distribution, failure examples, and performance when labels become available. Deep models may learn sensitive proxies or fail unpredictably outside training conditions. Document intended use and known limitations, minimize retained data, and design human review for uncertain or consequential predictions.