Deep Learning Mastery: Build your AI Foundation

Define human intelligence as perceiving, thinking, learning, and reasoning. Describe artificial intelligence as designed entities that imitate these abilities in robots, self-driving cars, and smart assistants like Siri and ChatGPT.

Explore how artificial intelligence encompasses multiple technologies, with machine learning as its current face, and how deep learning, a subset of neural networks, powers computer vision and natural language understanding.

Compare traditional programming with machine learning, where explicit rules struggle with unseen scenarios. Learn how machine learning uses lots of data and labels to infer rules and tag new activities.

The lecture explains image classification using images and labels to learn a mapping from x to y. It explains optimizing weights through loss minimization and data representation in deep learning.

Understand data representations in binary classification by transforming features like url length and frequency of special characters to simpler rules. See how deep learning automates these data transformations.

Explore how deep learning, a subfield of machine learning using neural networks, builds tens, hundreds, or thousands of layered data representations to classify digits from 0 to 9.

Machine learning maps input data x to labels y and learns by minimizing loss through iterative weight updates, while deep learning uses neural networks and multiple layers for richer representations.

Explore how traditional machine learning maps input data to labels with weights and a bias, minimizing loss via gradient descent to predict a human face.

Explore how deep learning transforms raw image pixels into meaningful representations using a single hidden layer, one-hot encoded ten-class outputs, forward and backward propagation, and gradient descent to minimize loss.

Learn how weights and parameters drive deep learning models, from dense neural networks with a hidden layer and output layer to matrix-based weight calculations and activation functions.

Explore activation functions and their role in turning linear logits into probabilities, from sigmoid in binary classification to ReLU in deep networks, and softmax for multi-class predictions.

Learn how loss and cost functions measure prediction errors, including mean squared error for regression and cross-entropy variants for binary and multi-class classification, with one-hot and integer labels.

Explore gradient descent and its optimizers, from batch to mini-batch and stochastic, learn how the learning rate alpha shapes cost-minimizing weight updates in deep networks.

Access and set up course materials by downloading the GitHub repo, reading readme, and using Google Colab with GPU or a local environment configured via requirements.txt (TensorFlow and Keras).

Explore TensorFlow, a Python-based open-source platform for tensor manipulation and gradients across CPU, GPU, and TPU, and learn how Keras provides a high-level API for rapid deep learning model building.

Master TensorFlow basics by exploring tensors and NumPy arrays, the low-level math behind deep learning, and how Keras offers a quick high-level API to build models with gradient tape.

Master Keras basics by building and training models with sequential and functional interfaces, using fashion-mnist, weighted layers, cross-entropy loss, and practical callbacks for robust training.

Master convolutional neural networks by learning how convolutional layers use filters to extract local and hierarchical image features, with padding, stride, pooling, and flattening before dense layers.

Build an image classifier with convolutional neural networks on the CIFAR-10 color image dataset using the Keras functional API, training, evaluating on test data, and producing predictions.

Explore data augmentation techniques in deep learning using convolutional neural networks to improve generalization on CIFAR data, including random flips, rotations, zooms, and translations.

Learn how transfer learning uses a VGG16 convolutional base to extract features from CIFAR images, then train a small top classifier for improved accuracy over training from scratch.

Master transfer learning by freezing the pre-trained VGG16 convolutional base, then train a new head on CIFAR, balancing frozen and trainable parameters for end-to-end fine-tuning.

Master transfer learning by freezing the earlier layers of the VGG16 convolutional base and fine-tuning only the last four layers, using a CIFAR_VGG16_select_layers model for end-to-end training.

Explore transfer learning by fine tuning all convolutional base layers of the VGG16 architecture, train end to end with our own dense head, and compare freezing versus unfreezing strategies.

Explore vanishing gradients and remedies like residual connections and batch normalization, then learn depthwise separable convolutions and conv 2D transpose layers for image segmentation and other vision tasks.

Builds a mini-xception like architecture for image classification on CIFAR dataset, using residual connections, batch normalization, and depthwise separable convolutions to achieve 77% test accuracy with about 50,000 parameters.

Explore image segmentation with the Oxford Pets dataset by building an encoder-decoder model using conv2d and conv2d transpose. Train on a 6390/1000 split to achieve about 86% validation accuracy.

Explore how to handle sequence data with recurrent neural networks, time series, and text. Compare RNN variants such as LSTM and GRU, and cover sequence-to-sequence, encoder-decoder, and transformers.

Explore how LSTM and GRU address long sequence memory, overcoming vanishing and exploding gradients, with techniques like layer normalization and gradient clipping for robust deep learning.

Build a deep learning time series forecast for Seattle bicycle counts using past 30 days to predict the next day, exploring dense, RNN, LSTM, and multivariate models in TensorFlow.

Explore the evolution of NLP from rule-based systems to deep learning, covering bag-of-words, RNNs, LSTMs, GRUs, transformers, and word embeddings for text representation.

Master sentiment analysis on IMDb reviews by building a bag-of-words model with multi-hot vectors, exploring unigram and bigram n-grams, and tf-idf using Keras text vectorization in a deep learning workflow.

Explore sentiment analysis on imdb reviews using sequence models with bidirectional lstm, compare bag of words, and introduce word embeddings and transformer architecture for deeper nlp mastery.

Explore the transformer encoder architecture, including self-attention and multi-head attention, learn positional embeddings, and see how context-rich representations empower modern NLP models and sentiment analysis.

Learn to build a transformer encoder for IMDb sentiment analysis, comparing fixed vs learnable positional embeddings and applying multi-head attention to classify reviews.

Explore sequence-to-sequence encoder-decoder RNN for machine translation, converting an input sentence into a context vector and decoding the target sequence token by token with past outputs.

Explore how transformer encoder and decoder blocks translate English to Spanish, compare with RNNs, and learn self-attention, multi-head attention, cross attention, and context-aware token representations.

Build an encoder-decoder translation model using GRU to translate English to Spanish. Learn data loading, vectorization, start/end tokens, and a train-validation-test workflow in TensorFlow and Keras.

Build a transformer-based English-to-Spanish translator with an encoder-decoder architecture and self- and cross-attention. Use learnable positional embeddings with causal masking and train on vectorized bilingual data.