GEN AI: Introduction to Deep Learning and Generative AI

Topics covered

• Introduction to Deep Learning

  • Difference between Machine Learning and Deep Learning.
  • Basic concept of Neural networks.
  • Real-world applications of Deep Learning.

• Artificial neurons

  • Artificial neuron inspired by biological neuron.
  • Weights, biases and weighted sum.
  • Mathematical representation of Neuron.

• Neural network architecture

  • Input layer, hidden layers and output layer.
  • Forward propagation.
  • Parameters and model architecture.

• Activation functions

  • Sigmoid, Tanh and ReLU.
  • Need for activation functions.
  • Role of non-linearity.

• Loss metrics

  • Purpose of loss function.
  • Mean Squared Error (MSE).
  • Introduction of Cross-Entropy loss.

Topics covered

• TensorFlow and Keras

  • Introduction of Deep Learning framework.
  • Tensor, layers and models.
  • Making Sequential Model.

• Building neural networks

  • Defining input and output layers.
  • Adding hidden layers.
  • Compile and train the model.

• Training process

  • Epochs and batch size.
  • Training and validation data.
  • Monitoring loss and accuracy.

• Optimization methods

  • Practical idea of gradient descent.
  • Learning rate.
  • Introduction of optimizers.

• Experimentation

  • Comparing performance by changing activation functions.
  • Model architecture and hyperparameters experiment.

Topics covered

• Introduction to Computer Vision

  • Representation of image data.
  • Image classification problem.
  • Traditional image processing and CNN-based learning.

• Convolutional Neural Networks

  • CNN architecture.
  • Convolution operation.
  • Filters and kernels.
  • Feature maps.

• Convolution parameters

  • Stride.
  • Padding.
  • Output feature-map dimensions.

• Pooling layers

  • Max pooling.
  • Average pooling.
  • To reduce spatial dimensions.

• CNN architecture

  • Convolution, activation and pooling layers.
  • Flattening and fully connected layers.
  • Hierarchical feature extraction.

Topics covered

• Image classification

  • Datasets like MNIST and CIFAR-10.
  • Image preprocessing.
  • Labels and class representation.

• CNN implementation

  • Creating convolutional layers.
  • Adding pooling and dense layers.
  • Use of Softmax output layer.

• Training

  • Training-validation split.
  • Monitoring loss and accuracy.
  • Tuning epochs and batch size.

• Evaluation

  • Model evaluation on test data.
  • Inspecting misclassified images.
  • Comparing model performance.

Topics covered

• Sequential data

  • Sequence and time-dependent data.
  • Text, speech and time-series examples.
  • Limitations of feedforward networks for sequences.

• Recurrent Neural Networks (RNNs)

  • Recurrent connections.
  • Hidden state.
  • Sequence processing.
  • Unrolling an RNN through time.

• Training RNNs

  • Backpropagation Through Time (BPTT).
  • Introduction of vanishing and exploding gradients.

• Long Short-Term Memory (LSTM)

  • LSTM architecture.
  • Forget gate, input gate and output gate.
  • Cell state and hidden state.

• RNN vs LSTM

  • Long-term dependencies.
  • Memory mechanism.
  • Typical use cases.

Topics covered

• Sentiment analysis

  • Text preprocessing.
  • Tokenization and sequence representation.
  • Positive and negative sentiment classification.

• LSTM implementation

  • Use of embedding layer.
  • To configure LSTM layers.
  • Generating classification output.

• Text generation

  • Concept of next-character or next-word prediction.
  • Simple RNN/LSTM language modeling.
  • Evaluation of generated text.

• Time-series prediction

  • Giving historical values as sequence input.
  • Time-window creation.
  • Predicting future values.

• Model evaluation

  • Training and validation performance.
  • Basic idea of overfitting.
  • Comparing predictions and actual values.

Topics covered

• Introduction to Generative AI

  • Difference between discriminative and generative models.
  • Concept of data generation.
  • Image generation applications.

• Autoencoders

  • Encoder and decoder architecture.
  • Latent representation.
  • Reconstruction loss.

• Variational Autoencoders (VAEs)

  • Probabilistic latent space.
  • Mean and variance representation.
  • Sampling from latent distribution.
  • Reconstruction loss and KL divergence.

• Generative Adversarial Networks (GANs)

  • Generator and discriminator.
  • Adversarial training.
  • Discrimination between real and generated data.

• Model comparison

  • Autoencoder and VAE.
  • VAE and GAN.
  • Generation quality and training challenges.

Topics covered

• Introduction to diffusion models

  • Diffusion-based approach to generative modeling.
  • Relationship between noise and data.

• Forward diffusion process

  • Gradually adding noise to the original image.
  • Noise schedule.
  • Generation of noisy samples.

• Reverse diffusion process

  • Noise predicting.
  • Denoising steps.
  • Image reconstruction from noise.

• Diffusion probabilistic models

  • Basic idea of probabilistic modeling.
  • Training objective.
  • Sampling process.

• Training and inference

  • Use of noisy images during training.
  • Iterative denoising during inference.
  • Sampling steps and generation quality.

Topics covered

• GAN implementation and experimentation

  • Implementing generator and discriminator.
  • Basic workflow of training loop.
  • Inspecting generated image samples.

• Pretrained diffusion models

  • Use of pretrained models.
  • Image generation workflow.
  • Practical introduction to prompt-based image generation.

• Image generation

  • Image generation from latent noise.
  • To compare generated outputs.
  • Experimenting with different generation settings.

• Evaluation

  • Visual quality inspection.
  • Basic idea of diversity and consistency.
  • Identifying common generation artifacts.

Topics covered

• Introduction to LLMs

  • Concept of language modeling.
  • Next-token prediction.
  • Applications of large language models.

• Text representation

  • Word embeddings.
  • Tokenization.
  • Vocabulary and token IDs.
  • Introduction to contextual representation.

• Attention mechanism

  • Need of attention.
  • Query, Key and Value.
  • Scaled dot-product attention.
  • Self-attention.

• Transformer architecture

  • Transformer blocks.
  • Multi-head attention.
  • Positional encoding.
  • Feedforward layers.
  • Basic idea of residual connections and normalization.

• Transformer workflow

  • Input tokens to contextual representations.
  • Attention-based information exchange.
  • Output predictions.

Topics covered

• Transformer model families

  • Encoder-only models.
  • Decoder-only models.
  • Encoder-decoder models.

• BERT-like models

  • Bidirectional context.
  • Masked language modeling.
  • Text classification and other NLP tasks.

• Decoder-based text generation

  • Autoregressive generation.
  • Causal attention.
  • Next-token prediction.

• Machine translation

  • Source and target sequences.
  • Encoder-decoder workflow.
  • Sequence-to-sequence generation.

• Fine-tuning LLMs

  • Difference between pretraining and fine-tuning.
  • Parameter-efficient fine-tuning (PEFT).
  • LoRA (Low-Rank Adaptation).
  • Comparison of full fine-tuning and parameter-efficient approaches.

Topics covered

• Prompting techniques

  • Zero-shot prompting.
  • Few-shot prompting.
  • Instruction-based prompting.
  • Prompt structure and clarity.

• Prompt experimentation

  • Comparing outputs of different prompts.
  • Task-specific instructions.
  • Output formatting and constraints.

• Prompt fine-tuning

  • Basic idea of prompt tuning.
  • Trainable prompt representations.
  • Compared to traditional fine-tuning.

• Other adaptation methods

  • Overview of fine-tuning approaches.
  • Practical context of PEFT-based methods.
  • Trade-offs of model adaptation.

• Evaluation

  • Assessing the quality of generated responses.
  • Consistency and task relevance.
  • Basic awareness of hallucinations and limitations.