Complete AI: Understand AI, ML, DL, GenAI, LLM, NLP (2025)

Set up your project environment by installing Miniconda and Visual Studio Code on Windows, then register Miniconda3 as the default Python 3.12 for seamless tool detection.

Explore data with exploratory data analysis to uncover how features like dual-SIM and Bluetooth influence price range, using counts, plots, cross tabs, and correlation insights.

Explore how cross-validation improves model evaluation, prevents overfitting, and compares training and testing strategies with scikit-learn's cross_val_score and five-fold schemes, including hyperparameter tuning.

Demonstrate hyperparameter tuning in scikit-learn, comparing manual, random search CV, and grid search CV for a KNN model, including train scores, test scores, and plotting.

Explore how to generate and interpret a classification report, including precision, recall, F1 score, support, and accuracy, and learn about cross-validation with grid search for KNN models.

Explore feature importance and how ram drives mobile price range, and learn to discard useless inputs to improve model performance via cross-validation and hyperparameter tuning.

Learn to save and load a scikit-learn model with joblib, compare Onnx, scopes.io, and pickle, and verify the loaded model yields the same predictions.

Apply one hot encoding to convert string categories into numeric features, enabling machine learning models to learn from car data like name, seller type, and transmission using a column transformer.

Use a one hot encoder with a column transformer to convert categorical features like car name, fuel type, seller type, and transmission into numeric inputs.

Explore how convolutional layers use filters to create multiple feature maps, apply nonlinear activation, and transition to pooling, flattening, and fully connected layers for image classification.

Explore rnn and other neural networks with lstm and gru, focusing on hidden states, weight sharing, and backpropagation through time, plus transformers, gnn, and autoencoders and key data practices.

Explore reinforcement learning by examining the environment, the agent, actions, states, and rewards. Learn how trial-and-error policies maximize cumulative reward across episodic and continuous tasks, with model-based and model-free approaches.

Explore how Bellman's equation underpins reinforcement learning by linking state value and action value to optimal policies, using Q-learning to balance exploration and exploitation amid deterministic and non-deterministic environments.

Explore policy and plan as they guide agent behavior and introduce the Markov decision process. Examine the Bellman equation, state transitions, and probabilistic paths to maximize rewards.

Explore how living penalty in reinforcement learning imposes a per-action negative reward, guiding the agent from wall-bumping to a direct, safe path toward maximum cumulative reward.

Explore q-learning intuition by linking states, actions, rewards, and the bellman equation, and learn how q(s,a) guides the best next move.

Discover how temporal difference updates the q-value in q-learning using reward, discount factor, and the maximum next-state q-value, reflecting the difference from the previous q-value.

Apply q-learning to a warehouse case study, guiding an autonomous robot through labeled blocks from A to L using states, actions, rewards, and q-values to find the shortest route.

Launch the Q-learning project for an autonomous warehouse robot, define 12 states and 11 actions, build the reward matrix to learn the shortest route.

Set up the lunar lander environment from gymnasium with an eight dimensional state vector and four discrete actions, and install the required packages to begin a DQN.

Build a three-layer fully connected neural network in PyTorch for the DQN project, using fc1, fc2, fc3 with ReLU activations and a forward pass from state size to action size.

The lecture implements a dqn replay memory with push and sample, selecting cuda or cpu, storing recent experiences and returning batched states, actions, rewards, next states, and dones as tensors.

Explore how the DQN agent's act function selects actions using an epsilon-greedy policy, converting state data from numpy to a tensor, evaluating the local Q-network, and balancing exploitation and exploration.

Train the dqn agent for 2000 episodes with 1000 steps per episode, using epsilon from 1.0 to 0.01, track 100-episode scores, and save the q-network when the average reaches 200.