Data Science & AI Masters 2026 - From Python To Gen AI

Install python using python.org, anaconda, or colab, then master python basics by exploring variables and keywords, data types, and simple operations with practical examples.

Master numpy, a Python library for fast numerical operations on multi-dimensional arrays. Install via pip, import as np, and use zeros, ones, full, and identity.

Explore descriptive and inferential statistics, and distinguish data types—categorical (qualitative) and numerical (quantitative)—to build foundational data science skills for gen ai.

Define population as the full group you want to study and sample as a subset. Use a country’s average age to illustrate why we analyze samples before drawing population conclusions.

Master the addition rule of probability by applying mutually exclusive and not mutually exclusive formulas, using die rolls and class examples to compute P(A or B).

Explore cumulative probability, the likelihood that a random variable lies within a range a to b, with examples from fair two-coin flips and dice sums.

Explore Bayes theorem as an extension of conditional probability, learn the formulas for P(A|B) and P(B|A), and see how probability of A intersect B underpins this approach.

Explore the Poisson distribution, a discrete probability model for the number of events in a time period, with lambda as the average rate.

Explore the normal distribution (Gaussian bell curve), its relation to the central limit theorem, the empirical rule, and the 68/95/99.7 percent ranges with mu and sigma.

Master hypothesis testing to make statistical decisions from experimental data, using null and alternate hypotheses. Learn how sample data assess hypotheses with a standard normal distribution and tailed tests.

Explore practical techniques to handle missing values in a churn modeling dataset using Python, pandas, and imputation methods like mean, median, mode, and forward or backward fill.

Explore outlier treatment in data cleaning, learn detection methods—box plots, histograms, scatter plots, and normal distribution rules—and decide when to remove, replace, or keep anomalies in time series forecasting.

Identify and clean invalid values in data by correcting formats, out-of-range values, and data types, and apply proper Unicode encoding when reading data with pandas to ensure accurate analysis.

Explore the two main data types, qualitative (categorical) and quantitative (numerical), with subtypes nominal vs ordinal and discrete vs continuous, plus practical examples.

Explore multivariate analysis to gain deeper insights by examining more than two variables, including combinations of categorical and numerical attributes within the exploratory data analysis framework.

Learn practical data cleaning in pandas by converting invalid types to numeric and addressing null values, including drop or impute decisions. Apply feature binning to tenure for clearer insights.

Learn to build an end-to-end EDA report for churn analysis, covering business understanding, data assessment, missing data handling, graph-driven insights, and final recommendations.

Master data query language dql basics, including select queries and creating tables. Learn to filter results with where, like, in, and to use aliases with as.

Explore SQL aggregate functions and how group by, count, min, max, and average create data insights from examples like telco churn, gender, and contract.

Learn how stored procedures in MySQL automate repetitive queries by saving code, accepting input parameters, and returning results with in and out parameters, including top players by goals.

Explore feature encoding to convert categorical x variables into numerical inputs for predictive models. Learn label encoding, one hot encoding, and dummy encoding, including their use and practical examples.

Learn the basics of regression as a supervised learning technique that predicts a continuous dependent variable from independent variables, covering simple and multiple linear regression, lasso, ridge, and polynomial extensions.

Practically apply regression metrics on a 50 startups dataset, computing mae, mse, rmse, r2, and adjusted r2 with sklearn, including train-test split and feature scaling.

learn how simple linear regression models the relationship between area and price with a linear equation y = mx + c, and uses euclidean distance to fit the best line.

Experiment with polynomial regression to fit non linear data, compare linear vs polynomial models across degrees two, three, five, ten, and beyond, using train-test splits, R2 scores, and visualizations.

Learn how log loss evaluates prediction probabilities in binary classification, with intuition and a formula, and compare it to metrics like confusion matrix and recall using fraud and spam examples.

Learn how the area under the ROC curve (AUC ROC) evaluates binary classifiers across thresholds, by plotting true positive rate versus false positive rate and relating to sensitivity and specificity.

Learn to build a k-nearest neighbors classification model in Google Colab, performing data cleaning, one-hot encoding with dummy features, train-test split, and standardization for scalable predictions.

Visualize decision tree classifiers and compare gini versus entropy to understand root, internal, and leaf nodes in churn predictions. Use plot_tree and graphviz to simplify trees, explore max_depth, and accuracy.

Bagging combines bootstrapped training samples of diverse base models with aggregation voting to reduce bias and variance, with random forest and extra trees as key examples.

Compare bagging and random forest ensembles, and observe how max features influence splits in each model. Visualize the trees using a generated classification dataset.

Explore ensemble methods for churn prediction, comparing random forest, AdaBoost, gradient boosting, and XGBoost using encoding, scaling, and train-test splits, with accuracy as the focus and hyperparameter optimization ahead.

Explore hierarchical clustering, an agglomerative method that builds a dendrogram illustrating the hierarchy of clusters from single data points to a global cluster, using euclidean distance.

Implement the practical PCA workflow in Python, including data cleaning, dummy encoding, and a train-test split, then apply PCA to reduce 30 features to principal components. Compare PCA with LDA, kernel PCA, and QDA, analyze explained variance, and evaluate logistic regression accuracy to understand PCA's impact.

Explore manual hyperparameter optimization on a breast cancer classification task with a random forest, including data prep, feature scaling, train/test split, and accuracy evaluation, with grid and randomized search upcoming.

Utilize randomized search CV to tune a random forest's hyperparameters—n_estimators, max_depth, min_samples_split, bootstrap—by sampling 100 iterations from 15,400 combinations and evaluating with threefold cross-validation to improve accuracy.

Pre-process and clean time series data by handling missing values, removing duplicates, and addressing outliers, then apply feature scaling, encoding, and feature engineering for forecasting.

Learn practical techniques for handling missing values in a churn modeling dataset with pandas, including dropping data, forward/backward fill, and imputing with mean, median, or mode.

Explore the mathematics behind the Maa model and how the moving average component uses past error terms to forecast.

Assess time series stationarity with the augmented Dickey-Fuller test, apply transformations for non-stationary data, and forecast with ARIMA models.

Explore end-to-end time series transformations and their inverses, focusing on log, double log, and log differencing, with numpy exp and cumsum techniques for restoration and testing.

Explore Facebook Prophet for time series forecasting, featuring fast additive regression with yearly, weekly, and daily seasonality, holiday effects, and robust handling of missing data and outliers.

Learn how to incorporate holiday effects into the Facebook Prophet model by adding a holidays data frame or built-in country holidays, then fit, forecast, and compare with earlier models.

Explore mean squared error, the metric for the average of squared differences between actual and predicted values using the MSE formula. Note that MSE, MA, and RMSE are distinct metrics.

Forecast stock prices using univariate time series with Facebook Prophet, applying Bayesian regression to historical data and volumes to generate predictions that inform trading strategies.

Explore the ReLU activation function and its role in mitigating vanishing and exploding gradients in neural networks, contrasting it with sigmoid and tanh, and leaky ReLU to prevent dying ReLU.

Understand forward pass and backward propagation in neural networks, illustrated with a churn prediction example. See how weights update through batch size, iterations, epochs, and learning rate using gradient descent.

Explore gradient descent and stochastic gradient descent, linking forward and backward passes to cost function decay, weight updates, and the journey from local to global minima in neural networks.

Explore how artificial neural networks assemble input, hidden, and output layers to model data; learn to set feature counts and hidden units, and compare with traditional methods on churn analysis.

Apply churn modeling on a real dataset by comparing traditional machine learning models with an artificial neural network, including one-hot encoding and feature scaling after train/test split.

Explore how epoch, batch size, and iterations govern neural network training, guided by gradient descent, forward and backward passes, and data diversity.

Build a convolutional neural network for x-ray image classification to detect pneumonia. Prepare grayscale 100x100 images, train with TensorFlow, and save the model.

Explore vanishing and exploding gradients in neural networks, and learn remedies like ReLU, RMSprop, LSTMs, and truncated backpropagation to stabilize learning.

Discover how long short-term memory networks overcome vanishing and exploding gradients using input, forget, and output gates with an internal state, enabling short- and long-term learning for sequences.

Explore pre-trained models, with a focus on Vgg16's 16 learnable layers, transfer learning, and its 1000-class image classification architecture.

Master natural language processing foundations and its role in generative AI, from text to numbers and embeddings, enabling computers to understand and generate language for chatbots and translation.

Explore the key NLP challenges, including pragmatic, lexical, syntactic, and anaphoric ambiguity, along with issues of standardization, ethical considerations, context understanding, and data sparsity.

Remove special characters to clean text data, reduce noise, and improve tokenization in NLP preprocessing, using regex, spaCy, and NLTK.

Learn to handle contractions in nlp. Expand them via the contractions library or regex and apply contraction handling during preprocessing with stopword and special character removal.

Learn how stop words affect NLP and how to remove them to boost text analysis. The lecture covers English stopwords and a Python example using NLTK for tokenization and preprocessing.

Learn how textual data is converted into numerical vectors for machine learning through vectorization. Explore one-hot encoding, bag of words, tf-idf, and word embeddings like Word2Vec, GloVe, and FastText.

Explore pre-trained word2vec embeddings from Google News vectors using gensim, examining 300-dimensional word vectors and concepts like cosine similarity, word similarity, and king minus man plus woman analogy.

Compare glove and fasttext word embeddings in a practical NLP session, loading pre-trained models with gensim and evaluating semantic similarity using the word similarity 353 dataset.

Apply LSTM to a language task by building a next word predictor from a text corpus, converting text to numeric sequences with a tokenizer and padding, framed as supervised classification.

Train a TensorFlow sequential LSTM model with embedding and tokenization, using pad sequences to map 119 samples of 33 features to 88 outputs, achieving about 97% training accuracy.

Discover gated recurrent unit architecture, compare GRU with LSTM, and learn when GRU or LSTM performs better, with emphasis on training time and fewer parameters.

Explore Bert configurations across base, large, tiny, mini, small, and medium models, detailing encoder counts, attention heads, and hidden units to guide model selection.

DistilBert reduces Bert to 40% of its size and up to 60% faster, preserving 97% of its capabilities. It uses MLM pretraining and three losses, trained on 16 GB data.

Albert, a light Bert variant, reduces parameters through cross parameter sharing and embedding factorization, introduces sentence order prediction, and enables faster training with shared encoder weights.

Explore decoder-only GPT as a generative pre-trained transformer, highlighting embedding, masked multi-head attention blocks, and the prompt-driven input that yields the next token output.

Learn how self-attention drives transformers by computing attention weights from queries, keys, and values via softmax, then apply masked multi-head attention with concatenated heads in the decoder.

Explore rag architectures for turning PDFs into text with large language models, chunking data, creating embeddings, and storing them in a vector database for context-based answers.

Explore LangChain, a framework to build apps powered by language models, connecting LLMs to data sources, embeddings, and vector stores for retrieval augmented generation.

Master prompt engineering for non-technical learners to interact with large language models like ChatGPT, using tokens, vectors, and memory across four components—instruction, context, input data, and output indicator.