Risk and AI (RAI) Certification Prep Course

Explore the origins of classical AI, also known as symbolic AI, from early computing devices to Turing’s ideas and the rise of conversational AI like ChatGPT and Gemini.

Contrast specific AI with general AI and AGI, highlighting domain-limited capabilities like driving, writing, or chess, and discuss whether true intelligence requires general competence.

Trace the birth of good old fashioned AI in 1956, with symbolic logic and rule-based systems, the Logic Theorist program, and the rise of game playing and Deep Blue.

Learn simple reinforcement learning with a Nim game using move weights. Update weights after each game to favor winning moves and leave a multiple of four.

Practice multi-step look ahead to set subgoals and explore future states for complex problems like Sudoku and the Tower of Hanoi, contrasting with simple trial-and-error learning.

Explore fundamental AI search techniques, including binary search, breadth-first, depth-first, and heuristic search (A*), with examples from number guessing and the traveling salesman problem to address combinatorial explosion.

Explore recursion through the Tower of Hanoi, learn base cases and breaking a problem into smaller steps, and apply to ai problem solving and tree traversals.

Master recursive adversarial tree search in AI using tic tac toe as a running example to achieve optimal play through exhaustive lookahead and base case evaluation.

Explore how complexity and combinatorial explosion affect game-playing algorithms. Learn how heuristics and reinforcement learning optimize move evaluation and self-play strategies through limited search and weight tuning.

Explore the limits of classical ai, including representation problems, frame problems, real-world uncertainty, and the brittleness of expert systems as researchers move toward neural nets.

Trace the birth of neural nets from Turing and McCulloch and Pitts to Rosenblatt’s perceptron, highlighting the shift from symbolic logic to probabilistic, learning-based neural networks.

Explore the artificial neuron as a weighted input decision maker with net input, weights, bias, and non-linear activation function, forming layers from input to hidden to output in neural networks.

Explore how deep learning evolved from LeNet-5 to AlexNet, leveraging CNNs and GPUs to beat symbolic AI on image recognition, and understand its impact on AI breakthroughs.

Deep reinforcement learning learns from scratch via self-play to outperform symbolic AI in games like Atari, Go, and chess. The lecture highlights generalized learning and move toward artificial general intelligence.

Explore why deep learning models remain inscrutable, as neural networks learn patterns through hidden layers and weights, producing accurate predictions yet lacking transparent, explainable reasoning.

Explore the ai revolution sparked by chatgpt and large language models, their human-guided training and limitations, including lack of true understanding, absence of general intelligence, and the stochastic parrot nature.

Explore the four types of machine learning—reinforcement, supervised, unsupervised, and semi-supervised—and how they learn, with practical examples and a discussion of risks in AI.

Explore principal component analysis (PCA) as a powerful unsupervised learning method for dimensionality reduction and data visualization, illustrated with village coordinates and stylometry in literature.

Explore the risks of inscrutability in deep learning, where black box models and biased training data leak bias into decisions, raising privacy concerns and driving explainable ai efforts.

Examine the risks of overreliance on AI systems, including undermining human judgment and vulnerability to adversarial examples. Learn mitigation through human oversight, testing, and governance.

Explore how machine learning trains models to recognize data patterns and make predictions for risk management tasks like credit scoring and fraud detection, and compare it with classical statistics.

Explore the types of machine learning, including unsupervised learning with clustering and PCA, supervised learning with classification and regression, semi-supervised and reinforcement learning, and parametric versus non-parametric methods, and EDA.

Explore the core steps of exploratory data analysis, including data collection, cleaning, visualization, and analysis, and understand data types, formats, and measurement scales for effective ML models.

Learn why data cleaning is essential for reliable analysis and how to fix inconsistent recording, remove unwanted and duplicate observations, address outliers, and handle missing data.

Learn how data visualization translates information into visual context using histograms, box and whisker plots, scatter plots, q-q plots, and correlation matrices to reveal distribution, outliers, and relationships.

Transform raw data into meaningful features using feature extraction, encoding, ordinal handling, and discretization. Learn one-hot encoding and dummy variable techniques and how to avoid multicollinearity.

Explore data scaling, including standardization and normalization, to bring features to a common scale, handle outliers, and improve numerical stability and interpretability for machine learning models.

Transform data to reduce skewness with log, square root, or inverse transformations. These methods alter distribution shape and can improve model performance and interpretation, unlike simple scaling.

Explore dimensionality reduction techniques like principal component analysis to compress data while preserving information, reveal yield curve patterns through PCs, and reduce multicollinearity for clearer models.

Master three-way data splitting: training, validation, and testing to train models and generalize. Avoid overfitting and underfitting, and address contamination, while exploring cross-validation and time-series splits for model selection.

Explore how programming languages and tools shape machine learning, with no single best option, and compare Python and R libraries like numpy, SciPy, pandas, scikit-learn, TensorFlow, Keras, and caret.

Explore unsupervised learning and clustering, including hierarchical, density-based, and partitional methods, and learn normalization, dendrograms, and K-means for real-world data segmentation.

Explore how the k-means algorithm clusters data into k groups by assigning points to the nearest centroid and updating centroids until convergence, using Euclidean or Manhattan distance.

Learn clustering performance measures such as inertia, the within-cluster sum of squares, between-cluster sum of squares, and the variance ratio criterion, and how centroids affect cluster quality.

Explore centroid initialization in k-means, compare multiple runs and k-means plus plus, and learn elbow, silhouette, and heuristic methods to select the right number of clusters.

Apply k-means clustering to NYSE and Nasdaq stock returns and treasury yields to identify two regimes, boom and bust, using the elbow method to set k=2.

K-means is simple, intuitive, fast, and scalable, with visualization for three features. It struggles with non spherical clusters, outliers, and high dimensionality; kernel functions or other alternatives may help.

Explore fuzzy k-means, a soft clustering method where data points have degrees of membership across multiple clusters, controlled by a fuzziness coefficient. Contrast with hard clustering's winner-takes-all approach.

Explore density-based clustering with DBSCAN and shared nearest neighbors, classifying points as core, border, or noise, and learn how threshold distance and minimum points form clusters and handle outliers.

Explore simple linear regression, a bivariate model linking y to x, with beta naught and beta one, and an unobservable error with zero mean, using ordinary least squares.

Learn multiple linear regression, modeling a target y with multiple features, and interpret coefficients as partial effects, with transformations, interaction terms, and OLS estimates for a best fit.

Explore how experience and a college degree shape hourly wages; the model uses a dummy variable and a quadratic term to show diminishing returns.

Identify common regression model misspecifications, including wrong or irrelevant features, multicollinearity, outliers, and heteroscedasticity, and learn detection and mitigation methods such as residual plots, Cook’s distance, and weighted least squares.

Explore stepwise regression as a wrapper feature-selection method that minimizes AIC to select the best feature combination, using forward and backward approaches with house price examples.

Explore classification problems in finance, distinguishing binary and multi-class tasks, and learn how logistic regression uses the sigmoid function to model probability between 0 and 1.

Explore discrete choice models with multinomial logit for multi-category outcomes, comparing choices to a baseline, and use ordered logit for ordinal data with cutoff points, noting selection bias.

Understand how linear discriminant analysis classifies data into multiple classes using class means, covariance, and priors, assigning new points to the most likely class.

Discover supervised learning techniques for classification and regression, including decision trees, K nearest neighbors, SVM, neural networks, and autoencoders, with emphasis on white-box interpretability in credit scoring and fraud detection.

Learn how regression trees predict continuous values by recursively splitting data to minimize residual sum of squares, using top-down binary splitting and interpreting leaf averages.

Explore classification trees as the predictive cousin of regression trees, using entropy and the Gini coefficient to split data and predict dividend payments from binary and continuous features.

Explore how pruning balances model complexity and accuracy in decision trees, comparing pre-pruning and post-pruning, including reduced error and cost complexity methods to improve generalization.

Explore ensemble techniques that combine multiple models to improve accuracy and reduce overfitting. Learn bagging, pasting, random forests, and boosting methods like gradient boosting and AdaBoost.

Explore k-nearest neighbors (KNN), a lazy supervised learning method that classifies or predicts by majority voting among the k closest training points using Euclidean distance, with a loan-default example.

Explore how support vector machines use support vectors to form a maximum margin classifier for high-dimensional data, with applications in text categorization, image classification, biological sequence analysis, and financial prediction.

Demonstrate how SVM separates yes and no loan decisions using income and savings, maximizing the margin with a decision function, and introducing soft margins and kernel extensions for non-linear boundaries.

Explore how artificial neural networks use feedforward architectures with backpropagation, weights, biases, and activation functions to learn complex patterns from input to output, with finance applications including risk management.

Explore how activation functions introduce non-linearity in neural networks, enabling hidden layers to learn complex patterns beyond linear transformations, with examples like sigmoid, ReLU, and tanh.

Explore a numerical example of a three-layer neural network that processes four inputs, applies weights and biases with ReLU and logistic activation to predict a binary class.

Enable backpropagation to let neural networks learn from mistakes by adjusting weights and biases after a forward pass and mean squared error loss, guided by the learning rate.

Learn how neural networks approximate functions via the universal approximation theorem and balance hidden layers, neurons, pyramid versus uniform architectures, with regularization and hyperparameters to prevent overfitting.

Overfitting occurs when a model shines on training data but stumbles on unseen data. Apply penalty-based regularization, dropout, and early stopping to reduce overfitting and improve generalization.

Explore convolutional neural networks (CNNs) that process grid-like data from images to time-series, using shared kernels over receptive fields to generate feature maps with ReLU activation and pooling.

Explore how autoencoders use an encoder and decoder to learn latent representations, reconstruct inputs through a bottleneck, and enable non-linear dimensionality reduction in high-dimensional data, unlike PCA.

Use semi-supervised learning (also called weak supervision) to leverage both labeled and unlabeled data for classification when labeling is costly, guided by clustering, smoothness, and manifold assumptions.

Explore semi-supervised learning techniques, contrasting transductive and inductive methods, with label propagation on existing data in a closed worldview versus open worldview for unseen data.

Self-training is an inductive semi-supervised wrapper method using labeled data to label unlabeled data with pseudo labels, iteratively adding high-probability predictions and noting batch labeling.

Explore co-training, a semi-supervised technique that uses two complementary views to train two models on labeled data and teach each other from predictions to improve robustness.

Unsupervised preprocessing uses unlabeled data to reveal structure and improve labeled learning through feature extraction (PCA or autoencoders), clustering with pseudo labeling, and pre-training before supervised fine-tuning.

Learn the core principles of reinforcement learning, where an agent acts in an environment to maximize long-term rewards through trial and error. Feedback guides action selection and policy development.

Explore the multi-armed bandit, showing how an agent balances exploration and exploitation among slot machines to maximize rewards, introducing policy, value function, and q-values.

Explore reinforcement learning strategies, balancing exploitation and exploration with epsilon-greedy and decay to learn the best options while still discovering potentially better rewards, leading to MDPs.

Master Markov decision processes by modeling decisions with states, actions, rewards, and the Markov property; evaluate future returns with discounting and optimize policy via state and action value functions.

Explore reinforcement learning approaches, contrasting model-based and model-free methods, then compare value-based and policy-based strategies, including q-learning, deep q-learning, policy gradients, and Bellman equations.

Explore Bellman equations in reinforcement learning to evaluate and improve policies, with value-based and action-based forms for state and action values, and learn dynamic programming, Monte Carlo, and temporal-difference methods.

Explore how the Monte Carlo method estimates state-action values from repeated episodes with random initialization, updating Q(s,a) via returns and an exponential moving average, noting finite horizon and slow convergence.

Explore the temporal difference method and Q-learning for estimating state and action values with rewards and next-state estimates. Compare with Monte Carlo, discuss curse of dimensionality and neural network approximation.

Apply ordinary least squares to derive the line in simple linear regression, predicting salary from experience by estimating the intercept and slope and minimizing residuals with RSS and MSE.

Learn how non-linear least squares extends OLS to fit non-linear relationships, minimizing the residual sum of squares using iterative gradient descent, with steps from initial parameters to stopping criteria.

Hill climbing is a simple optimization technique that starts with an initial guess and makes small adjustments to maximize an objective, though it may trap local optima and be slow.

Learn how gradient descent minimizes a loss or objective function by iteratively updating parameters with the learning rate, balancing batch, stochastic, and mini batch approaches.

Backpropagation updates network weights to reduce error with gradient descent and chain rule. Initialize weights, run a forward pass, compute error, then perform backward pass to update gradients until convergence.

Explore the computational challenges of training neural networks, including local minima and vanishing or exploding gradients. See how momentum, ReLU, batch normalization, and LSTM architectures speed learning.

Estimate the best parameter values for linear and nonlinear models, including logit and probit, via maximum likelihood, deriving probabilities, log-likelihood, and thresholded predictions.

Understand how overfitting arises when a model is too complex and memorizes training data, degrading generalization, and prevent it via simplifying the model, cross-validation, regularization, and early stopping.

Underfitting arises when a model is too simple to capture underlying data patterns, producing poor fits and biased estimates, as with a curved hedge fund size performance relationship.

Understand the bias-variance trade-off as the balance between underfitting and overfitting, and learn how model complexity affects generalization using linear, quadratic, and high-degree polynomial examples.

Explore the trade-off between prediction accuracy and interpretability in machine learning. Compare complex models with high accuracy to interpretable simple models.

Apply regularization to shrink predictor weights and reduce overfitting when many features are present. Use ridge regression and lasso with a lambda penalty and standardized data.

Learn how Lasso uses an L1 penalty on absolute coefficient values to shrink some to zero, enabling feature selection. Compare it with ridge's L2 penalty for regularization.

Elastic net blends ridge and lasso penalties to regularize regression, shrinking coefficients via l2 and selecting features via l1.

Explore cross validation and grid searches to optimize machine learning models by efficient data use, tuning hyperparameters with k-fold and leave-one-out schemes, and applying rolling window validation for time series.

Explore stratified cross-validation to preserve class distributions across folds for imbalanced datasets, improving reliable model evaluation and minority-class learning in classification tasks.

Explore bootstrapping, a sampling with replacement technique that creates training and validation sets from the original data, uses out-of-bootstrap data for validation, and compares with cross-validation.

Grid search tunes hyperparameters with cross validation to find the best lambda in lasso regression, then trains the model on full data and tests final performance.

Explore model evaluation for continuous outputs, compare metrics like mean squared error, root mean squared error, mean absolute error, and mean absolute percentage error, and understand outlier effects.

Evaluate logistic regression and a one-hidden-layer neural network for loan default prediction using training and test confusion matrices, calculating accuracy, precision, and recall, with notes on overfitting.

Explore how classification models predict discrete outcomes using thresholds, evaluate with confusion matrix metrics, and balance precision, recall, and ROC AUC for robust risk assessment.

Compare model A's linear regression and model B's tree-based regression for predicting house prices per unit area, using MSE, MAE, and MAPE to show why tree-based models often reduce errors.