Artificial Intelligence Course for Engineers & STEM 2026

Define a linear model that relates stress to strain via Hooke's law, estimate its slope and intercept as weights, and assess fit using residuals and a loss function.

Derive the closed-form solution for a simple linear model by minimizing quadratic loss with partial derivatives, obtaining the best-fit slope and intercept, then introduce gradient descent for larger models.

Explore model generalization, distinguishing interpolation within the training data from extrapolation beyond it, using a steel specimen example to show how predictions outside the elastic range fail.

Assess generalization with held-out data by training on a subset and testing on unseen points, avoid memorization, and apply Occam's razor to prefer simpler models.

Master google colab notebooks with markdown and code cells to visualize data, analyze residuals, perform gradient descent on mse loss, and explore ai-assisted edits with gemini.

Generate 400 synthetic P and L data with 2% Gaussian noise to predict gamma; split 70/30 for training/testing, scale inputs to order one, and use symbolic regression to recover equation.

Explore expression trees in tree-based genetic programming for symbolic regression, where operators form internal nodes and terminals become variables or constants to fit data.

Explore how genetic programming uses an evolution loop to generate random expression trees, evaluate fitness, select top performers, and apply crossover and mutation toward lower loss.

Discover genetic operations in symbolic regression, including crossover and mutation, to generate offspring by swapping subtrees and adding random variation, driving populations toward lower loss through natural selection.

Explore fitness and parsimony pressure in genetic programming, using mean squared error plus a node-based penalty to curb bloat and favor simpler, Occam's Razor–aligned models.

Set a population of 450 equations and evolve them over 30 generations with a validation dataset, using a five-item tournament, crossover, parsimony 2e-4, and linear scaling.

Explore neural networks in artificial intelligence, deep learning, using crack detection to contrast regression approaches and train with mini-batch gradient descent, evaluating with accuracy and recall in fnn and cnn.

Explore how neural networks mimic biology with layered structures, weighted sums, and nonlinear activations. Training adjusts weights to minimize loss, from perceptrons to backpropagation and the AlexNet breakthrough.

Explore how an artificial neuron computes a weighted sum of inputs, with learned weights and bias, and applies a nonlinear activation to produce a network output.

The sigmoid activation maps z to zero to one via 1/(1+e^-z) and faces vanishing gradient, while relu uses max(0,z) to introduce non-linearity and ease computation.

Explore the architecture of fully connected feed-forward neural networks, detailing input layers, hidden layers with ReLU activations, and all-to-all connections that enable the universal approximator.

Compare regression and classification: predict continuous outputs like stress or delta in regression, versus discrete class labels in neural networks for crack detection and vehicle types.

Explore mini-batch gradient descent using data subsets and stochastic batch selection to update neural network weights. Learn with Adam, adaptive per-parameter learning rates and epochs versus iterations.

Explore a feedforward neural network (FNN) for concrete surface image classification. Flattening 96x96x3 images into 28,000 features discards spatial structure, yielding 600k parameters across two hidden layers and motivates CNNs.

Explore how convolutional neural networks exploit local spatial relationships with learnable kernels sliding over images to produce feature maps, enabling translational equivariant features for crack detection with fewer parameters.

Global average pooling collapses each feature map channel into a single scalar by averaging spatial positions, with no learnable weights, reducing parameters before the classifier and avoiding non-linearity.

Explore cnn architectures that use weight sharing with three by three by three kernels on rgb images, featuring stacked conv blocks, max pooling, global average pooling, and logit-based training.

Explore how learned filters form a feature hierarchy in CNNs, from low-level edges to abstract shapes, and apply transfer learning by freezing early layers and retraining final layers.

Explore how accuracy alone can mislead with class imbalance and the accuracy paradox, and learn to use the confusion matrix, precision, and recall to evaluate models.

Explore how overfitting occurs when models memorize training noise and fail to generalize in deep learning, and learn regularization methods such as data augmentation, dropout, early stopping, and weight decay.

Implement a four-block convolutional neural network in PyTorch, using 3x3 filters, ReLU, max pooling, global average pooling to 224 channels, and a final linear layer trained with BCE with logits.

Train models on data, define architecture and loss, and compare them on the test set using a confusion matrix and accuracy. Results show CNN outperforms FNN.

Engineers apply neural networks, including feedforward networks, convolutional neural networks, and deep operator networks, to predict wave pressure distributions on hyperparaboloid surfaces, using regression on a 30-by-30 pressure grid.

Explore the big picture, learn about additional model families you should know, practice responsible AI coding assistance, and prepare for a capstone project to apply everything end to end.

Understand how loss functions define error, with MSE for regression and cross entropy for classification, and how gradient descent drives training. Use validation and regularization to improve generalization.

Explore a core set of machine learning methods for tabular data, including logistic regression, tree ensembles, SVMs, and PCA-based dimensionality reduction, plus clustering and anomaly detection.