Neural Networks and Deep Learning
Deep learning practice in industry has shifted from training models from scratch toward adapting, compressing, and serving pretrained foundation models as components that agentic systems call, with practitioners accountable for evaluating machine-generated training code and the behavior of the resulting models. The course therefore keeps its mathematical and architectural core but adds parameter-efficient adaptation, reproducible evaluation harnesses, provenance and drift monitoring, and packaging a trained model as a callable, documented tool.
Current description → proposed description
This course offers an in-depth study of neural network architectures and deep learning techniques that power today’s most advanced Artificial Intelligence systems. This course focuses specifically on the mathematical foundations, structural design, and training dynamics of neural networks. Topics include perceptrons, backpropagation, activation functions, optimization algorithms, regularization methods, and loss functions. Students will analyze and implement advanced architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, Transformers, autoencoders, and generative models. The course also covers techniques for improving training performance, reducing overfitting, and interpreting model outputs. Through hands-on coding assignments, students will build and fine-tune deep models on real-world datasets. Emphasis is placed on critical model evaluation, explainability, and preparation for deployment in research and production contexts.
This course offers an in-depth study of neural network architectures and deep learning techniques that power today’s most advanced Artificial Intelligence systems. This course focuses specifically on the mathematical foundations, structural design, and training dynamics of neural networks. Topics include perceptrons, backpropagation, activation functions, optimization algorithms, regularization methods, and loss functions. Students will analyze and implement advanced architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, Transformers, autoencoders, and generative models. Building on these architectures, students will adapt pretrained foundation models through parameter-efficient fine-tuning, distillation, and quantization, weighing these methods against prompt-based and few-shot adaptation under realistic accuracy, latency, and cost budgets. The course also covers techniques for improving training performance, reducing overfitting, and interpreting model outputs, along with automated evaluation harnesses, model and data provenance, and post-deployment monitoring for drift. Through hands-on coding assignments, students will source, clean, augment, and leakage-check real-world datasets and will build and fine-tune deep models on them, directing AI coding agents to generate training code while remaining accountable for verifying every result those agents produce. Students will also package a trained model as a callable, documented service registered as a tool that a domain-specific agent invokes under guardrails. Emphasis is placed on critical model evaluation, explainability, and preparation for deployment in research and production contexts.
What changes
- Pretrained foundation-model adaptation: PEFT, distillation, quantization
- Prompt-based and few-shot adaptation compared against fine-tuning
- Reproducible evaluation harnesses, model and data provenance, drift monitoring
- Trained model registered as a guarded tool a domain-specific agent invokes
- Student accountability for verifying AI-agent-generated training code
6 proposed outcomes, mapped to 10 program outcomes
Each outcome below is written to be observable and assessable, and each is mapped to the program learning outcomes for which it produces evidence.
Students will be able to derive the mathematical mechanics of deep networks - backpropagation through feedforward, convolutional, recurrent, and attention layers, gradient-based optimization, activation and loss function selection, and regularization - showing how each choice governs training dynamics, convergence, and generalization.
Deriving backpropagation, optimization, activation and loss selection, and regularization is direct application of the neural-network and probabilistic mathematics named in PLO 3.1, which the workbook assesses at exit in this course.
Students will be able to design a deep architecture appropriate to a stated task - convolutional, recurrent/LSTM, Transformer, autoencoder, or generative - including constructing and curating the training data pipeline that feeds it, covering sourcing from APIs and databases, cleaning, augmentation, and leakage-checked splitting.
Selecting and building a learning-based model for a stated task is the intelligent-system development in PLO 4.1 and applies the neural-network theory in PLO 3.1, while the explicit requirement to source from APIs and databases, clean, augment, and split without leakage is the data acquisition work in PLO 2.1.
Students will be able to refine supervised and unsupervised deep models iteratively on the evidence of a reproducible evaluation harness - including verifying agent-generated training and experiment code against that harness before accepting its results - using ablation studies, learning-curve and error analysis, calibration, robustness and subgroup testing, and post-deployment drift monitoring.
Ablation, error analysis, calibration, robustness testing, drift monitoring, and the required verification of agent-generated training code against the harness before its results are accepted are all the diagnostic metrics and iterative refinement specified in PLO 6.2, and the refinement is exercised on the student’s own supervised and unsupervised deep models as PLO 4.1 requires.
Students will be able to adapt pretrained transformer-based foundation models through parameter-efficient fine-tuning, distillation, and quantization, comparing each against prompt-based and few-shot adaptation of the same model on the same task under enterprise and embedded budgets for latency, memory, and cost.
Fitting transformer foundation models to enterprise and embedded deployment budgets is the industrial LLM utilization in PLO 6.1, and the required head-to-head comparison against prompt-based and few-shot adaptation obliges students to engineer and assess prompts for alignment and task fit as in PLO 4.2.
Students will be able to construct an embedding and inference service that exposes a trained model as a version-pinned, callable tool - with input validation, schema-constrained outputs, access control, and logging of every invocation - and register that service with a domain-specific AI agent, demonstrating a guarded end-to-end invocation in which the agent calls the model and consumes its result.
Registering the service with a domain-specific agent and demonstrating a guarded end-to-end invocation, with input validation, schema-constrained outputs and access control holding at the boundary, is the secure and flexible domain-specific agent integration in PLO 2.2 at the introductory depth the workbook assigns this course, and delivering the model as an embedding and inference API is the structured LLM deployment through APIs and embeddings in PLO 6.1.
Students will be able to evaluate the bias, privacy, explainability, provenance, and economic-access consequences of a deep model destined for deployment, presenting the assessment as a model card written for both technical reviewers and non-technical stakeholders.
Assessing bias, privacy, explainability, provenance, and economic access is the ethical and legal integrity outcome that the workbook states twice, word for word, under Christian Faith (PLO 1.1) and under Integrated Disciplinary Knowledge (PLO 3.2), so a CLO that carries one necessarily carries both; delivering that assessment as a model card addressed to both technical reviewers and non-technical stakeholders is the audience-appropriate communication of results and ethical considerations in PLO 5.1.
Program outcomes this course reaches
Filled cells are program learning outcomes with at least one supporting course learning outcome in this course. Sparse coverage is expected — no single course carries all twelve.
The workbook maps this course too
The proposal workbook's assessment map already assigns program learning outcomes to this course. The outcomes above were written to cover it.
| Program outcome | Workbook level | In this draft |
|---|---|---|
| PLO 1.1 Ethical and Legal Integrity in AI | I | covered |
| PLO 2.1 Data Acquisition for Real Problems | I | covered |
| PLO 2.2 Domain-Specific AI Agents | I | covered |
| PLO 3.1 AI Theory and Mathematics | D,AE | covered |
| PLO 3.2 Ethical and Legal Integrity in AI | I | covered |
| PLO 4.1 Learning-Based Systems and Agents | D,AE | covered |
| PLO 4.2 Prompt Engineering for LLMs | I | covered |
| PLO 5.1 Communicating AI to Any Audience | D | covered |
| PLO 6.1 LLMs in Industrial Settings | I | covered |
| PLO 6.2 Model Evaluation and Refinement | D,AE | covered |
I = Introduced · D = Developed · AE = Assessed at Exit