CS 320 · Year 3 · Semester 2 · 3 credits · AI & Machine Learning

Deep Learning

Bridging classical actuarial credibility with deep representation learning, hybrid CANN architectures, and probabilistic uncertainty.

Full practice exam · 30 Q · 120 min

Season 1 · 8 episodes

The Architecture of Risk

An ambitious actuarial data scientist at a struggling commercial insurer must overhaul the company's failing predictive models with deep learning architectures before a catastrophic solvency review shuts them down.

Protagonist · Maya Lin, ASA, Senior Predictive Modeler at Meridian Specialty Mutual

Setting · Meridian Specialty Mutual's actuarial headquarters in Chicago and state insurance department review chambers

Stakes · Meridian is hemorrhaging capital from severe adverse selection in commercial fleet auto and long-tail liability; failure to modernize their predictive rating and reserving models will trigger regulatory takeover and insolvency.

Pick an episode

Episodes

Season finale: Maya presents an end-to-end, ASOP 56-compliant deep learning platform with CANN architecture, transformer sequence modeling, and conformal uncertainty bounds to the board and state insurance commissioner, securing full regulatory approval and stabilizing Meridian's capital.

From the lecturer: Welcome to CS 320. In this course, we transition from classical parametric and tree-based paradigms into differentiable computation graphs, representation learning, and deep sequence models tailored specifically for actuarial risk. To master this material, you must balance mathematical derivation with computational implementation. Keep a clean notebook beside your keyboard: for every architecture we explore, derive the backpropagation gradients and tensor dimensions by hand before writing a single line of PyTorch code. Track tensor shapes explicitly at every layer boundary—matrix dimension mismatches are the primary source of silent failures in deep models. We will systematically build from multi-layer perceptrons up to Combined Actuarial Neural Networks (CANNs), entity embeddings for millions of postal codes, and transformer-based mortality forecasting. This material underpins the modern predictive modelling expectations of SOA Exam PA, CAS PCPA, and the emerging ASOP 56 standards for algorithmic governance and model risk management. Approach every network not as an opaque black box, but as a structured non-linear extension of the generalized linear models you already know and trust.

Cold open

At 7:14 AM on a Tuesday, Maya stares at Meridian's commercial auto loss ratio chart: 118 percent. A legacy generalized linear model has systematically underpriced high-risk urban delivery fleets, burning twelve million dollars of surplus in six months.

preparing lecture…

Transcript

Derive the backpropagation equations matrix-by-matrix using the chain rule on computational graphs
Interpret the error sensitivity vector delta across hidden layers and derive weight and bias gradients
Contrast traditional Stochastic Gradient Descent (SGD) with adaptive optimisers including Momentum, RMSprop, and Adam
Map neural network loss formulations directly to standard actuarial loss objectives such as Poisson deviance and mean squared error