While modern machine learning models often achieve superhuman performance on biohealth benchmarks, they frequently fail to generalize to new hospitals, patient populations, or biological contexts. This course investigates the theoretical and practical foundations of generalizable inference in biomedicine, focusing on the critical gap between predictive performance and mechanistic validity. We will examine how to build "world models" that leverage biological structure, enabling generalization beyond the training distribution. Key topics include: inductive biases (biologically relevant priors), causal representation learning (discovering latent state variables), hybrid models (combining mechanistic ODEs with neural networks), learning from interventional data (spanning high-throughput perturbation screens to policy learning from clinical interventions), and causal transportability. Students will engage with cutting-edge literature, dissecting success stories and analyzing "failure modes" where black-box models fall short of clinical reality.
1-3 units · Letter or Credit/No Credit
While modern machine learning models often achieve superhuman performance on biohealth benchmarks, they frequently fail to generalize to new hospitals, patient populations, or biological contexts. This course investigates the theoretical and practical foundations of generalizable inference in biomedicine, focusing on the critical gap between predictive performance and mechanistic validity. We will examine how to build "world models" that leverage biological structure, enabling generalization beyond the training distribution. Key topics include: inductive biases (biologically relevant priors), causal representation learning (discovering latent state variables), hybrid models (combining mechanistic ODEs with neural networks), learning from interventional data (spanning high-throughput perturbation screens to policy learning from clinical interventions), and causal transportability. Students will engage with cutting-edge literature, dissecting success stories and analyzing "failure modes" where black-box models fall short of clinical reality.
Offered in Spring 2027 at Stanford University.