Goals And Interpretable Variables In Neuroscience
summary
Argues that which variables become interpretable depends on the goal: four concrete construction methods — clustering by intercorrelation (parcellations), dynamic motif clustering (EEG), target-correlated selection (MVPA) and model-driven selection like causal feature learning — lead to different, sometimes incommensurable variable sets even in large-sample limits. Uses neuroimaging examples (fMRI, DTI) and climate and education analogies to show these choices produce conflicting models, and recommends encoding goal-aware criteria and metadata rather than assuming a single privileged variable set.
AI-written summary · about
chapters (11)
- 0:00Workshop introduction
- 1:53Interpretable ML cluster goals
- 3:28The state of interpretability
- 6:46Speaker introduction
- 9:21Approaches to interpretable models
- 12:01The problem of variable definition
- 16:00Four methods of variable construction
- 19:41Loss functions and values
- 23:55Risk of scientific fracturing
- 27:35Concluding thoughts on goals
- 31:35Q&A and discussion
related talks
48mInterpretability - now what?
15mInterpretable Machine Learning with R
26mStop Explaining Black Box Models, Use Interpretable Models Instead
35mReading the Mind of an LLM
41mIf We Want AI to be Interpretable, We Need to Measure Interpretability
32mIrina Rish, Research Staff, IBM T.J. Watson Research Center
21mWhy you should care about AI interpretability
32mTheoretic analyses of internal model dynamics in the brain