Primo: Practical Learning-Augmented Systems with Interpretable Models
summary
Primo presents two interpretable models PrAM and PrDT, a Distill Engine, and Bayes Optimization to auto-tune pruning and hyperparameters; it adds Monotonic Constraint and Counterfactual Explanation for debugging. A 31-feature MLP is replaced by a decision tree with seven leaves using only three features, cutting inference work to simple if tests and showing up to 7.9x reduction in tail IO latency in a storage case study. Claims include lower training/inference cost, improved robustness, and actionable model adjustments.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:34Machine Learning in Systems
- 1:04Challenges in Practical Deployment
- 1:29High Training and Tuning Cost
- 2:13Strict Inference Overhead Requirement
- 4:134 Opaque Decision Process
- 5:06Hard to Adjust
- 6:25Existing Solution
- 8:38Why Interpretable Models Work
- 10:11Primo Design
- 11:26Interpretable Models Training
- 12:54Post-Processing Optimization
- 15:56LinnOS: Performance Analysis
- 16:50LinnOS: Effectiveness Analysis
- 18:26Summary
related talks
01h 10mModels and Programs: Better Together
26mStop Explaining Black Box Models, Use Interpretable Models Instead
32mTaming the Machine: Basics of ML Models Training and Inference Optimization
42mReasonable AI
11mCollaborative Natural Language Inference
15mInterpretable Machine Learning with R
01h 30mToward Understanding In-context Learning
01h 38mTraining Large Language Models: Practices and Research Questions