Down the Wabbit Hole
summary
Formation builds polymorphic model handles in Haskell that abstract TensorFlow sessions and Vowpal Wabbit pointers so the same pipeline runs both frameworks. The code uses type families, a handle/bracket lifecycle, continuation-based VW serializers to safely turn byte strings into examples, and conduits plus lenses to run online SGD with periodic holdouts. An IPS holdout scorer estimates counterfactual rewards so different models are compared fairly; the system runs on EMR/Kubernetes and proves practical in production.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 0:37Best practices in machine learning
- 2:18Gluecode antipattern
- 3:44Handle pattern
- 6:46Polymorphic handles
- 11:05Simple example
- 15:04Configuring VW
- 19:09Example
- 24:52Action Dependent Features
- 26:08Type Parameters
- 26:47Example Type
- 27:35Has Pattern
- 28:33SGD Harness
- 29:51Type Signature
- 31:15Pipeline
- 31:51Learn
- 33:02Scoring
- 35:25Holdout
- 36:58Logger
- 37:43Learning Holdout
- 37:54Lift
- 39:06Machine Learning Architecture
- 39:32Questions
related talks
44mDown the Wabbit Hole
34mOnline learning, Vowpal Wabbit and Hadoop #bbuzz
19mThe Small Model Infrastructure Nobody Built (So We Did)
31mHugging Face ecosystem for Local AI/ ML
01h 03mScaling up machine learning without tears (and what do programming languages ha...
37mGetting Started with AI and WebAssembly
17mCode World Model: Building World Models for Computation
34mBeyond Transformers - Intro to RWKV Architecture & The World To...