Neural Trees: Using Neural Networks as an Alternative to Binary Comparison in
Talkat 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage '20) · 2020-07-13..07-14·
summary
Neural Trees implement a hierarchy of many tiny neural models that map keys to physical addresses instead of learning the data distribution, claiming up to 40% higher fan-out per byte versus B+ trees. They pack multiple single-layer perceptrons into 4KB blocks, retrain locally only on leaf splits with mean training time ~1 ms, and use preprocess/postprocess operators to bound inference error. Evaluated on SPC‑1 and weeks of production VDI traces, lookup paths shrink and lookups become comparable or faster without expensive global retraining.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 0:19Binary Comparison
- 3:00Technical Contributions
- 4:48Neural Networks
- 5:29Neural Tree
- 9:11Neural Network Model
- 13:57Future work
related talks
37mLearned Indexes: a New Idea for Efficient Data Access #bbuzz
14mRe-architecting Traffic Analysis with Neural Network Interface Cards
12mDistilling models using the linear representation hypothesis
21mNeural Turing Machines: Perils and Promise
14mBrain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...
44mFast Neural Networks… a no brainer!
38mModern B-Tree techniques
01h 51mUnderstanding Neural Networks (Somewhat)