RIANN: Real-time Incremental Learning with Approximate Nearest Neighbor on Mobile Devices
summary
RIANN is an on-device system that builds graph-based ANN indexes with per-point, dynamic construction parameters and uses a trained statistical prediction model (gradient-boosted decision trees) plus an analytical performance model to pick hyperparameters via simulated annealing. On a Samsung S9 it reports a 2.42× overall speedup over the prior HNSW baseline, and for 100 incremental batches satisfies 96.67% of timing/recall constraints versus 55.33% for the state of the art.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:40Approximate Nearest Neighbor (ANN)
- 1:14Applications
- 1:39Graph-based ANN Algorithms and HNSW
- 2:49Challenge 1
- 4:44Overview of RIANN
- 5:37Dynamic ANN Graph Construction
- 7:42Statistical Prediction Model
- 10:19Analytical Performance Model
- 12:56Experiments Configuration
- 13:34Evaluation Results
related talks
18mScalable Billion-point Approximate Nearest Neighbor Search Using SmartSSDs
38mNEURAL NETWORKS ON YOUR MOBILE PHONE. DIVING INTO THE NN API
17mUnleash All Cores: Asymmetry-Aware Scalable DNN Inference on Mobile CPUs
10man introduction to approximate nearest neighbors
06mIntermittent Learning: On-Device Machine Learning on Intermittently Powered System
48mHolder Homeomorphisms and Approximate Nearest Neighbors
39mAI/ML on mobile/android with TensorFlow
17mADR-X: ANN-Assisted Wireless Link Rate Adaptation for Compute-Constrained Embedded...