Ashfaq Munshi, ML7 Fellow, Pepperdata
summary
Pepperdata ingests ~300 metrics per task every 5s (one 570-node customer yields ~41M points/min; company-wide ~300 trillion points) and maps each series to GADF images, feeds them through Inception v3 embedding to get 2048‑D vectors, then trains a tiny MLP (two hidden layers ×50 nodes, softmax) for classification. Off‑the‑shelf pipeline hits ~89.8% on UCR univariate benchmarks, outperforms published multivariate baselines on robot sensor sets, achieves ~76–78% labels accuracy on 2k+ node production clusters, and enables ~30‑minute cluster load prediction.
AI-written summary · about
chapters (13)
- 0:00Start
- 2:52Example Time Series
- 3:22Characteristics of our TS
- 4:18Problem
- 7:07Time Series and Images
- 9:23Example GADF Image
- 9:40Time Series and Bag of Patterns
- 12:16Accuracies for a subset of UCR
- 13:33Training Time Comparison
- 14:33Approaches and Data Set
- 17:425-Fold Cross Validation Error
- 19:48Accuracy per Label on PD Dataset G
- 21:28Summary
related talks
29mScaling to Millions of ML Models to Solve the Problems of SRE and Security
41mPeer Beneath the Surface of the Earth
52mDevil in the Haystack
21mMaja Kabiljo, Software Engineer, Facebook Inc.
13mBuilding Scalable AI/ML Platforms for Industrial IoT
30mScaling Machine Learning at Discord: Our Journey with Ray and Anyscale
33mAnalytics in the age of the Internet of Things #bbuzz
34mScaling AI Like a Pro-PepsiCo’s LLM Deployment Strategy on K...