Resource Utilization as a Metric for Machine Learning
summary
Shows concrete wins: MNIST: 26% → 93% after simple normalization in two epochs, swapping mean-centering also works; optimizer comparison finds SGD ~54% after one epoch vs RMSProp ~85% after one epoch; activations give sigmoid ~79% vs ReLU ~86%; a learning-rate decay callback reduces lr over iterations and stabilizes accuracy. Resource experiments convert an NVIDIA autopilot model from ~132k params with 8.7 GB RAM to 30k params and ~800 MB using downscaling and HSC filtering. Delivers an Indian Sign Language recognizer built for low‑resource devices that runs but degrades in crowded scenes.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:20Welcome
- 3:00Follow along
- 3:30What is Machine Learning
- 4:02Example
- 4:24M9XT
- 5:08Analysis
- 6:08Normalization
- 11:28Optimizers
- 13:20Activation
- 14:41Learning
- 16:58Digital Encoder
- 19:47Network Architecture
- 25:44Expectations
related talks
19mResource Management Innovation for AI Workloads in Cloud Computing
33mRunning Machine Learning in Production - a Journey to Success
49mMobiconf 2018: Vladimir Alekseichenko, DataWorkshop
42mPost-Training Small Language Models: the adventures of a practitioner
16mTowards GPU Utilization Prediction for Cloud Deep Learning
17mMaximizing Utilization for LLM Accelerators
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
15mUSHER: Holistic Interference Avoidance for Resource Optimized ML Inference