The Model Efficiency Pipeline: Enabling Deep Learning Inference at the Edge

Explore a comprehensive keynote presentation from the tinyML EMEA 2021 conference focusing on the model efficiency pipeline for enabling deep learning inference at the edge. Delve into the challenges of deploying AI applications on low-power edge devices and wearable platforms, and discover a systematic approach to optimize deep learning models. Learn about Hardware-Aware Neural Architecture Search, compression and pruning techniques, and state-of-the-art quantization tools. Gain insights into mixed-precision hardware-aware neural architecture search and conditional processing as future trends in efficient edge computing. Examine real-world examples, key results, and practical applications across various domains, including video processing and semantic segmentation.

Introduction
About Qualcomm AI Research
Challenges with AI workloads
Model efficiency pipeline
Challenges
DONNA
Fourstep process
Example
Blocks
Models
Accuracy predictor
Yields
Linear regression
Evolutionary search
Evolutionary sampling
Finetuning
Results
Model pruning
Unstructured pruning
Structured compression
Main takeaway
Quantization research
Quantization
Recent papers
Adaptive rounding
AI model efficiency tool
Key results
Highlevel view
Mixed precision
Mixed precision on a chip
APQ
Running networks conditionally
Classification example
Multiscale dense nets
Semantic segmentation
Dynamic convolutions
Video processing
Skip convolutions
Video classification
Summary
Questions
Sponsors