30 Golden Rules of Deep Learning Performance

Siddha Ganju

Deep Learning Machine-Learning

See in schedule

“Watching paint dry is faster than training my deep learning model.”
“If only I had ten more GPUs, I could train my model in time.”
“I want to run my model on a cheap smartphone, but it’s probably too heavy and slow.”

If this sounds like you, then you might like this talk.

Exploring the landscape of training and inference, we cover a myriad of tricks that step-by-step improve the efficiency of most deep learning pipelines, reduce wasted hardware cycles, and make them cost-effective. We identify and fix inefficiencies across different parts of the pipeline, including data preparation, reading and augmentation, training, and inference.

With a data-driven approach and easy-to-replicate TensorFlow examples, finely tune the knobs of your deep learning pipeline to get the best out of your hardware. And with the money you save, demand a raise!

Type: Talk (45 mins); Python level: Beginner; Domain level: Beginner

Siddha Ganju


Siddha Ganju, an AI researcher who Forbes featured in their 30 under 30 list, is a Self-Driving Architect at Nvidia. As an AI Advisor to NASA FDL, she helped build an automated meteor detection pipeline for the CAMS project at NASA, which ended up discovering a comet. Previously at Deep Vision, she developed deep learning models for resource-constrained edge devices. Her work ranges from Visual Question Answering to Generative Adversarial Networks to gathering insights from CERN's petabyte-scale data and has been published at top-tier conferences including CVPR and NeurIPS. She has served as a featured jury member in several international tech competitions including CES. As an advocate for diversity and inclusion in technology, she speaks at schools and colleges to motivate and grow a new generation of technologies from all backgrounds.