Machine Learning for Software Developers (and Knitters)

Our first model mistook a cat for a knitting pattern. Learn how we iterated to over 90% accuracy and what it reveals about real-world machine learning.

#1about 5 minutes

Inspiring real-world applications of AI and machine learning

AI is being used for scientific breakthroughs like protein folding, creating art with GANs and neural style transfer, and optimizing business logistics.

#2about 4 minutes

Why developers must understand the Software 2.0 paradigm

The shift to "Software 2.0" means developers will train software instead of writing explicit instructions, requiring them to handle probabilistic outputs and confidence levels.

#3about 3 minutes

An overview of the three-layer AWS AI/ML stack

The AWS stack is structured in three layers to cater to different skill levels, from pre-trained AI services to the comprehensive SageMaker platform and foundational infrastructure.

#4about 6 minutes

Introducing the KnitML project to reverse engineer knitting

The project aims to use image classification to identify knitting stitch patterns from a photograph, simplifying a complex reverse-engineering problem.

#5about 9 minutes

Crowdsourcing and labeling training data for the model

An automated email pipeline using SES, Lambda, and S3 was built to crowdsource images, which were then labeled by volunteers using SageMaker Ground Truth.

#6about 5 minutes

Training the first model with Amazon SageMaker

After simplifying the problem to binary classification, the first model was trained using the SageMaker console, but the initial results were highly inaccurate.

#7about 4 minutes

Improving model accuracy with data augmentation and tuning

Model accuracy was improved from 45% to 95% by applying data augmentation techniques, adding non-knitting "clutter" images, and running a hyperparameter tuning job.

#8about 5 minutes

Rebuilding the model with Amazon Rekognition Custom Labels

Amazon Rekognition Custom Labels provides a simpler, managed alternative to SageMaker for building custom image classification models with a built-in UI and less expertise required.

#9about 6 minutes

Comparing SageMaker and Rekognition for custom models

While both services achieved comparable accuracy, SageMaker offers more control and cost-saving options for experts, whereas Rekognition prioritizes ease of use and speed for non-specialists.