AWS AutoML for Vision

Factored built an AWS AutoML platform for computer vision that improved accuracy 20% over Vertex AI and sped up model deployment.

Key Takeaways:

Deep Learning Model Training.

Machine learning engineers often struggle to create, train, and improve deep learning models, especially in computer vision, when working with smaller datasets. A client wanted to build a platform where users could input their dataset and desired task, and the platform would automatically train an efficient model. The primary challenge was ensuring that the platform could enhance model performance with limited data and perform synthetic data generation to improve results.

Deep Learning Models Improvement.

• Factored engineers implemented various state-of-the-art semi-supervised and self-supervised learning techniques to improve model performance using smaller datasets.
• We also implemented style transfer and generative techniques to perform synthetic data augmentation.
• Lastly, we developed complex TensorFlow routines to optimize GPU usage. The platform outperformed Vertex AI in several image benchmarks.

Improved Performance Over Other Vendors.

• 20% higher accuracy compared to Vertex AI in image processing benchmarks.
• Significant reduction in time-to-deployment and increased model accuracy.

Continue Reading