Everything but the Training: 3 Ways to Enhance TensorFlow with MATLAB

Abstract: As Deep Learning becomes more prevalent across industries, there is a growing need to make it broadly available, accessible, and applicable – not just for data scientists but to engineers and scientists with varying specializations. While open-source frameworks like TensorFlow are widely used to train deep learning networks, other key aspects for building consumer and industrial applications remain challenging, including:
• Curate labelled datasets for supervised learning, including data augmentation and generation
• Apply traditional signal and image processing techniques to assist deep learning, and
• Integrate models with embedded or enterprise systems.

MATLAB is well-known for its strength in traditional engineering and scientific applications like image and signal processing, controls, and wireless system design. This talk demonstrates how MATLAB works with various deep learning frameworks to make it easier to develop, deploy, and maintain AI-powered applications for many industrial applications. Learn how MATLAB can help you:
• Automate ground truth labeling for image, video, Lidar and sensor data
• Apply physical models and simulations to augment training data, develop control algorithms and test the integrated system, potentially with hardware in the loop
• Generate high-performance C++ and CUDA engines for embedded system and cloud deployment

MATLAB helps you accomplish this by interoperating with TensorFlow and other deep learning frameworks at different points of your workflow:
• Data interoperability (like Apache Arrow) lets you preprocess signals or automatically label data in one framework while training models in another
• Model exchange formats (like ONNX) or importers let you evaluate and optimize a model trained in a different framework
• Compilation and automated code generation makes it easy to integrate models in cloud environments and with embedded or enterprise systems

Bio: Shounak Mitra is the Product Manager for Deep Learning at The MathWorks Inc. He has over 7 years’ experience in leading teams and developing products rooted in Artificial Intelligence, Computer Vision, Natural Language Processing, and Statistical Modeling across industry and academia. He holds two Master’s degree from University of New Hampshire – one in Mathematics and Statistics and the other in Structural Engineering with a research focus on applying machine learning principles for vision and vibration analysis. Currently, he focuses on core command line algorithms for deep learning, interfacing MATLAB with 3P frameworks like TensorFlow and PyTorch, building apps and tools for networks designs etc.