MATLAB, which provides integration with deep learning frameworks,
Supporting all processes including development and distribution of AI models
Reduce optimization time and effort by providing an integrated environment


The global spread of COVID-19 has accelerated digitalization in almost all industries. As a result, many companies are rushing to introduce AI into their businesses. According to Gartner, the average number of AI projects for companies is expected to increase from 4 in 2019 to 35 in 2022.

However, the AI knowledge and experience of experts in each industry who must design and implement AI systems are still lacking. On the 16th, Jongnam Kim, an application engineer at MathWorks, held a live demo titled ‘Linking MATLAB and Deep Learning Open Framework’ on YouTube.

In the demo, Vice President Kim Jong-nam explained the workflow of calling MATLAB from PyTorch to train an AI model and exporting it back to MATLAB to generate code, and the workflow of calling a TensorFlow neural network architecture from MATLAB and developing an AI model.


Connecting MATLAB and Deep Learning Frameworks
Deep Learning is a type of machine learning that uses multilayer neural networks. Neural network algorithms used in deep learning include DNN, CNN, RNN, RBM, DBN, etc., and deep learning frameworks provide various deep learning algorithms with verified libraries and pre-training. Engineers can use this to develop core algorithms for problem solving.

TensorFlow and PyTorch are deep learning open frameworks that support neural network design. Engineers can import models designed with these external deep learning frameworks into an integrated environment called MATLAB to perform optimized analysis, training, and C/C++ CUDA code generation.

This is possible because MATLAB can interoperate with Python, and importers can be used to bring Keras-Tensorflow and Caffe models into the MATLAB environment.

MATLAB supports ONNX, an industry standard that supports the interoperability of AI models. Therefore, AI models from PyTorch, Caffe2, MXNet, Core ML, and TensorFlow that are compatible with ONNX can all be imported into the MATLAB environment.

MATLAB has a variety of toolboxes for automated data labeling and preprocessing in engineering and mathematics. Engineers can perform data preparation tasks such as labeling and preprocessing with MATLAB, and then export the prepared data to an external deep learning framework. Engineers can use the data prepared with MATLAB in the deep learning framework to complete the development of AI models.

This AI model can also be imported back into MATLAB. Engineers can then use the AI models imported into MATLAB to perform subsequent tasks such as analysis, scenario-based training, code generation, and deployment to targeted embedded hardware.

MATLAB supports: △Code Generation and Compiler, △Model Analysis through Visualization, △Model Debugging, △Checking network structure and layer-by-layer properties through Deep Network Designer and Network Analyzer, △Parameter selection for deep learning network training and network retraining, etc.

In addition, it provides system integration through △Simulink. Engineers can integrate AI models developed with external deep learning frameworks into higher-level systems such as reinforcement learning or autonomous driving, and conduct verification through simulation and testing.

Manager Kim Jong-nam used PyTorch in the demo. He preprocessed data for a speech recognition AI model in MATLAB and passed the data to PyTorch through the MATLAB engine interface. And I opened MATLAB with the MATLAB engine interface in PyTorch and checked the functions and operation results.

After training the AI model in PyTorch, we exported the model to the ONNX version, then imported the ONNX file in MATLAB, verified the model, and generated code.

We also demonstrated how to import a TensorFlow AI module developed in Python into the MATLAB environment. Manager Kim showed that the design of the neural network architecture corresponding to the AI module can be confirmed in the MATLAB environment. We also initialized the module and entered data to conduct AI model training.

“MATLAB supports the entire workflow for developing AI models,” said Vice President Kim Jong-nam. “It provides guides and tools to help AI models installed on purpose-built embedded systems achieve optimal performance.”