This post is from Laura Martinez Molera, Product Marketing Manager for Machine Learning and Data Science, here to discuss Machine Learning latest features. 

We have just launched the 2^nd release of the year, R2020b. I know it’s not easy to keep up with so many new features, so I wanted to highlight the most important updates for Machine Learning and Data Science. If you want to learn more details about them and the rest of the new features, check out the Latest Feature page.

AutoML

AutoML, or Automated Machine Learning, is a great feature for MATLAB users. This feature is intended for beginners of ML (or those with less machine learning experience) to automate certain steps of the workflow, getting to a higher accuracy in less time. It also makes experienced AI developers more productive by automating routine and time-consuming steps, allowing them to focus on further optimizing the model manually.

AutoML applies to the following steps in the machine learning workflow:

• Data Exploration and Preprocessing by identifying variables with low predictive power and highly correlated variables that should be eliminated.
• Feature Extraction and Selection by extracting features automatically from signals and images and identifying those with high predictive power.
• Model Selection and Tuning by identifying the best performing model and automatically tuning model hyperparameters at the same time.
• Preparation for Deployment by transforming high-level machine learning code into lower level languages like C/C++ for deployment on embedded devices with limited memory and low power consumption.

Here you can see a step-by-step example about building optimized models to classify human activity. If you want to check more details about AutoML check out this page. In addition, I encourage you to check out this short video that walks you through applying AutoML.

Model Interpretability

The second feature I want to introduce you to is model interpretability. This feature is intended to alleviate the “black box” nature of machine learning models, since its representations of knowledge and decision-making aren’t always intuitive. Interpretability introduces the option of its adoption in domains where black box approaches aren’t feasible, including safety and risk management, Finance and Medical applications.

Many interpretability methods analyze how variables impact model predictions. For example, the partial dependence plot shows the marginal contribution of one specific variable on model score. If the response changes significantly, it indicates the variable is important, thus “explaining” one aspect of the model. If these important variables align with the human expert’s expectation, we can say that we “understand” the model, increasing trust in its predictions. Partial Dependence Plots (PDP) have been available in MATLAB for regression, and with 20b are available for classification as well.

Another popular interpretability method available in 20b is LIME, Local Interpretable Model-Agnostic Explanations. LIME allows you to analyze the model behavior near a point of interest by approximating a complex model within the local area of interest and offering the simple model’s parameters as explanation for model behavior.

If you want to learn more about this feature, check out this video and this discovery page with all the details.

Simulink

The last feature I want to highlight is the new library of Machine Learning blocks for Simulink. In this release we will support SVM (support vector machine) Classification and Regression. With Simulink you can integrate your models with a larger system that could include components such as controls, dynamic models, sensor fusion, and computer vision. Simulink supports Model-Based Design of complex multi-component systems, including simulation of system-level performance on hardware, and facilitates deployment to hardware with embedded code generation.  

The Simulink model below represents an implementation of a human activity classifier using the Simulink block for SVM Classification. This model accesses signals from your mobile device’s accelerometers, specifies modules to calibrate and normalize the signal, extracts features, feeds them through a classification algorithm, and finally, displays the human activity as an output: running, standing, walking, laying, etc. You can check the full example here.

If you want to learn more about any of these features, you can register for the next webinar to see new examples and ask questions in a future Live Q&A.

Thanks to Laura for updating us on the important new features for Machine Learning. Have a question or comment for Laura? Leave a comment below.