Qeexo

Depending on the application, data, and use case, there may not be one model that performs best for all metrics: accuracy, latency, and memory size. Qeexo AutoML allows you to build multiple models all at once, so you can compare the performace results and select the best model that suits your needs.

“Instance Length” is the time duration in milliseconds of each window on the sensor data stream that we want to classify. “Classification Interval” is the time gap in millisecond between running each classification. You can let Qeexo AutoML choose these values for you by selecting “Determine Automatically”, or you can enter them manually.

A UMAP plot is a graph displaying the “Uniform Manifold Approximation and Projection”, which visually shows how separable the classes under consideration are with respect to the selected group of features. It is a 2D plot and represents each class as a cluster of points in a unique color. You want to have these clusters as far away from each other as possible for the best classification. You can visually inspect UMAP plots for your selected sensors and feature groups. We recommend choosing sensors and feature groups that result in maximally separated clusters.

“Group Labels” is a handy tool if you are collecting data for multiple classes and then want to group a few classes into one. For example, if you are classifying human activities as WALKING, RUNNING, BIKING, and IDLE, you may have collected data of human subjects sitting, standing, and working on the computer. You can group sitting, standing, and working on the computer classes as ‘IDLE’. During live classification, if human subject is sitting, standing or working on the computer, they will be classified as ‘IDLE’.

Before grouping labels, we recommend building a model with all labels and then checking the Matthews Correlation Coefficient (MCC) table in the Results section. The pairs which have low MCC values are generally not separable. We recommend grouping these labels for best performance.

When CNN and ANN models are selected in the Model Settings screen, the “CONFIGURE” button will become click-able, and you can configure various parameters. For a detailed description of the parameters, please refer to the Qeexo AutoML User Guide.

We expect the model to perform better with Hyperparameter Tuning turned ON. However, model training may take longer since the training algorithm has to explore the multi-dimensional parameter space.