Just like humans, AI algorithms have their own strengths and weaknesses. This is what helps us decide which human may be better suited for a specific job and which algorithm is the best candidate for a specific problem dataset. However, just like humans coming together to work in teams do a fabulous job, AI algorithms combined can enhance the accuracy of results in specific cases. This also applies to algorithms used for edge analytics.

Commonly used algorithms in edge analytics, by functionality, are 

• Classification
• Regression
• Object detection and segmentation
• Anomaly detection
• Clustering
• Dimensionality reduction
• Transformation

The assumption in this article is that users are familiar with these algorithms. While each of these algorithms works splendidly in specific use cases, many can be paired to build more accurate and powerful solutions. This article focuses on the different approaches to combining these algorithms. 

This is where the fun and creative part comes into play. You can build powerful and unique solutions with a deep knowledge of these algorithms and an understanding of business processes. The four common approaches to combining algorithms for edge AI are:

• Ensembles
• Cascades
• Feature Extractors
• Multimodal models

Ensembles: It is a widely used approach in data science. A set of ML models is fed the same input data, and the output is combined mathematically to attain more accurate results than the ones generated by individual algorithms. You need to understand the ML models and their behaviors to finalize the optimal set of models.

Cascades: In this approach, models are leveraged in a specific sequence in the pipeline. The results from one model may be used to trigger the next model. In the context of Industry 4.0, an example can be that an algorithm on a smart camera on a dock gets triggered when another algorithm detects that the door is opening.

Feature extractors: These are typically a combination of an embedding model that reduces dimensionality in a data set and then feeds that data as input data to another model. So, suppose you are trying to develop an algorithm to identify damaged packages. In that case, you can use the output from a feature extractor to train the model, reducing the data needed and training time. This approach is used widely, and many pre-trained feature extractors are available in the open-source world.

Multi-modal models: A multimodal model takes multiple data types as inputs simultaneously. An Industry 4.0 context example can be when an algorithm takes vibrations audio and accelerometer data together. The powerful application of this category is in the fusion of sensor-generated/captured data, combining different data types.

In my opinion, the differentiating capabilities in the Industry 4.0 context can only be built using combinations of algorithms and processes. Vanilla approaches like predictive maintenance have become almost commoditized, and there is a need to evolve capabilities in this area.