The Case for AI at the Edge

Artificial Intelligence (AI) is becoming increasingly commonplace as organizations seek to bring effective decision making and operational efficiencies to business in ways that transform how humans and machines work together. As this transformation takes shape, the advantages of edge-based AI implementations over centralized or cloud-based models is changing how AI tools are being deployed and used. Edge-based AI brings decision making right to the device while eliminating the costs, security risks, latency, and information bottlenecks associated with centralized or cloud-based models. With decreases in costs and increases in performance for microcontrollers capable of providing AI and Machine Learning (ML) inference at the edge, businesses now have the wherewithal to reduce costs and improve efficiencies by decentralizing operations and decision-making processes.

Within the cloud-based “big data” approach to AI, sensors are largely treated as unintelligent devices lacking localized data processing capabilities. Cloud-based AI requires large volumes of raw sensor data sent across high bandwidth networks to cloud-based systems for processing by server executed algorithms. These systems use traditional AI frameworks such as Google’s TensorFlow, Caffee, Apache Spark, and others to analyze and process flowing streams of raw data originating from sensor devices spread across multiple potential networks. Such systems require complex manual interaction and data science expertise to create ML algorithms that analyze and execute AI processes. They also inject inherent risk in the form of network latency; the time delay that occurs as large volumes of raw data move across a network and are processed by server-based algorithms before responses are sent back to local devices where decision-based processing occurs. In contrast, edge-based AI brings intelligence to the data collection source by utilizing ultra-low power microcontrollers to manage sensor output. Processing occurs instantaneously on the device using local ML algorithms, enabling instantaneous decision loop management while also eliminating latency and other risks associated with cloud-based AI methodologies.

Furthering recent advancements in low cost sensors and microcontrollers, new AI algorithm automation tools combine to offer organizations and developers with even modest budgets the ability to create “smart sensors” quickly and easily. Previously, the ability to truly capitalize on the capabilities of intelligent IoT applications was limited to only a handful of use cases where the investment in time and effort to hand-code algorithms for high-volume consumer IoT devices justified the time and expense. Today, new AI algorithm automation tools like SensiML’s Toolkit Suite allow developers with domain knowledge within their field of expertise, but little or no knowledge in the field of data science, to rapidly generate ML algorithms designed specifically for the microcontroller environment where they will run. SensiML’s Toolkit Suite solves two critical problems that make IoT AI and ML a feasible reality for organizations lacking the financial wherewithal to staff data scientist personnel or to maintain expensive cloud servers.

Sizing ML Development for Microcontroller Deployment – Microcontrollers have substantially different design constraints than desktop or server computing applications. And, while the “on-premise” computing of edge-based devices offers significant advantages over cloud-based computing, creating ML algorithms for these devices require distinct firmware management competencies unless developers use intuitive ML automation tools such as SensiML’s Toolkit Suite. For example, working with microcontrollers requires some knowledge of the microcontroller’s processors (CPUs), program memory, data memory, timers and counters, I/O port, serial communication interface, clock circuits, and interrupt mechanisms. However, SensiML’s Toolkit Suite simplifies the hardware management process by allowing the developer to select the target environment for which he or she is creating the ML algorithm. By providing developers the ability to specify the target environment up front, the conceptual knowledge required to work with microcontroller firmware is simplified. And, developers have the assurance that the SensiML Toolkit Suite will only return ML algorithms that fit and function within the specified microcontroller environment.

Automating the Data Science Effort – In simple terms, data science is the process of extracting data in various forms and converting it into knowledge. Organizations or individuals can then use this knowledge to make informed business decisions, improve operations, or enhance aspects of everyday life. While the concept of data science is straight forward, the process of collecting data, performing exploratory data analysis, identifying and labeling metadata, modeling, evaluating, deploying, and testing data models varies extensively based upon business and functional objectives. By necessity, Big Data analytics tools need to be robust enough to handle very different analytics, data visualization, and database management tasks depending on the data model to be created. As a result, the quality of data models and resulting ML algorithms often relies heavily on the skill and tenacity of the data scientist(s) creating those models. And, because many organizations go through extensive, time consuming, and laborious processes to obtain an operational algorithm they tend to confuse complexity with value. The data model and the process required to update and maintain that model becomes the “secret sauce” of their business. They incorrectly assume that just because they took a circuitous and time-consuming route to create their data model(s) that others will have too as well.

SensiML’s Toolkit Suite changes this business model by allowing data scientists, or even developers who otherwise lack a data scientist background, to capture, identify, define, and label data elements using SensiML’s intuitive and easy to use Data Capture Lab. From there, SensiML’s Analytics Studio analyzes the data files and returns candidate algorithms designed to fit and function within the target hardware environment based upon the accuracy and efficiency of each algorithm. Developers can receive a compiled data model that they reference within their application, or they can receive human readable code that they embed directly into their application. The final SensiML Toolkit Suite application is SensiML’s Test App. This application allows developers to test and validate newly created or updated data models against previously untested data files. The SensiML Toolkit Suite is designed and licensed to allow developers, technicians, and testers to collaborate on the same project simultaneously. And, because the SensiML Toolkit Suite is a SaaS product and service, all data models are readily available for local or distributed team interaction. The SensiML Toolkit Suite provides an intuitive interface that walks developers through the entire data science workflow process from identifying the target hardware environment to providing the ML algorithms that will run within that environment.

Learn how quickly you can add intelligent sensors to your IoT end-point applications.

Download the Free Trial Version of SensiML’s Analytics Toolkit Suite