November 6, 2017

Concept Networks: Combining Subject Matter Expertise & Machine Learning to Build Industrial AI

Most of the chatter around AI focuses on problems such as data analysis (data mining, customer segmentation), prediction (churn prediction, fraud detection, product recommendation) and perception (facial recognition, language translation, identifying birds).

Typically, those kinds of AI problems are finite, constrained and relatively low in risk. Nobody dies if you recommend the wrong book to a customer, misidentify a warbler or mistranslate a word from Greek into French.

But there’s an untapped group of large and complex physical systems in which modern machine learning technologies can be highly effective; areas such as autonomous transportation, manufacturing processes, supply chain logistics and advanced robotics. Applying AI to these types of systems, which we refer to as Industrial AI, essentially means using machine learning to automate the control of a physical system (ie: manufacturing line) or optimize the decisions and actions of an enterprise system (ie: supply chain).

While applying machine learning to these systems present huge business opportunities, they pose their own special kinds of challenges:

  1. The risks of automating physical systems are much higher. And there are often real physical dangers, such as when robots work alongside humans.
  2. Available real-world data is often limited. Training AI models in the real world is more difficult than training models in the lab. The cost of running experiments with real-world systems can be prohibitive.
  3. Physical spaces are large and complex. Training robots to execute difficult tasks requires time and money. Robots don’t learn tasks and skills on their own; they need to be taught, just like humans.

Leveraging Domain Expertise to Solve Complex AI Problems

These types of challenges highlight the importance of understanding your business problem before developing an AI strategy, both to understand potential risks and to be able to most effectively break down the problem into smaller, more manageable pieces.

Concept networks allow subject matter experts to break down a large, complex problem into smaller sub-concepts. An AI model can learn to solve each sub-concept before combining all of the trained sub-concepts to solve the end goal. The subject matter expert is able to break down a complicated problem and teach an AI model how to solve it just as a human would, piece by piece.

Tackling large problem spaces with concept networks leads to a number of benefits for enterprises:

  1. Speed - Faster training of AI models with fewer physical risks.
  2. Reusability - Organizations will build up a library of concepts which can be reused to solve a number of problems, which is almost always more cost-effective than rewriting code from scratch.
  3. Scaled expertise - Domain experts can distill and codify their knowledge and experience into programs at a high level, without becoming machine learning experts.
  4. Explainability - The AI model will make decisions based on the conceptual hierarchy that it’s been trained on, so there is a modicum of insight into why a model is making certain actions

Concept Networks in Action: HVAC

Let’s look at an example of how subject matter experts would use concept networks to control an industrial-scale HVAC system in an office building. The expert wants to train an AI model to control the building’s environment, providing maximum comfort for workers while keeping energy costs minimal

In the old days, you would have set the temperature controls on the building’s HVAC system, walked away and hoped for the best. But with concept networks, you can train an AI model to optimize for multiple smaller concepts at once, all of which contribute to solving a larger problem.A subject matter expert might determine that a model needs to learn about concepts like inside temperature, outside temperature, humidity, sunshine, cloud cover, time of day, day of the week, season of the year, how many people are in the building and where they are located in order to make the best decisions on how to set the temperature of the room.

Those concepts can be seen as nodes in a network; changing the condition of a node will send ripples across the network. By first learning how to optimize for each individual concept, the system can then learn how each decision affects the rest of the environment (ie: turning on the air conditioner to lower the temperatures inside a building will also shorten the life of the air conditioner, raise electricity costs and require more frequent changes of air filters throughout the HVAC system) and ultimately learn how to leverage all concepts to most effectively heat the room.

Applied Deep Reinforcement Learning

Concept networks are just one feature of the Bonsai Platform - along with Gears and Machine Teaching - that brings  deep reinforcement learning closer to the enterprise.

To better understand how concept networks result in intelligent control of industrial systems,  check out how we taught a robotic arm to stack and grasp blocks using 5 low-level concepts. You can also head over to our Getting Started page to learn how you can incorporate concept networks and reinforcement learning in your own organization.

Always. Be. Learning.

Stay up-to-date on our latest product news, AI industry highlights, and more!