Machine Teaching, the not-so-secret AI weapon
Deep Neural Network-based (DNN) AI solutions often behave like a black box from an AI practitioner’s perspective. The system is trained using labeled data and converges, starting from randomly assigned coefficients, to the ones that the system learns as being the most effective to solve a particular problem. In modern DNNs, the number of coefficients used in these networks ranges from a few million to, for the most advanced ones, several billion.
However, it’s mostly impossible at the AI practitioner level to know what a neuron does and why. Trying to understand why the DNN does what it does is referred to as explainability. AI explainability is an advanced research field, and although some progress has happened – e.g., in vision AI-, DNNs remain mostly black boxes for the external world.
To help with these explainability issues, Microsoft researchers have developed the concept of Machine Teaching (MT). MT harvests human expertise to breakdown the problem that the AI is supposed to solve into smaller, more explainable ones. As Mark Hammond, Microsoft GM for business AI, in his blog post on machine teaching, wrote: “[MT] relies on people’s expertise to break a problem into easier tasks and give machine learning models important clues about how to find a solution faster. It’s like teaching a child to hit a home run by first putting the ball on the tee, then tossing an underhand pitch, and eventually moving on to fastballs.”
In real-life business use cases, AI engineers will leverage process expertise from operators and engineers to design the solution.
Machine Teaching key concepts
Contrary to what its name might suggest, machine teaching is less about specific Machine Learning (ML) algorithms and is more about leveraging expertise from process practitioners to help design AI solutions.
A recent development in the AI field, MT optimizes AI design by harvesting human expertise. It impacts multiple aspects of the AI solution design and training.
Specifically, with Autonomous systems where MT is most commonly used, this expertise is used to define four key aspects:
- Solution parameters selection
- Application scope
- System design
- Training parameter (aka the so-called “reward function”)
Solution parameters selection using Machine Teaching
In real life systems, the number of potential variables that an AI controller might need to consider could be very large. However, in most cases, only a subset of those parameters is significant for controlling a system.
Let’s start by using a car analogy to illustrate that concept. There are hundreds, if not thousands, parameters that impact a car’s gas consumption, from tire pressure to gasoline quality to driving habits and more. However, during normal operations and once the tire pressure is set, and the gas tank is filled, the principal elements to estimate gas consumption are traffic and the driver’s skills.
A similar approach is also relevant for control systems, supply chains, or any other advanced system an AI would control. By leveraging the expertise of the people that have designed and are managing, operating, and maintaining the system, AI practitioners can select the most relevant inputs and outputs necessary to build this AI control system
Similarly, the same experts will be able to define the scope of those inputs and outputs. Their expertise will help specify the parameters ranges the system should support.
The scope can be as simple as defining mean minimum and maximum values for each parameter. However, these parameters could also be correlated so that a simple range definition would not be enough.
For instance, to come back to the car example, one could define the scope as two distinct situations: city versus highway driving. In manufacturing, these could, for example, be regular vs. failover modes.
The illustration below illustrates, in a theoretical situation, how three variables could define multiple appropriate operating scopes.
Machine Teaching-led system design
After the process experts define the core parameters and the applications scope (or scopes) they will help design the AI architecture. They will, for instance, help AI practitioners decide how many AI agents the solution will require.
For the car example, one could build two different self-driving AI agents: one for city driving and one for highway driving.
Leveraging Machine Teaching to select training parameters
Machine teaching primarily use case is for Autonomous Systems. In Autonomous Systems, AI agents are not trained using the traditional machine learning approach of using pre-labeled training data sets. Autonomous Systems use Deep Reinforcement Learning (DRL) to train the AI agent. DRLs leverage so-called reward functions that drive the AI trial and error learning process by sending positive or negative feedback after each action the AI performs.
But how to define this reward function?
With Machine Teaching, process experts use their experience to define heuristics that are then translated as a DRL reward function.
In our car gas consumption example, a heuristic could be that if a traffic light is red at a distance, the driver should start slowing down early instead of hitting the brake right before the car reaches the traffic light.
Learn more about Machine Teaching
Learn more details about machine teaching concepts in this video, the second of our 5-parts series on Autonomous Systems, as well as the additional resources listed below.
- Learn about Autonomous Systems
- First blog post of this series: Autonomous Systems 101
- 5-parts video series on Autonomous Systems: https://go.nealanalytics.com/AS101
- Blogpost from Mark Hammond, Microsoft GM for Business AI
- For a more in-depth review of Machine Teaching refer to this Microsoft article
- Wall Street Journal article on how Pepsi is making the perfect Cheetos using Autonomous Systems