Machine Teaching to Improve Artificial Intelligence
November 17, 2020
We provide machine learning models with volumes of data expecting that they will identify patterns or learn relationships within the data. This can be an expensive and time-consuming process.
However, an approach exists that leverages human abilities and expertise to guide machine learning models to find solutions much faster. This is called machine teaching. We’ll explore this approach in this article.
Table of Contents
Machine teaching vs machine learning
Benefits of machine teaching
Applications of machine teaching
You’re only required to be familiar with basic machine learning concepts. Here’s a link to an article that offers an introduction to machine learning.
Research around the field of machine learning today takes a technocentric approach. Meaning it centers this approach on technology and promotes its importance. It addresses technical and functional goals, often at the expense of human values and ethics. As a result, this perspective may lead to some level of disconnection between the growth of artificial intelligence as a field and the users of the technology.
To curb this disconnection, an alternative approach is emerging. This approach is placing the focus on domain experts and end-users as opposed to machine learning experts. This concept is referred to as machine teaching. Machine teaching looks to get knowledge from people as opposed to data as the sole source.
The idea behind machine teaching is adamant that any information processing skill that can be taught to a human can be easily taught to a machine. To develop solutions to these problems, machine teaching uses human expertise and human abilities to assist machine learning models in identifying crucial clues on how to generate a solution faster.
A human being, who is a domain expert, acts as a teacher to a machine-based learner. An important result of this is that domain knowledge is instilled into the resulting machine learning system. Machine teaching provides optimal training data for driving the learning algorithm towards the desired output.
How it Works
Machine teaching combines the expertise of human domain experts with machine learning to automate the management of machine learning algorithms. A goal of machine teaching is to reduce manual intervention in the learning process.
The teacher has a central role in this process:
The teacher ensures an effective model is generated by improving the transfer of knowledge to the learning algorithm.
The teacher is actively involved in the collection and labeling of data. They ensure an optimal dataset is collected by filtering data based on domain expertise or intuition. They decide which dataset is best used to carry out a task.
Let’s go through the machine teaching process. It has three stages; building, training, and deployment.
Building stage. This stage involves the creation of the machine teaching program. We then connect the program to a training simulator that is domain-specific. The role of these simulators is the creation of adequate training data. The training data generated by simulated environments cover various use cases. These simulations mimic real-world conditions and scenarios to train the model effectively.
Training stage. A training engine handles this phase. It automates the generation and training of machine learning models. It combines these models with neural networks and models containing domain-specific expertise. The latter results from the building stage.
Deployment stage. Trained models are then deployed to their desired use cases.
Machine teaching vs machine learning
To better understand machine teaching, it’s crucial to distinguish it from traditional machine learning.
Traditional machine learning takes a learner-centered approach. What this means is that the focus is on a machine-based learner to identify relationships, generate insights, and make predictions from vast amounts of data. This implies that the key source of knowledge for traditional machine learning models is sizeable amounts of data.
But, with machine teaching, it shifts the perspective to focus on the “teacher”. The teacher in this case is a human who possesses expertise in a domain. The teacher leverages their experience, abilities, and domain expertise to train machine learning models. The involvement of the human element as the teacher allows the breaking down and decomposing of useful concepts to assist machine learning models to be trained efficiently.
Traditional machine learning often involves the use of lots of labeled data to train a machine learning model. It also requires expertise in the form of data science or machine learning experts for each task. The availability of both is limited therefore the opportunity to build models is available only to those with the aforementioned expertise.
Machine teaching operates a bit differently. Since the focus is on imparting domain knowledge into machine learning models to drive their learning. This means that a wide range of knowledge is available to help train the model. The machine teaching process doesn’t need one to be a data scientist or for one to have a large labeled data set to build effective machine learning models.
Duration of training
As we mentioned before, traditional machine learning involves the use of a lot of labeled data to train a model. The learner-centered approach means the model is exposed to a lot of data and it’s up to the model to identify patterns and relationships in data.
This approach may not be the most effective one, since the datasets may not be optimal for the training of the model. There might be data that is of little value to the training of the model that ends up unnecessarily prolonging the training time of the model.
With machine teaching, the training process might be more efficient since it gives the model an optimal dataset to learn from. It gives the model hints and clues on where to look for relationships in data. Compare this with a scenario where a model is given a very broad or general dataset. We expect this model to identify patterns in data with no clues or guides.
The model in this context would end up taking longer to train in comparison with one that already knows where to look and what to look for in data. Compared to traditional machine learning, the time taken to build a model can be much shorter using machine teaching.
Benefits of machine teaching
Democratization of knowledge
Machine learning development has been in the hands of machine learning experts for far too long. It would previously be very difficult for anyone who is not a data scientist or machine learning expert to build a model.
Machine teaching gives non-machine learning experts the tools to not only build machine learning models but also teach the models to perform better. It makes machine learning technology simpler for anyone to use. The know-how to develop and use machine learning models is thus opened up to non-machine learning experts.
Machine learning models are used to solve many problems in many domains. They are trusted and believed to have the capability to impressively solve a wide range of problems. Machine learning implementation is unlimited.
For instance, we can implement machine learning in the medical sector. It can be impactful in the transportation sector. There are use cases in finance. Agriculture is benefitting from machine learning technologies.
It would be helpful if those developing the models also had the domain knowledge needed in the sectors where those models would be needed to be implemented.
However, machine learning experts may have domain knowledge but may not be domain experts. Which may lead to useful but not the most effective models.
No one understands a given domain better than a domain expert. Machine teaching allows domain experts to be involved in the machine learning process. This facilitates the development of custom-made models.
These experts can offer perspectives, methods, and insights a data scientist with general domain knowledge would otherwise miss. The models developed as a result feel like they are tailor made to solve the problem.
Both pointed mentioned above result in greater innovation. Non-machine learning experts benefit immensely from the machine teaching approach.
Creativity is subjective. Since creativity and innovation cannot be monopolized, statistically, the more the people (outside of machine learning practitioners) who can develop machine learning solutions, the greater the opportunity for innovation.
Putting this power in the hands of people with different perspectives, skills, and approaches to solving problems allows them to be creative and diverse in developing solutions.
They know the pain points, what can be optimized, and where solutions are needed most. Therefore, such a person is the most suitable to develop a machine learning solution for their particular domain of expertise.
Accessible machine learning
A key benefit of machine teaching is that it is making machine learning accessible to as many people as possible. We can argue that the greatest barrier to entry to machine learning development is the need to have machine learning expertise.
Circumventing this barrier thanks to the democratization of knowledge and development of machine teaching tools opens up the field of machine learning to millions of people. People with little to no machine learning skills can be involved in developing machine learning solutions.
Automation of model building and management
The automation of models is valuable in an enterprise setting. Consider a scenario where an enterprise uses a traditional machine learning approach to automate insurance assessments. First, such a solution would have to be developed by a machine learning expert.
Second, the labeling of data in the form of insurance documents, for example, would involve the use of many people. In regards to maintenance and tweaking the system, the enterprise would have to outsource an expert.
With a machine teaching approach, a person within the enterprise would identify the definitive features in insurance assessment documents. The person’s expertise would then be translated into a machine learning model to carry out the insurance assessments.
The model would learn how to carry out said assessments. It would also learn to deal with new scenarios that would previously require the involvement of an expert. We’ll explore an industrial example later on. The company would save a lot of time and money with machine teaching.
Applications of machine teaching
Below are a few examples of how machine teaching uses data and human domain knowledge to carry out desired tasks.
These examples are represented as demos by Microsoft. Check them out for different perspective.
Computer numerical control (CNC) is a technique that’s used to automate the control of machine tools like drills and boring tools. It’s done via software that is embedded in a microcomputer that’s attached to a tool. Spinning tools are used to cut metal as well as plastic parts.
However, during cutting, the friction between the machine and the target leads to a reduction in precision. An expert operator of the machine has to routinely recalibrate it (manually) until it’s precise again.
With machine teaching, we may train a model using the machine operator’s expertise to automatically calibrate and recalibrate the CNC machines. The model may end up being more accurate than a human operator. This makes the process faster and more accurate compared to physical human intervention.
Heating, ventilation, and air conditioning (HVAC) systems in buildings consume a lot of energy, often ineffectively. They also struggle to maintain safe CO2 levels. This may cause occupants to experience a range of discomfort.
Consider the HVAC system of a smart building. Designed to save energy, manage the levels of CO2, and keep the tenants safe and comfortable. This may very well be a possibility if an expert can train a model to achieve this.
We may teach a model to supply heat when ever-varying energy cost is cheapest. The model may learn the optimum time of day to freshen the air and vary the temperature in the building. This would improve the quality of life in such a building.
Motion control is used to automate a wide range of machines in an industrial or manufacturing setting. We can apply automation to oil drill rigs.
You often find an operator manipulating the drill using a joystick or tablet device. This is done to avoid obstacles and to keep the drill on course as it goes underground.
Instead of always depending on a human being overseeing this process, we can train a model to follow a predetermined drilling plan and control the drill.
Thanks to the expertise imparted by a human expert, a model can drill faster by figuring out the optimal path of drilling.
This ends up being a faster and more accurate process that spares humans of the tedious and high pressure tasks. Similar approaches may apply in other industries such as mining, automotive, and discrete manufacturing.
It’s a Wrap
Overall, machine teaching adds value to the machine learning process. It’s a paradigm that makes AI more accessible to more people as they involve domain experts in the model building process. This allows for the creation of better, more efficient models.
We need to welcome more non-experts in the creation and implementation of AI technology with machine teaching and considering the evolution of AI and technology.
References and Further Reading
L. Holmberg, P. Davidsson, C. M. Olsson, and P. Linde, “Contextual machine teaching,” 2020 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom Workshops), Austin, TX, USA, 2020, pp. 1-6, doi: 10.1109/PerComWorkshops48775.2020.9156132.
L. Holmberg, P. Davidsson and P. Linde, “A Feature Space Focus in Machine Teaching,” 2020 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom Workshops), Austin, TX, USA, 2020, pp. 1-2, doi: 10.1109/PerComWorkshops48775.2020.9156175.
Peer Review Contributions by: Lalithnarayan C
About the authorCollins Ayuya
Collins Ayuya is pursuing his Masters in Computer Science, carrying out academic research in Natural Language Processing. He is a startup founder and is passionate about startups, innovation, new technology, and developing new products. Collins enjoys doing pencil and graphite art and is also a sportsman and gamer.