April 23, 2018 § Leave a comment
Göte Nyman, University of Helsinki, Finland
Ossi Kuittinen, SimAnalytics Ltd Helsinki, Finland
Machine Learning (ML) as part of AI is now familiar to us all, from managers and workers to engineers and designers. We know the secret of ML: it’s amazing potential to learn from data, huge amounts of it, which sometimes is a guided process but can also proceed without significant human intervention during the learning process. However, human contribution and collaboration remain necessary for any ML system design and implementation and especially in securing optimal ML system learning and its continuous improvement.
(Creative Commons licence image)
What is the process where skilled human professionals and a ML system join their competences for a learning journey? How should we characterize the teaching, learning and interaction, where the ML system, human designers, engineers and workers act and learn together to come up with the best possible ML solution to the target environment?
With the fast improving ML we are suddenly confronted with a new kind of a teaching-learning-collaboration challenge or paradigm, where the competences, roles, learning processes, and responsibilities of the participating people must find their effective form and content, together with the ML system.
One could think that human-technology problems are nothing new and that the introduction of ML is just one phase like any other major leap in technology. But it may come as a surprise that we face a new kind of an interactive learning paradigm which changes fast. We call the work with ML as ‘triad of learning and collaboration’, consisting of at least the following components/actors (We use the concept of ‘triad’ in a slightly transformed form as borrowed from the learning approach where ‘the individual’, ‘the learning community’, and ‘the authentic use of the objects’ are separated entities of the learning situation; Paavola & Hakkarainen, 2005). A ML system that learns, interacts and teaches is a complex creature, indeed and this makes all the difference:
- The ML system as a whole – its sw, hw, system integration, computational models used, interfacing and its UI
- Operator-professionals specialized in the target domain and facing the planning, design and implementation of the ML system
- Designers and engineers responsible for the ML system design, implementation, process quality, and its continuous tuning.
What is special in this triad, is that all three ‘partners’ interact, teach and learn from each other, from the beginning of the ML project, often simultaneously. Furthermore, while there are a plethora of theories and models of learning, including computer-assisted learning, for example, the theory of very intelligent human-ML-system, interactive learning – where a human teaches a machine and the machine teaches humans – does not exists. It is time to launch its development so that it can keep up with the pace of fast advancing machine learning. There is already an intense discussion on the possibility to look inside a working deep neural network, one of the ML application, to improve its interaction with its operators but this only starting. Furthermore, it is highly likely that due to theoretical and technological advances, the current AI models will undergo major changes within the next decade or so.
We have summarized the triad learning relationships in a simplified form in the figure, where the instances of ML-human learning occur: 1. between the User/Operator and the ML system, and 2. between the Designer/Engineer and the ML system. The two instances are very different in nature and require relevant organizing and management, because of the differences in the human competences involved. Both learning processes can occur over the whole development and maintenance cycle of the ML system, with different weights over time and they are of utmost value to the success of the ML implementation.
Figure. A simplified scheme of the learning/teaching relationships and some of the learning contents in the ML implementation.
Design and introduction of the ML system
The design and project plan for the implementation of the ML can best be described as co-design (some form of it, from weak to strong). The designers/engineers introduce the candidate ML model for the management of the target system – an organization, industrial plant or other relatively large-scale enterprise where the benefits of ML are expected to cover the costs. This is a complex learning challenge requiring the participation of the designers and relevant specialists who know and work with the target system and know the processes and who will then work with the ML-system. It is dangerous to apply a linear and fast-to-the-implementation approach in such a complex endeavor.
While the mathematical and technological features can be too complex for all users/operators to follow, it is necessary to introduce the ML in a way that its basic learning and its teaching characteristics can be understood and that a fruitful interaction in co-design becomes possible. In other words, a valid, ML model-in-practice must be introduced to the participant operators and other personnel. Usually there is no unique or single ML model available for this purpose; hence the introduced model must be formulated and tailored accordingly, to match the competences of the participating personnel and the specific context of ML implementation. At the beginning, it is important to relate the ML concept, its functions and measures as closely to the target system model ‘as-is-model’ as possible so that the participating people can relate the ML system functions and its measures, the metrics used and the UI to the target system with which they are high-class professionals.
During piloting, testing and implementing these new concepts become critical since they make up the language with which the operators and users ‘negotiate the system’ and try to understand how it works. The operators learn a new ‘mental model’ of it by which they know what the ML system can and cannot do, what risks are included, how it can best be taught and how to communicate about it with the designers. These concepts guide the operators in their observations, too: what to follow, what to search for and how to react to certain observations of the ML system state and behavior. This is a most demanding challenge and mistakes made and vague concepts introduced can lead to blind spots in the development project, valuable knowledge of the ML system performance lost, underperformance and the costs becoming high. It is a very serious learning task for both the ML system providers and the personnel acquiring it.
New paradigm of interaction:people teaching ML when ML teaches people
With its computational power the ML system can shoot out practically unlimited amount of data: new measures if its status and dynamics, holistic and specific component state information, and any aspect of its measured performance, fast and slow in nature, depending on the environment. Overflow of information is a real practical risk and possibility and it must be dealt with by means design, and human learning and collaboration. Some of the ML data is totally new to the people specialized in the ‘old’ system having its own, perhaps historical control systems and performance metrics.
Some of the ML data can be significant and even crucial for the system performance, some not, and during the piloting, testing and implementation process the personnel learns to recognize what is relevant information and why, and then guide the designers in representing it in the UI. This data environment then becomes the de facto world where the personnel learns to work – with the (partly or totally) new information.
The ML data must be made accessible for easy observations and its background processes amenable for easy, immediate and relevant control whenever necessary. An explicit interaction loop (measure-control-measure) must be provided which is either automatically controlled by the ML system, and only made visible to the operators or which is amenable for human control. The observation of these (numerous) control loops is an essential part of human-ML learning and it will introduce new tasks, practices and responsibilities to the personnel.
What is new in this triad-entity is that the ML system teaches both the Operators/Users and the Designers/Engineers in various ways, some of which are new. How does this interaction take place and how should this learning be considered and conceptualized as it concerns both the personnel and the ML system? This is a fascinating learning theoretical challenge which cannot be dealt here with now, but we will return to it later, especially after having collected practical experiences again from the work with SimAnalytics Finland.