Duc Pham, Mechanical Engineering, The University of Birmingham
Believe it or not, this article is about innovation. However, before talking about innovation, allow me a diversion.
Readers of this column may remember my interest in intelligent systems, i.e. systems that have a capacity for learning among other attributes. Machine learning is not new, the term machine learning having been coined in a classic article by Arthur Samuel at IBM nearly sixty years ago.[1] The field of machine learning has enjoyed a renaissance in the past decade. It has captivated much popular attention thanks to the spectacular successes of machine learning programmes like IBM’s Watson and Google DeepMind’s AlphaGo. Machine learning is also one of the main technologies underpinning Big Data analytics and Industry 4.0.
Machine learning is conceptually quite simple. There are three main categories of machine learning modes: supervised, unsupervised and reinforcement learning. Supervised learning is usually associated with learning a classification or regression task. With supervised learning, a teacher provides example inputs to the learner (a computer) and tells it what those inputs represent or what outputs it should produce in response to them. Once it has been trained, the computer will be able to classify new inputs (i.e. those it was not previously taught to classify) or generate appropriate outputs in response to the given inputs.
Unsupervised learning is a clustering technique. The teacher again provides example inputs to the computer but, this time, does not tell it what those examples represent or what outputs it should produce in response to those inputs. The computer has to discover any structure in the inputs by itself and group them into clusters with similar characteristics. At the end of training, the computer will be able to place a new input into the correct cluster based on how close the input is to the items in that cluster. Because the teacher only provides raw examples and does not teach the computer anything, unsupervised learning could be regarded as learning without a teacher.
Reinforcement learning, which is inspired by behaviourist psychology, usually involves interaction between the computer and a dynamic environment. The computer has to execute some task in that environment – for instance, steering a car or playing chess – and learns through receiving feedback from the teacher. That feedback is a reward if the task was performed well or a penalty if it was badly done. The teacher does not tell the computer which action to take. Instead, it must discover for itself, mainly through trial and error, which action will yield the maximum reward.
The aforementioned three modes of machine learning closely reflect how humans learn. For instance, children learn to recognise the letters of the alphabet by being shown the different letters and told what they are. Botanists learn to categorise plants into species by examining many examples of plants and grouping those with similar characteristics together. A baby learns to walk by discovering the movements enabling it to advance towards a goal without falling; cheering by its parents serves as the reward for walking straight while the pain due to a fall is the penalty for an incorrect set of movements.
Now, let us return to the theme of this article. We have seen that people learn from either being told, or discovering for themselves, what is right and what is wrong, or what works and what does not. For those of us interested in learning to innovate, museums that display historical examples of innovations can provide a rich learning resource. While most museums tend to celebrate successes, I recommend a visit to the Museum of Failure in Helsingborg, Sweden (https://www.museumoffailure.se/). There, visitors will find an eclectic range of innovative flops, from branded lasagne and “pens for her” to game consoles and Twitter-only mobile devices. In the words of the museum curator, “Innovation requires failure. Learning is the only process that turns failure into success.”
http://www.birmingham.ac.uk/schools/engineering/mechanical-engineering/index.aspx
[1] Some Studies in Machine Learning Using the Game of Checkers, IBM Journal, Vol. 3, No.3. July 1959.