Why is machine learning [ML] important for your business? If you work at Nokia, your Chairman can explain it to you in a one hour presentation he developed over six months of research. Risto Siilasmaa helped make his network smarter. Everyone needs to know if ML can help with their business problems, but first they have to understand the basics, says Siilasmaa.
- Digitization has created an explosion of information
- ML is based on models like logistic regression, which can be fairly easy to understand
- ML is fitting the model to the data
- ML is neural networks learning from data sets
- The more high quality data, and computing power, the fewer mistakes ML will make
- In a large neural network you can have 100 million parameters in a single layer
- Flawed outputs can happen if human oversight confirms incorrect ML conclusions (human oversight becomes very important)
- A neural network first learns from a data set (time consuming) and then can be tested against other data sets
- The important work is done by systems of ML systems
- Machines are still getting faster and more tools are being developed
- The data we are helping create (e.g. through use of speech recognition) is feeding AI corporations
- ML can be tricked if you know the underlying algorithms
- Remember: Garbage-in, Garbage-out
- Big question: What data will we need in the future to make better decisions?
- Business and human work is moving to — Low Predictability + High Complexity
- ML can help to experiment faster and better in order to deal with Low Predictability + High Complexity
- The future of work: First experiment … then develop a strategy
In a similar vein, Dave Weinberger says that in a radically unpredictable world, the way forward is to — “Embrace unpredictability and practice unanticipation.”
“If the internet has changed our practical approach to the future, machine learning is providing a conceptual framework for understanding why unanticipation works.
Traditionally, to predict the weather, a model would need to be built that includes the determining factors, such as air temperature and moisture, and their interrelationships. Likewise, to estimate the next quarter’s profits, information about the number of salespeople, the number of leads, marketing costs and so on would be included and connected via formulas.
But machine learning doesn’t start with generalized models. Rather, it builds its models based on oceans of data without any sense of the factors the data represents or how those factors interrelate. It iterates on the data, looking for statistical relationships among them, building a model of connections so numerous and complex that we often cannot understand exactly how a machine learning application comes up with its results.
This lack of explicability raises many important issues about ensuring that machine learning’s outcomes are fair. But the success of machine learning in using models without generalizations is leading us to acknowledge that the future is determined by the unknowable and chaotic interaction of a universe of particulars, each affecting every other simultaneously.” —David Weinberger 2020-02-10
The banking industry sees similar challenges in developing the optimal blend of automation through ML with human skills to create trusted environments.
“Building the necessary trust requires increased awareness and transparency around how the AI is being used, the decisions it makes and the opportunities it brings — this is the essence of ‘responsible AI’ and ‘explainable AI.’ People who understand and can explain AI decisions — for example, how machine learning is used within credit scoring, how the systems were trained and how the process is controlled — are highly prized employees in this environment. Moreover, maintaining diversity among the people who are helping to develop AI programmes is important in ensuring that unconscious biases aren’t built into the outputs.” —PWC Banking CEO Survey 2019 (PDF)
[Update] The Citi GPS report on Technology at Work v 4.0 identifies two areas where machine learning can “accelerate & enhance decision insights” — 1) reduce time to predictions & information retrievals, and 2) reduce time to detection & mitigation.
In the book, Only Humans Need Apply, the authors identify five ways that people can work with machines. They call it ‘stepping’. I have added the current competencies I think are needed for each adaptation.
- Step-up: Directing the machine-augmented world — Trans-disciplinarity
- Step-in: Using machines to augment work — New Media Literacy, Virtual Collaboration, Cognitive Load Management
- Step-aside: Doing work that machines are not suited for — Social Intelligence, Sense-making
- Step narrowly: Specializing narrowly in a field too small for augmentation — Cross-cultural Competency, Design Mindset
- Step forward: Developing new augmentation systems — Novel & Adaptive Thinking, Computational Thinking
ML is just one aspect of how we will have to learn to step with the machines.
More on machine learning — Machine Learning Explained