The challenges with artificial intelligence are not about math.
Every article advising you about what is machine learning, deep learning or artificial intelligence tends to emphasize how much math is needed for building these algorithms. Sure, it is university level math that is required, but when a company looks to apply these technologies the math is never the problem. Many of the really hard problems like natural language processing (NLP) or computer vision (CV) are solved with existing cloud components, so no math is involved. On the other hand, the custom-built algorithms for linear or logistic regression and neural networks are quite straightforward for an expert and require no novel mathematical wizardry from case to case or company to company.
The difficulty is neither with the data. Surprisingly small amounts of data can be used for initial teaching and verification and business case calculation for a new application. Additionally, many problems can utilize reinforced learning where most of the learning happens after the algorithm is put into production.
So what is the challenge then? The challenge is the people and the solution is design.
Remember when the experts realized AI is not going to replace jobs and persons, but to augment them? It’s not robot or human. It’s both. And these are the questions solved today.
On the first abstraction level, we can ask how do we design services that utilize AI? And how can we design a new role for a person, whose job is being augmented by AI? We need to think of ways the user interacts with the machine and how trust in the decision making of each party is conveyed between them. It isn’t enough that an algorithm outputs a recommendation. Depending on who needs to act on that recommendation, a very different kind of communication on the grounds for that decision need to be conveyed. We need to be careful of both over and under relying on the machine. At the concrete level, it’s about showing all the alternatives and not just the best option, showing confidence intervals, spelling out reasoning, using colors, graphs and so on.
But it gets more interesting when we zoom out. Since it’s a learning algorithm, it can also learn to use the best method of communication with the current user. It should also learn how much this particular user needs help. Is it a good idea to present the user with a proposal all the time or only when the confidence level and criticality of the decision are very high. We get an algorithm that learns to automate the optimal amount of the work.
Next, let’s think about how every job has many kinds of activities. We can often apply AI to multiple of these and if each of the algorithms adjusts themselves as described above, we are effectively using a multi-algorithm AI to learn which parts of the current user’s job should be automated with AI. Very meta-level AI and design implementation.
What happens at the company level when individuals and their jobs are thus augmented? The AIs get personalized and each persons’ job gets automated based on their strengths and weaknesses. We get a huge boost in role – person match and the jobs inside the company start to be unique. This will require totally different kind of leadership and HR.
These changes to the way people and organizations work can’t be solved by mathematicians and machine learning experts alone. Solid world-class design is the key ingredient.