The computer press is just full of it. Each week, AI can do something more. It's the fastest growing branch of computing. Google says that, in the future, it won't be a search engine company. It will be an AI company. China is aiming to lead the world in AI and it investing billions into it.
What is AI, how can you use it, and what should you do to become one of those companies that gets into the press because it's using AI to great effect, rather than because you are being beaten by AI-based competition.
Let's look first at what AI is. There are many definitions of AI. However, the one is from McKinsey breaks AI down into…
Computers are taught or self learn how to recognize things. For example, animals in pictures taken by conservation groups, the lanes on a freeway in a self-driving car, the "digital finger prints" and user behaviour of someone trying to hack a bank, a computer disk that's likely to break down in the next 16 hours, possible cancer on a 3D image of a lung, a student failing to reach their potential from their “digital footprint” and old man possibly falling over in their home.
Wow – lots and lots and lot of uses for machine learning, once the machine has figured out how to do the recognition – the “training”. Initial training is hard work, both for the humans who prepare the training data and the computer that's trying to create the models from the training data. A lot of research is going into making machine learning's training phase easier. For example, it takes hundreds of thousands of examples to train a machine learning system what a bird that can fly is. Us humans can do it much, much more quickly. What do we do that machine learning systems don't?
And once the machine learning has done its initial training, it uses ongoing data to update its models. For example, a system that predicts the failure of sewage pumps might be successful in those predictions 92% of the time. But in 8% of cases, it either misses a failure or predicts a failure when none is imminent (false negatives and false positives). What is it about the machine learning model that causes these false predictions - these anomalies? Hopefully the system can learn and get better. By itself.
Scary? Elon Musk and Stephen Hawking certainly think so. If I remember my teenage reading correctly, the book "I, Robot" by Isaac Asimov also explores this concept. If we give a machine learning system a goal and lot of power, it might come to the unfortunate decision that humans get in the way of achieving its goals, and that if it could get rid of those humans, it would reach its goal faster. Hence Isaac Asimov's Robot laws …
Back to reality. So, machine learning systems are split. There is the inference part - the part that takes in the data and figures out what's going on. And there is the learning part. There could, and probably will, be many inference parts, often "at the edge" (I've written about this before in this blog post). There will probably be just one learning part at the "core".
In order for the learning part to continue to update its model, it needs to get "what happened" data from the inference parts - the "edges". The amount of data that the edges receive can be very high indeed. In order to keep data transmission times and costs reasonable, researchers are looking at ways of limiting the edge-to-core traffic. For example, HPE has created a "distance deep learning framework" that where edge-to-core chatter is limited to those situations where the edge inference engine is unsure. For example, if it can recognise a car with 90% confidence, then there will be no communication with the core. On the other hand, if it has only a 20% confidence that it is seeing a car, it will ask the core to train on the data that it is seeing. The core will then send an updated inference model to the edge.
Machine learning is by far the largest use of AI today, and will probably be so in the future.
