Artificial Intelligence – Dull delivers.

It’s the boring stuff that works.

• As the big ambitious moonshots crash and burn, it’s the smaller, simpler and deadly dull projects that are likely to see real success in the short to medium-term.
• Everyone wants to hear about how AI will cure cancer or drive all our vehicles, but it is the really boring things like saving money on electricity and basic automation where the money is going to be made.
• DeepMind’s ambitious project to use its AI with the National Health Service in the UK came to very little and IBM had to end sales of Watson for drug discovery due to its low return on investment.
• However, at the same time, DeepMind has had great success in cutting Google’s electricity bill for air conditioning in its data centres and Watson is very good at scanning specific biopsies.
• Neither of these are breakthroughs but they both provide a substantial improvement in efficiency and outcome for the tasks that they are being applied to.
• So why are the big moonshots such as autonomous driving and drug discovery failing?
• RFM research (see here) indicates the single biggest reason is an overreliance on the deep learning technique upon which almost all artificial intelligence being developed today is based.
• Deep learning, at its heart, is a system for statistically separating characteristics of data such that when these characteristics occur again, they can be recognised.
• While deep learning can appear to be clever and intuitive, it has no real cognitive understanding of the task it is being asked to perform.
• This means it can’t think outside of the box meaning that the historical data set upon which it has been trained must be an absolute predictor of what will happen in the future.
• Furthermore, the algorithm needs to be shown every combination and permutation that is possible before it can be relied on with 100% accuracy.
• This means that the ideal task needs to have both a finite and a stable dataset in order for deep learning to work well.
• This is where the problems occur as the data sets for moonshot tasks such as driving cars and drug discovery are neither of these two things which is why they have failed.
• On the contrary, the really boring jobs such as saving money on air conditioning costs, tuning basestation antennas and scanning specific biopsies are ideally suited for deep learning as their scope is very limited and they have a finite and stable data set.
• This is why it is these tasks where the return on investment will be found in the short-term rather than the big glitzy projects which are going to continue to struggle.
• This also highlights the need for new techniques other than deep learning as the limitations of this technique are quickly being reached.
• The bad news is that these look they are years away as some of the greatest minds in this field have been working on this for years with no real success to date.
• Despite this, the hype and funding continue to march on meaning that the 3^rd AI Winter remains very much on the horizon.
• Disappointment, disillusionment and falling investment will be the hallmarks of the 3^rd winter just as they were of the first two.
• Winter is coming.