Predicting a year's worth of trends in a field as explosive as artificial intelligence is a bit like guessing what a human would do if you gave them every superpower all at once. Would they heat up their latte using eye lasers or rob a bank with a cucumber? Sometimes machine learning projects can feel a bit like that; where shall we point this crafty set of algorithms and monster compute power? The adage was to 'predict the future by creating it' whereas now we warehouse the past (data), chose the appropriate algorithm in the present (AI) and observe a black box output in the not-so-distant future. It's fitting that we're in a place where predictions are possible which is back where we started. The first documentation of the 'predict the future' concept was in Dennis Gabor's "Inventing the Future" where he actually said:
Rational thinking, even assisted by any conceivable electronic computers, cannot predict the future. All it can do is to map out the probability space as it appears at the present and which will be different tomorrow.
Often the genesis of an idea holds the most truth because the human brain (recommender engine) is quite good when we ignore our internal editor. Now, let's take a look at the current state of machine learning, ingest some data and see what we can predict:
1. Machine learning will keep learning and will teach us more than we thought we would ever know. As a plethora of industries that never considered AI start to apply ML (machine learning) models to their business, the algorithms will become better refined and the early adopters will see exorbitant cost savings. It is already starting with consumer education and customer service, where historical data applied to the right algorithm is teaching us valuable lessons about FAQs, so consumers can spend less time waiting and more time doing. This method is drastically decreasing customer churn and increasing cost savings by ten times. Machine learning models will continue to tackle repetitive and structured tasks to add efficiency into hybrid human/AI systems based on solid strategies.
2. The hype of AI will fizzle out and the real work of problem solving will begin. The mainstream's first instinct was to allow their imaginations to run wild with the promise of 'new intelligence accessible by all'. Now that expectations have been managed (yes, sex with robots is Westworld fiction...), we can get down to the real business of bettering business with the help of augmented intelligence. Machine learning is not a panacea, but it can be a powerful tool when matched to the right screw. Augmented intelligence is the AI of the hour, because a hybrid approach that couples machine learning and human input is driving initial results for new adopters. Technically, it's supervised learning and it's the first step before you start working with unsupervised learning or even ensemble models for that matter. We are in an exciting new position (no euphemism intended) where new minds are getting excited about AI and trying to figure out how they can find a use case for their daily activities. We partly have the hype machine froth to thank for this increased interest and, if you can only avoid the blatant 'AI Washing' that is occurring, we will see exponential growth; focused growth.
3. Chatbots will learn to know their place in the mighty hall of machine learning as mechanisms for data collection. Developers will stop building boring text-based decision trees and start considering the point. The disappointing initial launch of bots that lacked understanding will step aside as builders lick their wounds and iterate. Just because you can do something, doesn't mean you should. Developers have been desperately trying to stuff specific conversation threads of knowledge into bot journeys, but they forgot they were talking to humans. Humans have two opposable thumbs (often) and love nothing more than challenging another's intellect, especially an overly-confident chatbot. When our company IV.AI built The Red Queen for Sony Pictures, we used AI at the beginning to build character traits and a package of animation using image recognition to make the journey visually stimulating. You won't be smarter than a human who wants to prove you wrong, so do not try. Use other strings in your bow and other realities of existence such as time to wrangle your audience and delight their senses.
4. We will start to see job losses due to the efficiencies that AI drives. Budweiser shipped 50,000 cans of beer over 120 miles in an autonomous truck. Everyday we are seeing real impact with our customer service clients. Significant impact where AI is driving 12X efficiency with some models, equating to 40% headcount savings. Where do these people go for work?
We should expect to see job losses for hard skills where machines add efficiencies and therefore diminish the demand for a human brain to manage the task. The practical impact we've seen often registers as positive with staff that appreciate skipping structured mundane tasks to focus on personable intuitive interaction with customers, but the job satisfaction and rate of jobs created will not equal the losses and steps need to be taken. The Obama White House released a report on the state of the industry that concluded that AI will ultimately lead to higher average wages and fewer work hours, thus supporting the benefit of AI in the workforce.
5. Reinforcement learning will find it's feet. Every machine-learning geek in the world was shaking with excitement as AlphaGo beat Lee Sedol, one of the leading Go players. Reinforcement learning is an A/B test on Adderall. This allows a machine to learn without examples (how-to) or instruction. Imagine your first kiss; a reinforcement learning-enabled set of computer lips could figure out how to navigate a smooch by trial and error with the goal of achieving a positive outcome: the perfect kiss. This approach coupled with deep neural networks allows us to tackle complex problems. How complex?
Let's look at a number. John Tromp, one of the creators of the Tromp-Taylors Go rules, computed that the exact number of legal positions in a 19x19 game is 208 168 199 381 979 984 699 478 633 344 862 770 286 522 453 884 530 548 425 639 456 820 927 419 612 738 015 378 525 648 451 698 519 643 907 259 916 015 628 128 546 089 888 314 427 129 715 319 317 557 736 620 397 247 064 840 935 and Lee Sedol still won one of his four games with the machine, which goes to show that even when we get it right, we still have room to grow.
And robots! I didn't even mention robots. How silly. But, we're all out of time. Machine learning is entering a great summer after many AI winters and this time, it holds more promise than ever as more people pour their beautiful ideas into the space.