AI: What Happened and What’s Next
We just finished a banner year for practical machine learning and artificial intelligence. And as we start a new calendar, it seems an opportune time to review what has happened in the last twelve months, and to also look ahead and think about what is next.
In 2016, voice-based assistants really took off. The Echo (and other Alexa-powered devices) are now estimated to be in tens of millions of homes, and all of the major technology players have made substantial capital and personnel investments in speech and language services. These include Apple’s Siri, Microsoft’s Cortana, and Google’s Now. And with the launch of Home, Google is aiming to put a physical presence in the house, which had been the sole domain of Amazon and the Echo for many more months than one might have expected. Recognized as a major opportunity, nearly everyone wants to be in the business of voice so that they can play host to the next ecosystem and platform for consumer integration.
Chatbots have also continued to proliferate, with many more companies launched (or having released new products) that will have machines communicate directly with customers. Some examples include GrowthBot (by Hubspot), Lemonade, and Talla. A few of these bots have been quite successful, and they are becoming more prevalent for use in customer service and marketing. However, most have been unfortunately somewhat disappointing and poorly designed. For example, Microsoft endured a rather disastrous release of their twitter bot, named Tay, which was attacked and nefariously trained by internet misfits.
Deep learning moved from downstage-left to upstage-center. TensorFlow has exploded in popularity and is becoming the dominant infrastructure for artificial neural networks. Meanwhile, Spark, the extremely popular framework for distributed computation, released v2.0 and has widely expanded mllib. Autonomous driving continued to make significant gains toward road-readiness; and near the end of the year, Amazon quietly released Lex, an AWS-hosted suite of services for speech and language processing. This gambit, in direct competition with Azure ML and Cloud Machine Learning, will ramp up the intensity of the ongoing machine learning as a service (MLaaS) land war.
Despite all this success (and more unmentioned), there is still much yet to come.
Despite expanding in number at an impressive clip, there are still many chatbots that operate at a substandard level of quality. Errors and misses abound, and users can be easily frustrated with what they’re given. However, there are some big (and some quiet) successes in the wild today (you may be surprised at how many customer service accounts on Twitter are not supported by humans), and the depth and capability of these bots are continuously improving. I suspect that 2017 will see even greater growth as compared to 2016, and will come with an increased average level of quality and customer satisfaction.
I expect voice-based assistants, like the Echo and the Home, to continue to expand in the rate of adoption and in capability at a very rapid pace. Voice-based conversational interfaces are one of the biggest ongoing land wars in machine intelligence, and someone is going to win and dominate, much like Apple did with the iPhone and the app store. You should expect major third-party integrations to be announced, and a wide array of native offerings to be released from both Google and Amazon. I suspect that at least one other major technology player will release a standalone in-home device this year (the rumor: Microsoft). And I expect that eventually, and possibly this year, someone is going to make a dedicated play for the commercial world, putting specialized voice-based systems in our stores and our offices.
While getting smarter, autonomous cars still have a lot of technical challenges to overcome before they can handle unconstrained and wide use. However, this is probably the biggest ongoing land war in machine intelligence (above voice-based assistants) and there is almost certainly going to be a road-ready product ready for open testing sometime in 2017. You won’t be able to buy it, of course, and it will be aggressively and closely monitored. But in various cities around America, there will be cars without drivers. nuTonomy will start road testing very soon in South Boston; Uber is planning a self-driving test fleet in Pittsburgh later this year; and Toyota is already deploying cars onto the streets of Cambridge. Coupled with the biggest opportunity for economic disruption, this will be the key area on which to keep a close eye.
Azure ML was one of the first serious attempts at commoditizing machine learning, and it didn’t go spectacularly well. The tools offered were not sufficiently advanced to appeal to the power user, and were not sufficiently simple to appeal to the non-engineer. However, you can expect those offerings to steadily improve, especially now that Amazon, along with Lex, has officially entered the field of MLaaS (machine learning as a service). Google Cloud made a tepid foray last year with Cloud ML, but the scope of their offering will almost certainly expand, and rather quickly. By the end of 2017, I expect all of the major cloud providers to substantially expand the quality and depth of their MLaaS products, and prices will race to the bottom.
Amazon recently announced the availability of FPGA instances in AWS EC2. Fast becoming a new favorite for their flexibility, low power consumption, and capability to (when engineered correctly) very quickly process data, this integrated circuit is the latest in a long line of compute diversity from cloud providers. Hardware is nearly always getting cheaper and more accessible, which has been a key driver behind the machine learning renaissance froofm the last few years. Cloud solutions will continue to expand their offerings (and get even cheaper), and it will become increasingly fast and affordable, allowing anyone to play.
And why not finish with some wild speculation? I think quantum computing will have a major breakthrough in 2017, and materially impact the machine intelligence landscape. Google has been pouring money into quantum computing research, and they have unveiled plans for a so-called “quantum supremacy.” A major breakthrough in quantum computing would most certainly have a profound impact on the early research around quantum neural networks, which would produce amazing efficiency gains over alternative arrangements, and further accelerate the progress of the fast-moving field of deep learning.