“The question whether machines can think is as relevant as the question whether submarines can swim.”- Edsger Dijkstra
Artificial Intelligence is the hottest it has ever been and everyone wants to get in on the action. From bots to pop-culture to Google’s Go to Facebook’s image recognition, AI is more than a buzzword or the subject of a scary movie these days. With so many data points in the space, it is worth taking a moment to step back and calibrate our own neural networks as to what is actually happening: to separate the fact from the fiction in the real world right now.
Fiction: Deep Learning is the End-All-Be-All of Machine Learning
Deep Learning is an enormous development in the field of AI and Machine Learning, without question. Google was able to develop a single algorithm that given only pixels and score as its inputs mastered 49 different Atari game scenarios. Deep learning was also a major influence in the AlphaGo program that defeated the greatest Go players in the world.
In response to that feat such hyperbole as “Deep learning is killing every problem in AI” have become commonplace in even the most respected academic journals. It is important, however, to remember that Deep Learning also has its problems, namely speed and transparency.
Deep Learning is a black box. In unsupervised Deep Learning, because we are not training the data injected into the system, it is extremely difficult to parse what is happening behind the scenes. Even in supervised systems, the process is abstracted as it becomes an optimally weighted non-linear combination of layers. This makes it extremely difficult to understand why a decision was made the way it was. This may not seem like a big issue if the AI is controlling your Newsfeed or scheduling your next meeting, but if it is making a medical diagnosis or driving your car it is. “Ultimately, the user should be able to understand the answers to questions like ‘why did you do that,’ ‘when do you fail,’ or ‘when can I trust you.’” -DARPA
Deep Learning is Slower. The learning process goes many layers deeper than with “normal” neural networs. With each layer connected by a non-linear combination of the layers beneath them. This translates to multiple impacts: more data is required for training, the algorithm takes longer to create, it is slower to run, and it requires massive computing power (and thus usually connectivity to the cloud).
Fact: There are use cases that support many different forms of Machine Learning
Even in the subset of supervised learning, there are many different algorithms (boosting, random forests, bagging, SVM’s, etc.) that vary in performance based on their applications.
In general, if you want to optimize for speed from observation to decision, expert systems are usually the fastest. Even Fei Fei Li, the Head of Stanford’s Artificial Intelligence lab, gave up on trying to program computers to understand images and labeled them as a child would. As IoT takes hold of the world, there can be embedded intelligence into devices to make “smart” decisions, but by necessity these devices would need to run light-weight algorithms. Techstars Mobility company Algocian, for example, seeks to turn any camera into a smart camera. By embedding trained algorithms into a Nest security camera, it can restrict the data sent up to the cloud for processing, not only reducing false positives but also lowering computing costs by over $40/camera annually.
That said, for the Googles of the world who have an endless supply of data, top tier engineers, the time to develop and test their algorithms, and data centers to do the computing, Deep Learning can be used to create massive value.
Fiction: General AI is Revolutionizing the World
There is this undertone for every difficult problem these days to machine learn your way out of it. It is not uncommon to hear things like “machine learning’ is a revolution as big as the internet or personal computers” or that “the AI and machine learning revolution is upon us.”
There are many sectors that Artificial Intelligence is being applied to, but each solution is unique. The idea that you can throw an algorithm at a problem and have it solved, is at this point restricted to those 49 Atari games. Algorithms need to be tweaked and retrained to be applied to a new use case, not even Deep Learning is amenable to doing that itself.
At their current state, machine learning algorithms work well so long as the general framework is the same (ie they are computer games with pixels to define locations and scores to optimize), but cannot easily be piped between frameworks…yet.
Fact: Verticalized Artificial Intelligence is making a real impact today
Currently, the most successful business model for machine learning is to take a specific problem and learn the solution to make the workflow seamless. Textio focuses on a specific space: job postings. It then learns optimal language to help improve postings to get better applicants. Second Spectrum took on the task of helping an NBA team by learning from made and missed shot locations over their past two seasons. Using the highly focused training data, they give actionable insight into improving how the team is run.
Uptake has turned itself into a unicorn by making verticalized AI a repeatable process in the IoT space. Rather than trying to create a machine learning solution that fits everyone, they took a hands-on approach and embedded themselves into each industry and company they work with. As a result they are getting unique data to each use case to train their algorithms. They in turn are building up a massive data network effect while providing value along the way. If general artificial intelligence is in our near future, they will be well positioned to make that transition.
Fiction: Artificial Intelligence is Replacing Humans
Every week another post comes out about how artificial intelligence is going to put everyone out of work. People are scared that even skilled workers will not be immune to this round of automation. AI is great at learning individual activities well, but very few jobs consist of one activity. Humans are flexible, they can easily be retrained, algorithms are specialized and expensive to retrain.
Individual activities, especially highly replicable and optimizable activities, are prime to be taken over by AI. Scheduling a meeting, for instance, requires optimizing a calendar and sending back and forth emails. X.ai is working to make that an automated process, but the assistant is not going to lose his job for it — he will have more time to perform the rest of his tasks and meetings will not accidentally slip through the cracks.
In medical diagnostics, there is a sentiment that doctors could be replaced with automated systems, there are already machine learning techniques for image recognition after-all. There are currently around “35,000 people trained and licensed to annotate all of those observable imaging features.” Each image has thousands of observable features for the machine to examine. The amount of time required to collect and train the algorithm and validate its superiority with statistical significance would be astronomical. Deep Learning would mean operating without being able to answer “why?” which would be unacceptable for health questions.
Fact: Artificial Intelligence is Augmenting Humans
While those doctors do not need to worry about being replaced anytime soon, “computer aided detection and diagnosis of medical images” is already being used in the field, making the doctors more efficient and accurate. In a similar vein, Textio does not pose a threat to the recruiter writing job descriptions, it merely aids her in getting top talent in the door.
The result is people and companies are getting more efficient. Amazon combines data, human talent, and machine learning to run one of the most optimized companies in history. It “has at least 21 data science systems, including several for supply chain optimization, an inventory forecasting system, a sales forecasting system, a profit optimization system, a recommendation engine...These systems are intertwined with one another and with human strategists to create an integrated, well-oiled machine.”
Andrew McAfee said “The Industrial Revolution was when humans overcame the limitations of our muscle power. We’re now in the early stages of doing the same thing to our mental capacity.” AI is augmenting humanities capacity to learn and process information, making us better by letting us optimize to our strengths.
So, is the submarine swimming? Who cares, it’s moving us forward a lot faster than we could on our own, and we’re steering it.
