“Most AI today is about preprogramming a machine. Our way is to program them with an ability to learn for themselves. That’s much more powerful; that’s the way biological systems learn. “We refer to it as an Apollo programme, a Manhattan project…” – Demis Hassabis (Deep Mind CEO and Founder)
75 years ago, some of the smartest men in the world gathered together in the desert and, for better or worse, the world was changed forever. Many in the tech world believe that we are on the edge of a similarly momentous change, but many others are responding with understandable scepticism.
Last year, we were all told that blockchain would change the world, the year before that VR, and before that 3D printing… these things remain incredibly promising technologies, and all of them have notably failed to upend the world and bring the sky down on our heads.
So now as we read about dire predictions of 60% unemployment, killer drones, and dying industries, we can be forgiven for thinking that all this robot stuff is being oversold.
Humanity has met no shortage of revolutions throughout history, and humanity has generally soldiered on through all of them fundamentally unchanged.
This time may be very different…
This series of articles is intended to discuss machine learning specifically in the context of marketing, but we need to start by discussing why this latest revolution might end up being a bit more revolutionary than the rest…
And the best way to start that start is to be clear on what artificial intelligence and machine learning actually are.
AI – Artificial Intelligence
AI is best considered more as an objective than a method. The goal of AI is to produce systems that will respond to stimulus in the same way as a true intelligence might. Sometimes you will hear reference to Strong AI, or AGI (Artificial General Intelligence), which are hypothetical AI systems that would be at least as versatile and powerful as a human brain. But AI systems can be much less complicated than this.
The key is the appearance of intelligence.
If you have played a video game in the past few decades you may well have come across references to a Videogame’s “AI”. This is simply the part of the program that is responsible for making characters within the game react in an appropriate way to the players actions. This is as valid a usage of the term as any others.
Typically in this kind of AI there is no actual “learning process”, rather the developer is providing the software with a series of decision trees, using their own native intelligence to predict scenarios and provide the illusion of intelligent response from the software.
If player jumps forwards, enemy jumps back.
If player punches, enemy blocks.
This can prove surprisingly convincing because computers can run the instructions much faster than humans, and randomness can be coded into the system. A skilled programmer can even create the illusion of the software “learning” in response to a user’s actions, but ultimately the system is relying on hard coded decision trees that will not change without more external interference. In essence the system can “borrow” intelligence from its creator.
Machine Learning however refers to a different approach to creating AI, building a system that learns for itself how to complete a specific goal by iterating through possible solutions to a problem and then adjusting its method according to results.
If player punches, try jumping forwards
The program is still using decision trees, but it is creating and pruning them itself, keeping the ones that further its goal, and losing the ones that reduce its effectiveness. This is a process of trial and error, but it’s a process that a computer can often perform very quickly in real time, especially if you train it against another AI system. After enough repetitions you may very well end up with a system that can beat human competition.
What has changed?
The theory underpinning Machine Learning is by no means new, but actually making use of it is very taxing computationally, to the point that hardware development has only recently advanced to the point that it can produce useful results for many types of problem.
However, having now become viable, machine learning is distinct from previous approaches to AI in that it can produce better solutions than humans would be able to within the scope of a given task. Software is no longer “borrowing” intelligence, but rather learning for itself.
Moreover a successfully trained system can generally perform tasks much faster than humans, demands no breaks or salary, and can be freely copied, providing the potential of applying effectively infinite manpower to a problem. Meanwhile the increasingly digital nature of our society has moved many work tasks into the electronic world, with advances in robotics and vehicle automation extending the plausible reach of a virtual workforce.
All of this has the potential to put those companies developing this technology, traditional tech giants such as IBM, Apple, Amazon, Microsoft and Google (as well as international companies such as Baidu) in a formidable position.
Never in the history of mankind has there existed a single technology with the potential to simultaneously impact on as many different aspects of human society and industry in as short a period of time. Likewise, it is probable that no other key technology has ever been as demanding to research and develop and as reliant on very high levels of expertise to properly apply.
Also new is the tremendous concentration of funds, capability, legal influence, data consolidation, and sheer business muscle of modern tech giants such as Google. Tech companies have already become notorious for stripping entire university departments of machine learning expertise, and with salaries now starting in the six figures range, few smaller entities will find it easy to acquire the skills that they need to even begin duplicating this research.
While these tech powerhouses will certainly make this technology available to others, via the release of software, API’s, and hardware, they will retain a tremendous amount of control over skills, development, and application.
They will be in an unprecedented position to watch how other are using their tools, and identify opportunities within existing industries, and then leverage their technical advantages to gain a stake in those markets. Likewise, they are best placed to identify and engage new business models that would have been impossible without unfettered access to an infinitely replicable, tireless, undemanding, and potentially superhuman workforce.
In the above quote, Hassabis is talking as much about the quality of the people involved in the Deep Mind project as its impact, nevertheless there is not a shred of hyperbole in his choice of metaphor, capitalism is about to go nuclear, and the superpowers are already delineating their borders.
With the computers themselves doing much of the work, the focus of machine learning as a discipline is in studying the most effective processes for computer programs to learn, ways for them to apply learnings from previous problems to new ones, tying processes together to engage complicated tasks, and ensuring that programs are placed properly on-task. But most significantly for marketing and media applications, they need to ensure that it is provided with properly categorised training data.
If you want to train a computer to, for example, describe what is happening in a video of people on a busy street, you need an immense amount of footage representing that street in as many different contexts as possible (camera angle, time of day, city, location, season of year etc. etc.), and all this existing data needs to be accurately annotated to allow the system to begin to understand what it is looking at.
Data has been likened to the “oil” of the twenty first century, and machine learning is a large part of the reason why. Provide a machine learning system with enough training data and it can find a solution to almost any problem. Without enough data, it will likely produce solutions that are fatally flawed, possibly in a way that won’t be immediately obvious. For example, the “unbeatable” computer game system came unstuck when one of the human players discovered that simply crouching in the corner of the screen would force the computer to fall to its death. If train your video viewing robot solely on summer street views, the end product is likely to work exactly until the point that it snows….
At present comparing machine learning systems to humans is like comparing a diesel water pump to a human with a hand crank. It might be able to shift more litres of water per minute, but it’s going to struggle to change a diaper or open a tin of tuna to feed the cat, and it’s going to take a heck of a lot of data to change that…
As many a feline, or new Mom will appreciate though, unlike with a human, you only need to teach a computer to change a diaper or open a can once.
We will talk in more depth about the implications of all of this in the final article of this series, which will discuss some of the risks and limitations of applying machine learning. But before that we are going to take a more detailed look at near term applications for machine learning within the marketing world.