There is Now a Formula for Machine Intelligence Innovation
As Shakespeare understood more than four centuries ago, much of human identity, thought and culture has been rooted in the idea that we are unique. Although novelists, film-makers and technologists have often imagined advanced inter-galactic civilizations and super-capable futuristic machines, such thinking has always been the realm of science fiction.
What a piece of work is a man! How noble in reason, how infinite in faculty! In form and moving how express and admirable! In action how like an angel! In apprehension how like a god! The beauty of the world …
– Hamlet II. ii
But the belief in our uniqueness is now becoming its own fiction. While Human Intelligence (HI) is wondrous and remains deeply mysterious, it is not magic. Similar capabilities exist all around us, and many skills once seen as exclusively human can now be done better by machines that are infinitely replicable. There is also mounting evidence that life on other planets is even more likely than the sheer size of the universe always suggested.
These realizations are expanding our understanding of intelligence. While this paper will focus on Machine Intelligence (MI), society’s understanding of Animal Intelligence (for which we use the abbreviation AI) is also improving rapidly. Of course, we have always known that animals often see, smell, hear and navigate far better than we do, but we now know that they also have complex memories, emotions, identities, languages, problem-solving skills and social relationships. Jennifer Ackerman’s recent book, The Genius of Birds, makes these points about birds, but similar evidence has been amassed for wolves, whales, dolphins, octopuses, apes, elephants and many other species, especially by Frans de Waal.
From an MI perspective, animal capabilities are also not magic, and science is beginning to figure out how these talents might someday be engineered, greatly expanding our sense of the extraordinary skills that robots and other machines could eventually have. Imagine a machine with the night vision of an owl, or the scent-tracking of a hound. Looking ahead, we will increasingly need to think about intelligence in an integrated HI, AI and MI context.
(Terminology note: we prefer the term ‘machine intelligence’ to ‘artificial intelligence’ because there is nothing artificial about it. For example, we don’t refer to industrial machines as ‘artificial strength’. Using AI to mean ‘animal intelligence’ provides an additional useful twist.)
But why all the fuss about machine intelligence right now?
Sceptics might concede that all of the above is theoretically true, but they would also point out that there have been many previous predictions – going at least as far back as the 1960s – of an imminent machine intelligence era, which look silly in retrospect. And since there are still significant MI obstacles that must be overcome in basic areas such as Natural Language Processing, it is fair to ask why we think that this time will be different. Based on our research, we believe that, although the tipping point has not yet been reached, the code for MI innovation has been cracked. It has three main components:
1) Big Data. It has now been convincingly demonstrated that large unstructured data sets can be used to develop powerful machine intelligence capabilities, without specific subject matter expertise, or even human intervention. Many of the most important MI initiatives today – such as language translation and image, facial, activity and emotion recognition – are based on predictive analytics that get more accurate as the data behind them gets richer, and the internet is making larger and more relevant data sets more available than ever before. In these MI applications, human subject matter experts such as professional translators and psychologists are not only not necessary, they often get in the way of purely algorithmic approaches.
An influential example of this has been the image recognition project, ImageNet, which has created a database of some 14 million labelled images that can now be used to train machines to recognize just about any thing. In 2015, Microsoft demonstrated how deep learning (see below) could be used to enable computers to recognize these images as well as or better than humans. Similarly, Facebook enjoys a huge head start in facial recognition because it can already match our names and faces, just as Google has important advantages in machine translation because it has aggregated the best set of multilingual documents.
Looking ahead, new and established MI companies will use millions of internet images, videos and podcasts of people smiling, laughing, frowning, talking, arguing, holding hands, walking, playing football and so on as the basis for unprecedented Emotion and Activity Recognition capabilities. MI is now clearly among the most important Big Data applications.
2) Software and hardware advances. For decades, machine intelligence researchers have predicted that neural networks and parallel processing would be important MI development tools because they more closely resemble the way the human brain works and because they enable machines to learn. However, until the last few years, progress in both areas was slower than in many computer science fields. Happily, this is now changing, with the emergence of new software and hardware architectures that are particularly good in MI applications.
While deep learning is one of today’s hottest IT buzzwords, its meaning is often poorly understood. Deep learning is basically the latest generation of neural network design. It is called deep because there are more layers of processing than in the past. Although this is a highly technical and mathematical field, the basic idea is that the use of additional layers of abstraction enables tasks to be broken down more finely, and this enables a greater capacity for detailed analysis and self-improvement. This multi-layer approach was used in the recent triumph of Google’s AlphaGo program over Lee Sedol, one of the world’s best Go players.
Both neural networks and deep learning are computation-intensive, and real MI applications can overwhelm traditional systems. Fortunately, new hardware designs have emerged at both the individual system and cloud level. Many MI developers now use hardware that includes Nvidia’s GPUs (Graphical Processing Units) which can greatly accelerate neural processing speeds. And when even more computational capacity is required, the almost unlimited cloud resources of Amazon, Microsoft, Google and others are available at affordable prices. Taken together, deep learning software and parallel processing hardware now provide a powerful MI platform.
3) Cloud business models. The ability to leverage Big Data and the availability of much more capable hardware and software mark major steps forward in the MI journey. But as important as these computer science advances are, the emergence of powerful MI business models is arguably the single biggest reason that the MI field is so energized today.
We are essentially seeing the merger of machine intelligence with cloud economics. This merger will prove fundamental to the innovations of the future, but it is still not sufficiently recognized. Before the cloud, most AI work was isolated and relatively high cost. However, as shown in the figure below, MI advances can now take advantage of the full panoply of cloud capability – including 24x7 availability, rapid global deployment, variable costs, continuous improvement, real-time data, effectively zero marginal cost, easy integration with supporting web services, venture capital funding, and winner-take-all market tendencies.