First Steps to a Machine Brain

The race to Artificial Intelligence is a grueling, sweaty marathon that human beings have been running for over 80 years. We’ve faced steep climb after steep climb, the gradients cranking higher every time we run into a new funding gap or technological roadblock. Recent advances in deep learning and neural networks are leading some to wonder if we might, finally, be approaching the finishing line.

We were all impressed when Google’s DeepMind computer beat a human at the game of Go. This was the first breakthrough of artificial neural networks that captured popular imagination. Despite its simple rules, the popular Chinese board game is mind-bogglingly complex, and possesses more possibilities than the total number of atoms in the visible universe. Mastery of Go is supposed to be the ultimate expression of human intuition — so when world champion Lee Sedol was beaten by a computer program, Google’s groundbreaking AI made headlines worldwide.

And from predicting the outcome of the US election to rushing an incapacitated driver with a blood clot to hospital, the achievements of artificial intelligence continue to excite.

But really, if our finishing-line goal is a general artificial intelligence that can solve humankind’s biggest problems (climate change, inequality, resource depletion) then current techniques are running with their legs tied. Neural networks are inspired by the human brain but in no way think like a human brain. In truth these are largely deterministic systems, which means that the same sequence of inputs will result in the exact same output, each and every time. Their edge over regular, rules-based software is that they can be made to learn the rules using inputs with known outputs. This evolution of software is truly powerful, but is it enough?

We are right to begin dreaming about the possibilities such feats arouse and it is true we are living in exciting times. But if we are to make the first steps to a machine brain, we must be wary not to commit the first sin of AI research: the mistaking of complex behavior for intelligence.

Humor, Creativity and Wisdom

Consider the female digger wasp. Before she brings food into her burrow, she drops it off at the entrance and goes inside to check for intruders. Only once satisfied her nest is safe will she bring in the food. If, while the wasp is busy securing her burrow, a meddling human moves the food a couple of inches, the wasp will move the food back to its original drop-off point and repeat her burrow-check all over again. Because the wasp is not capable of remembering she just checked the nest, she can be made to repeat this cycle of behavior more than 40 times. Complex and premeditated it may be — but her behavior is not intelligent.

Learning from large-but-limited datasets, most commercial applications of artificial intelligence exhibit cognitive shortcomings similar to the wasp. You might ask why this is important: Is it not enough that we have the complex behavior? If a computer can automate a human task, does it really matter if that computer doesn’t exhibit real human intelligence?

Well, yes. Eminent AI thinker Jack Copeland contrasts the behavior of the digger wasp — and by extension most 21st century AI — with that produced by human intelligence. In the same situation, a human would probably wonder who has been messing about with their food — but they would not repeat the ritual of checking their home was free from invaders. We know we don’t need to do this again because our intelligence has naturally evolved to enable not only experiential learning but also the extrapolating of those experiences to new situations. It is obvious the burrow is still safe — without a moment’s thought, we would adapt our behavior.

Copeland literally defines intelligence as the ability to adapt one’s behavior to fit new circumstances, and this is what made AlphaGo so special; through deep learning, it could extrapolate its experience of playing Go to new in-game situations. Ask it to attend to any other task outside of the game, even one as simple as booking the flights for your next holiday, and it would not be able to do so. It must be equipped with tools, connected with the appropriate interfaces, trained again from scratch and possibly rewritten in large parts to do anything in the non-Go world. Google did announce another software to do some of these things, but that, too, is very specialized.

Robert Epstein, a research psychologist points out that the human brain does not store facts or data, like a computer does; it continuously edits itself and relives experiences, like an artist painting a landscape. Our unique combination of experiential learning, ability to abstract concepts and deft extrapolation leads to seemingly-incomprehensible phenomena like humor, creativity and wisdom.

If we are to build a truly intelligent artificial intelligence, we need to build something like a machine brain, and to build a machine brain we need to start being more ambitious.

The Incredible Wisdom of Human Thought Process

Trying to imitate human thought process is intoxicating.

When we’re building artificially intelligence natural language systems, a key Coseer design principle is to encapsulate every point of data in the form of an idea — concepts, rather than keywords. This is still nothing close to the processing power of a human brain, but even a relatively simple emulation leads to incredible results: Accuracy shoots up. Latencies and training times collapse.

All this with a relatively simple emulation of human thought process. Word2Vec, ELMo, BERT, GPT — all popular natural language processing models — adopt the same philosophy to translate each word into an abstract representation of ideas.

Designing algorithms to mimic human thought processes makes business sense for real world problems even today: finding actionable stock market insights from three million documents everyday, assisting doctors through oncological pathways, seamless knowledge management are just some examples. AI is already helping largest and most innovative organizations across sectors. I know this first-hand.

First Steps to a Machine Brain

To continue in our journey towards an artificial general intelligence, let’s ask two simple questions:

• First, what happens when all our applications can learn not only from the data they see, but from the experiences of other applications, and also from the data each application sees?
• Second, what happens when each application (and all of them collectively) can learn from the repository of collective human knowledge, available in the form the Internet?

To achieve any success on these two questions, a language must emerge that can translate insights from every context to a common set of binary code. Current approach to neural networks is woefully short on this. Each model must cater to a specific problem only, and with very limited data sets.

Sure, humans also have specialists. Doctors and military generals are trained differently. However, the differences are at a level too abstract for a neural network to ever achieve. Both may have the same political views, sense of duty, familial commitment or investment acumen. Even at basic level, both may be excellent drivers, adept conversationalists, and masters at Go. One would need a different neural network to even scratch the surface on each of these.

The long-term goal is bold: such a development will set technology free to pursue true intelligence. It will let AI apps benefit from the knowledge that already exists. Applications will ponder on their own time to bring more accurate decisions that the applications are more confident about. And each time that happens, every other application will become smarter.

And that would just be a baby step towards a true machine brain.