How to transform data into knowledge
27 February 2018 | Written by Filippo Scorza
The human brain and artificial intelligence are much more similar than what you might think: both, in fact, use algorithms to process data. From videogames to AlphaGo, here’s how the machines are learning to think.
In the spring of 2000, two crazy and visionary scientists at MIT tried to associate the connections coming from the retina of a ferret not with the corresponding visual cortex, but with the auditory one, and vice versa.
Practically, the eyes of the small rodent were connected to the auditory cortex and the ears to the visual cortex.
Seems like Frankenstein stuff, you’ll say …
That ferret was “rewired” and, for this reason, it would have gone against an almost infinite series of difficulties and disabilities.
But that didn’t happen.
What happened, surprising the researchers, was a spontaneous re-mapping of the brain areas of the ferret: a few weeks later, in fact, the animal was able to see and hear again. Thanks to cerebral cortex’s ability to remodel itsefl, the signals coming from the two organs of sight and hearing were functioning again.
Ben’s story is very similar.
Ben was a blind child from birth, who independently developed an eco-localization system: just like bats or dolphins, Ben could hear the echo of the sounds emitted to identify a close obstacle. Snapping his tongue and listening to the return echo, Ben managed to move independently within unknown environments and even skateboarding!
All this to underline how our brain always uses the same algorithm to learn and create knowledge, regardless of the type of data it receives.
The large amount of data that we process everyday, after all, is nothing but a set of electrical signals emitted by our neurons and exchanged within multiple nodes through synapses: the learning mechanism can not be developed except through a repetitive “strengthening” method.
Repetitiveness and iterative research are, moreover, the two key components of natural evolution, which constantly tries to solve a particular problem through attempts: it’s a sort of process similar to “learning by doing”, which allows us to obtain a result through cyclic experimentation.
The same approach applies when we talk about machine learning: algorithms able to learn through the experience and through the processing of large amounts of data. In its early days, machine learning represented, according to Turing, the path that would lead to the development of computers with an intelligence similar to the human being, able to learn independently: it would be enough to have plenty of data to “educate them” or “to train them”.
After all, to develop an artificial intelligence, we need only logical gates: thousands of NOR (the NOR operator, the negation of the result of the OR operation, returns 1 if and only if all the elements are 0, while it returns 0 in all the other cases) to be used for construct microprocessors that are increasingly performing and capable of processing huge amounts of data in a relatively low time. The videogame market, thanks to the “gamers”, has allowed and stimulated the acceleration of this hardware production process: the request of increasingly complex scenarios, more realistic render and real time processing for the challenges on the net, led to the need developing processors dedicated to computing such data (matrices).
That’s the beginning point from where the road was opened for all the modern algorithms dedicated to “deep learning” which, until a few decades ago, were not able yet to produce reliable and valid predictions.
AlphaGo, software developed by Google Deep Mind in 2014, is probably the key example of this whole process linked to the evolution of “learning machines”.
How did AlphaGo learn to play Go?
Through training and supervised learning, based on the analysis of human players: the software has learned by “watching” humans play Go and by tring to imitate them, basing his decisions on a database of about thirty million moves.
Giving “life” and ability to learn to these machines, have we opened Pandora’s box?
Do we have to be scared?
Or should we rejoice at the new possibility and exploit these algorithms to analyze our genes, discover any diseases, and develop predictions about possible side effects of the investigational drugs?
The only certain consideration is that the challenge of the future is no longer just technology and information technology, but rather social and ethic.