Subscribe to our Newsletter
Be the first to know about every G4A opportunity!
Sign up for our newsletter and get more of G4A straight to your inbox.
If you’re not a tech person, this article is for you.
Last May, the world watched in wonder and terror as Google showcased their new assistant technology:
Let me reinforce something crucial — the assistant has generated those sounds. This is not a bank of pre-recorded sentences the technology could draw from. Each ‘hm’ and ‘ehh’ was chosen and crafted by a software.
How did it do it?
The humans who created this piece of technology don’t exactly understand the totality of processes that happens inside the machinemind while it takes all those speech decisions. The algorithm taught itself how to do it! The superpop term around this process has been slammed in our ears repeatedly for the past few years: machine learning.
The thing is, computer algorithms are very prominent in our lives and their presence only increases: choosing the prices of items in our supermarkets, selecting the political ads we’ll watch, competing against us in games, keeping track of our money transactions and all sorts of uses.
Yet, for the uninitiated in the software magics, they remain shrouded in mystery. That’s why we’re here today to give non-techy people a basic idea on how they work. But we don’t believe in theory per theory, so we found two interesting examples to illustrate our point:
Netflix’s algorithm choosing what you should watch.
Turbine’s algorithm grouping different cancer-fighting drugs and their interactions in the body.
With those wildly different examples, we expect to give a broad vision on the infinite possibility space of machine learning.
You may think you know the Netflix catalog very well, but you know who knows it best? The algorithm.
Whenever Netflix asks itself the basic ‘what should I recommend to this person’ question, it fires a complicated process.
The way Netflix sees its own arsenal is very different from ours: it knows the genre, the exact length, the full cast, the music style of the soundtrack, the colour of the poster and many other datapoints from each and every piece of media it contains.
So when you’re about to be recommended something, it first checks everything you’ve already watched to find hints on your personal preferences.
Then it can connect your profile to a bigger database: Globally, 88% of people who watched Dark also watched Stranger Things. In Spain, the value decreases to 79%. In Majorca, it increases to 82%.
Netflix’s recommendation is not a mere “you like horror, you should watch more horror” human logic — the scope of its precision is far beyond the human capacities.
Turbine’s work is not that different. Imagine that you asked the algorithm ‘what do the patients that respond to this drug have in common’.
As Netflix knows its catalog datapoints, Turbine’s bot knows its own: how the thousands of proteins interact inside our bodycells and especially how they interact with each other.
Then there’s the user information — Netflix grabs your watch history and connects to its previous catalog knowledge, Turbine needs to be fed the specifics of the cancer drug and cancer cells: their genetic code, cell type, treatment dose, etc.
Finally, pattern recognition: in 89% of cells with this specific protein mutation, this drug combination effectively killed the lung cancer cells.
Of course a well instructed doctor can understand the effects and uses of many drugs, but here are millions of possibilities and combinations. It’s nobody’s fault, it’s just more than a human can process with a meat brain.
Right. The whole best answer gig was neat, but that’s perhaps not the most interesting part of this whole machine learning jazz.
A couple of years ago, I heard the difference between data, knowledge and wisdom. I know I seem to be falling on a tangent here, but stick with me.
Data is the raw information: this object is green, this object is round.
Knowledge is the processing of data in a way that’s useful: oh that’s a watermelon, they are good for eating.
Wisdom is the contextualization of the knowledge: watermelons are fine on their own, but dreadful when mixed with chicken and bread.
Now back to our example machines! Remember when I said the Netflix algorithm analyses your view habits and compares them to macro patterns? Well it does a little more than that. It generates those patterns.
The humans who made Netflix added all those data collection points to every user: what do you watch, how much you watch, where do you live, how many times you log in every week etc. When they had enough data, they started to make sense of it. By transforming all those isolated data points (Anna watched Rick and Morty, Josh watched Westworld, Achmed watched Breaking Bad) into knowledge (87% of users who watched The Avengers, also watched Batman), it transforms it into something useful.
Let’s say somebody in Rio de Janeiro posts a viral funny image comparing Star Trek and The Truman Show. The shows have different genres, cast, style, were made in different decades and share basically no attributes. Even so, the algorithm will correctly recommend Brazilians one show after the other simply because the data demonstrates people are clicking. It didn’t need any social media representative in Rio, it just fed information for itself.
The same overall process works for Turbine: the more people survive, the more people will survive. Every drug developed using Turbine and tested in trials increases the datapoints in the system and makes it work better for everyone else.
It’s a refining process — if now our best guess is to mix two different drugs, with more data we may be able to safely mix six different drugs and increase everybody’s chance of survival. In Netflix’s case, the information gets hidden from other parties, but in Turbine’s not so much. Other medical research facilities and specially other medical bots can use the patterns they find to dig even deeper into increasingly specific and custom-made treatments.
In a way, every survivor from any disease can help every human live longer as long as the data is collected and properly contextualized. Cancer may be our example here, but Turbine’s algorithm could be cross checked with diabetes, HIV and pneumonia bots to become even further intelligent and useful.
The things that algorithms can’t reach right now is wisdom. As I mentioned before, wisdom is knowing tomatoes are fruits, but not adding them to the fruit salad.
I know sometimes it seems to us that the computers are already intelligent — specially when we’re faced with something like that Google video, but that’s not yet true.
Algorithms are great in teaching themselves how to better answer the questions we give them, but they are as good as the questions we ask. Most of the tech energy nowadays is destined for sales: from advertising to exchanging cryptocurrencies.
G4A is happy to have Turbine as one of its alumni. It’s important to think about the future of mankind when dealing with the new powerful tools we have in our disposal — and having machines that cured cancer really sounds like one of those sci-fi golden futures, doesn’t it?