It works, although we don't know why...

Layer I of a neural network with two neural layers . The axons of the neurons reflect the term " z 2 " of function f(1). This is a fractal function and mimics the actual fractal topology of our neurones interconnections. Credit: Thomas Kromer,Zentrum für Psychiatrie,Münsterklinik Zwiefalten, Germany

Sometimes I run into a news that is very interesting at several levels. The first, of course it is the news itself. And them there is what the news stimulates in terms of thinking.

This is the case for this news from MIT: researchers at the MIT have been able to achieve similar image recognition capability of primates using a neural network. 
This is a very interesting news, per se. 

Recognising images is one of the easiest thing for our brain, as well as for primates' brains (and possibly for brains of many species). We look into primates brain because they are very similar to us in terms of perception of objects and their conceptualisation (it is much more difficult to correlate to the perception of the environment that a fly surely has: it knows the way to negotiate a space when it flies and it clearly knows what looks good and bad but it is difficult for us to say if we look different from a fly perspective from a cow...).

At the same time, recognising images has always been a challenge for a computer. It requires very hard work (computation) and the results are not comparable to our abilities.

A branch of artificial intelligence has been working on this problem, essential if you want to have a robot that can negotiate and "understand" an environment. And progress have indeed been made. But no clear solution has been found.

The paper published by MIT researchers on Dec 18,2014, describes the approach to image recognition based on "deep neural networks". These are a combination of software and hardware that can be stimulated by data and change their behaviour (the output) based on the data crunched in the past and the feedback on how good they have been in crunching their data. 
A "normal" computer either recognise and image or it does not recognise it. If it does not recognise it it will not get any better if you are resubmitting the same image over and over. Not so with a neural network. If it fails in recognising an image and you tell it so, the next time it will try something different eventually recognising it. More than that. It will apply the learned skill to other images and will get better and better over time.

By increasing the processing capabilities (using relatively cheap graphic processors taken from game consoles) and by training the deep neural network with million of images the researchers at MIT have been able to increase its skill to the point that it compares to the capability of a primate.

Interestingly, they have first look at the reaction of primate brains (by inserting electrodes in various parts of their brain to find out which parts of the brain would react to a certain image) and have coded the deep neural network to mimic those responses. Then they have trained the deep neural network by proposing (and correcting) millions of images.

Now, this is very interesting, a real progress in image recognition. But what was in a way even more interesting to me is the comment of Charles Cadieu, the leading author working at MIT McGovern Institute pointing out that the neural network works really well, although they don't know why!

We have achieved a point where we can start to mimic a brain although we don't really understand what is going on. That, on the one hand, will let us progress in a variety of fields where brain like capabilities are desirable, but on the other hand makes it difficult to expand and increase the quality of results since we don't know what is making them good.

I found this quite fascinating.

Author - Roberto Saracco

© 2010-2020 EIT Digital IVZW. All rights reserved. Legal notice. Privacy Policy.