Stephen Hawking: Machine Learning is Scary

SkynetAn eye-catching piece appearing in today’s edition of The Independent featured the thoughts of luminaries from the scientific world – renowned physicist Stephen Hawking, U.C. Berkeley computer-science professor Stuart Russell, and MIT physics professors Max Tegmark and Frank Wilczek – about the potential perils of artificial intelligence. Inspired by the new Johnny Depp flick Transcendence, the scientists said it would be the “worst mistake in history” to dismiss the threat of artificial intelligence.

Success in creating AI would be the biggest event in human history,” the article continued. “Unfortunately, it might also be the last, unless we learn how to avoid the risks.”

The professors wrote that in the future there may be nothing to prevent machines with superhuman intelligence from self-improving, triggering a so-called “singularity.”

One can imagine such technology outsmarting financial markets, out-inventing human researchers, out-manipulating human leaders, and developing weapons we cannot even understand. Whereas the short-term impact of AI depends on who controls it, the long-term impact depends on whether it can be controlled at all,” the article said.

With all due respect to the good professors, I don’t recognize the AI and related machine learning technology they seem to be concerned about. Sure there have been advances in the capabilities of statistical learning algorithms that power such consumer facing deployments such as self-driving cars, a computer winning at Jeopardy! and the digital personal assistants Siri, Google Now and Cortana. But as a data scientist who builds applications with so-called “intelligence” I can safely say we’re not at risk from the scenarios described.

Here is an excellent documentary called “The Smartest Machine On Earth” that tells the story of Watson, IBM’s famous Jeopardy-winning supercomputer, and delves into how IBM used machine learning to make its creation into a game show champion. I think it give a pretty accurate depiction of the level of AI that’s possible today and the foreseeable future and there is certainly nothing to be feared.

The article continues with a discussion of weighing the benefits and risks of self-aware AI:

Although we are facing potentially the best or worst thing to happen to humanity in history, little serious research is devoted to these issues outside non-profit institutes such as the Cambridge Centre for the Study of Existential Risk, the Future of Humanity Institute, the Machine Intelligence Research Institute, and the Future of Life Institute. All of us should ask ourselves what we can do now to improve the chances of reaping the benefits and avoiding the risks.”

With ideas like this, it seems like Hollywood is dictating the purported state-of-the-art in machine intelligence. I can’t this emphatically enough: we are NOT at risk for some kind of a SkyNet take over with “intelligent” Terminators exacting revenge on the human race. The machine learning I know, the algorithms I work with day-in and day-out, are simplistic, delicate, and require significant human supervision to be useful at all. And contrary to the recent marketing department blurbs from big data analytics vendors – data scientist are still a very much needed part of the equation. To think that the stochastic gradient descent algorithm I use for some machine learning applications will one day become sentient, is indeed SciFi.

skynet_humor

So to the good Professor Hawking et al., I just have to say, please chill out and read a good book on statistical learning to more fully understand where we are with practical AI. I’m a bit disappointed with Hawking’s perspective. I met him once years ago at the Pacific Coast Gravity Meeting and was in awe, but now? And John Connor, if you’re out there somewhere, can you please get one of your Terminator friends to come back in time to take care of Johnny Depp?!

For more check out the insideBIGDATA guide to Machine Learning.

 

Sign up for the free insideBIGDATA newsletter.

 

 

 

Comments

  1. The technology is neutral but I am quasi-certain that there are be misuses (deliberately in the present tense) .
    In reality the issues this brings forth are an extension of some of the issues of the previous century namely population ‘shaping’ through advertising/media to use a loose enough term.
    The real problem is in fact in the sophistication of the technology that the poster is trying to downplay, combine with ubiquitous surveillance / data beyond any data scientists wild dream, There is a problem.
    It is wrong and is a typical case of diversion to think that the hypothetical installation of a technology that would objectively decide to turn around and destroy humans because we’re not doing very well as a species (which is somewhat true but still a ridiculous narrative inspired from I,Robot or something like that) is what’s happening.
    Now back to the fact:
    - I don’t know how much machine learning the poster has done , but there are unsupervised learning algorithm being developed that are really quite effective, typical machine learning was just about statistical modelling which we all know has it’s limitation but we’re coming up with better ‘deeper’ information encoding techniques with auto-encoders. (I will honestly say that I haven’t developed anything yet with this but I know that ‘Deep Mind’ had something along these lines worth 239 Million pounds to Google).
    - Even the most basic machine learning algorithms like decision tree learning for example , can be used to automatically ‘make sense’ of the tonne of data poured into the big corporations servers , which do not always have our best interest in mind, but rather our direct and most ruthless exploitation under the banner of effective cost/benefit analysis seldom that usually doesn’t take any ‘human’ factor into account.

    - If i may , summon a little comparison for illustrative purposes, imagine the dumbest possible machine learning algorithm that is a simplex, couple of coefficients into a unit that under specific conditions emits again in one layer. If for example the input is a preprocessed image , i can bring this analogy to a military official that gets a stimulus when certain parameters are in a situation. this reaction is up to fire the drone and rain hell on evil enemy of the supreme and god blessed land of the freest free democratic spirits that have ever roamed past or present existence. Since the military official is human , he has spikes and neurochemistry as well as morals perhaps and other parameters , that at would at the very least make him tired! the simplex for example would only tire when the seemingly endless supply of energy dries up. in essence this can be a cataclyst for further inequality/unjust domination in the world that is already largely driven by geopolitical/resources scarcity distress.

    - Finally , The poster is suggesting that a quantum mechanics pioneer like Stephen Hawkins cannot grasp something as ‘simple’ as he himself describes as statistical learning/modelling.

    In sum , advertising and industro-military complex is where this innovation would see purposes that are way beyond mere protection of the nations involved of danger but something more along the line of undetected enslavement of mankind, we shouldn’t attack the technology since someone is always going to create it , but come up with safeguards , full TRANSPARENCY as a minimum of a nation that respects human progress and have a iota of decency learned its lessons from all the dictatorships and harm we’ve done to each other. So at least to shut up the critics once and for all if they are so certain they can do what they do without any judicial reprisal (I am perhaps the person that has been the most negatively affected by this technological innovation) and I am passionated by technology yet I am starting to dread it.

    Think about what’s sacred for your children that you at least had at some point in your life if you lived pre 2000-2010′s (exact timeline impossible) , if you fail to realise the importance of stepping up and standing up to these overblown and quite moronic bullies that operate no different than a mafia, then you should at the very least stop calling yourselves a land of justice (i am addressing the politicians not the people) where it used to be that each man’s house is his castle.

    I apologize for the big rant, if you’ve read this far, know that I am being very diplomatic in spite of how angry i might seem (you can gauge this to be relative to where you are in the spectrum of opinion ,psychology manipulation by ‘the planners’

    with grudges and discontent from london

    • @H: I appreciate where you’re coming from, many people are incensed by the privacy concerns surrounding big data and its enabling technology machine learning. I actually agree with this point of view to a degree. But that’s not what this article is about. The professors are not talking about invasion of privacy, they’re talking about the rise of self-aware AIs who could endanger their creators. That’s completely different, and I believe that Mssr. Hawking could understand the current technology’s limitations, just as you could, if you just worked with the technology for period. I think then you’d agree how limited it is in the terms of the original concern. True, unsupervised learning can discover new, previously unknown knowledge from data, but it takes a trained data scientist to interpret the results. For example, take a look at an agglomerative technique like hierarchical clustering, or even something like k-means – you need to hand hold the process to achieve results and even then, the results are subjective and require domain experience to make actionable decisions. To think that a machine can suddenly spring into independent thought is just SciFi. I believe the trouble with notions like what the professors originated is that they’re not founded on experience with actually using machine learning. And using machine learning is not just firing up R and passing a few arguments to hclust() to see what comes out. Fully understanding machine learning requires a background in mathematics, statistics, probability theory, computer science, combinatorics, etc. Without that background you won’t have a sense for what’s possible, and more importantly what’s not. Further, making policy decisions about AI without this background is like urging legislation about climate change, either pro or con, without really understanding the science.

Resource Links: