You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
Looking at what you can do with deep learning today and thinking "let's scale this up to AGI" is basically the equivalent of watching a magic trick and thinking "wow magic is real! im gonna start a magic company that uses magic spells to generate infinite value"
Magic tricks, just like successful deep learning solutions, are engineering: they only do one thing, they are only possible within a specific context, and they require hard work. There's no free lunch.
I am curious: what do you think of the approach of @singularity_net using OpenCog to sort of or orchestrate a bunch of small services ?
Most DNN systems are trained in one domain, but to extent domain is broad, they still seem to do pretty well. Engineering isn’t trivial, and there are limits to what current nets can do, but wasn’t it scale, compute, & bigger data that proved out some long-standing hypotheses?
Q:(putting aside optimization technique) If AGI is achievable, will the perceptron be the fundamental building block? Or do you think some other unit, whether more or less complex, is needed? Will symbolic logic need to be incorporated, or is statistical approach sufficient?
Thanks for saying it loud. We have not been able to fully comprehend human intelligence yet and we are already claiming to be on the path of creating an AGI.
What remains untold is the time it takes to craft an architecture (I am not sure if there is a theoretical explanation of why certain architectures work) and time to fine tune hundreds and millions of parameters.
There are no free lunches and in fact, there is a theorem: https://en.wikipedia.org/wiki/No_free_lunch_theorem … In layman terms, an algorithms perform that performs well for a *given task* is not universally better than any other algorithms.
Just like WW2 and the imaginary armies using magic though, while things are only have assurance of properties within a given context, you can get practical usable results in entirely different context if all you care about is having interesting new results.
i.e. you can get workable results from deep learning with no clue how deep learning works.
A guy from the IBM Watson team I met this week said it best: if it's written in Python, it's Machine Learning. If it's written in PowerPoint, it's AI.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
