Thoughts from Your Humble Administrators - Jan 8, 2017

What we have been thinking last week:

I (Arthur) am traveling, thus the late issue.

  • Our group was once called simply "Deep Learning".  Of course, all of you have read many articles about "Deep Learning is/is not a hype" type of articles.  So you may wonder how to call B.S. in those articles.
  • There is a trick - Does the author mention that deep learning's view point is to "automatically learn a representation", instead of relying on human experts or "feature engineering".   If an article fails to mention this point,  then that article doesn't worth its salt.
  • "But hey Arthur, isn't that just your definition of what deep learning is?"  I know you will say that. 🙂  Not really.  This point happens to be mentioned by Hinton in Lecture 2 of his Coursera class, In Chapter 1 of "Deep Learning" written by Goodfellow.   Goodfellow went on to explain one of the original "deep" means that there is a "deep hierarchical representation".   That's why DNN is good, CNN is  good, RNN? If you expand it, it is deep too.  So perhaps that's why it is also part of "deep learning".
  • My point is: I guess it doesn't really matter if someone is for/against deep learning, what you want to look at is their arguments.  Do they know what they are talking about?  Hint: most of them don't, even authors we allow them to post in this forum.  But we only ask for relevancy, so if the OP has opinions, we let it go.  You are warned though on the validity of some of the blog posts in the forum.
  • CES 2017 is happening.  To AIDLers, perhaps the most relevant is all intelligent agent and virtual assistants.   At least the me, the one which leads the trend recently is Amazon Echo/Dot which is enabled by Alexa.    Not only Amazon's teams ASR capability should be on-par with Apple, MS and Google.  Alexa leads in terms of long-distance speech recognition (based on beam-forming) as well as keyword wakeup (i.e. user can trigger recognition by saying "Alexa" instead of pressing buttons in old iOS.)  Those are impressive features, and tough to work well technically.
  • One thing to point out, while ASR is very impressive in many of these virtual assistants,  the dialogue system is still lacking the "soul".   This says for all chatbots - working for specific domain is okay, but you will quickly notice that it is not real. Jose Diaz asked us if one day a universal real virtual assistant is possible, I can only say it is in a universe far far away.

Must read: all posts about Mario.  Why? We love him since we were young! 🙂


If you like this message, subscribe the Grand Janitor Blog's RSS feed. You can also find me (Arthur) at twitter, LinkedInPlus,  Together with Waikit Lau, I maintain the Deep Learning Facebook forum.  Also check out my awesome employer: Voci.

Leave a Reply

Your email address will not be published. Required fields are marked *