More seriously, writing this was a rare joy. I'm grateful to my (equal) collaborator Jacob Steinhardt and to all of our senior mentors (too many to list here) for holding the bar high, and to the community that makes it safe to bring forth critical discussion like this.
-
-
এই থ্রেডটি দেখান
-
And to the ICML Debates organizers who let us obsess over details to get the message right even at the expense of pushing back posting the paper live until we had a final polished paper we could feel good about sharing.
এই থ্রেডটি দেখান
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
I would really appreciate if you (or someone else) would tweet or write up interesting points of discussion (especially disagreement) that come up.
-
Yes I'll try! Also I just posted to Approximately Correct so if anyone wants to leave longer-form comments / have an async debate, the floor is open. Thanks for the feedback Hal!http://approximatelycorrect.com/2018/07/10/troubling-trends-in-machine-learning-scholarship/ …
-
On #4, I enjoyed a talk recently by
@alaviers (UIUC) on range of skills. One of the points she made in early on was: we often anthropomorphize things (even like "robot manipulator"->"arm"), and this leads to confusion and difficulty when we want to talk about how robots _differ_. -
Thank you! And thanks for pointing me to this post.
I write more about verbiage here:https://medium.com/@alaviers/robotics-automation-and-dance-d93589d60224?source=linkShare-69708ece50e0-1531184141 … -
The mind body connection in human motion that you describe is amazing and so true. I hadn't thought of how it might impact robots.
- কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Is "fully differentiable" mathiness, or misuse of language or both? "Manifold" is also an old favorite (referring to anything but manifolds in the usual math sense). Why is "end-to-end" preferable and over what?
-
Truth is it can be all three! We tried to present what felt like a useful taxonomy, but it's surely not the only way to describe the landscape / the categories are not mutually exclusive. For example, mathiness is often used to disguise speculation as explanation.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Is there an alternative to dropbox where this can be seen? Unfortunately, some of us behind the corporate firewall have Dropbox blocked for security reasons. :(
-
Yes, here are your two options: On Approximately Correct: http://approximatelycorrect.com/2018/07/10/troubling-trends-in-machine-learning-scholarship/ … On arXiv: https://arxiv.org/abs/1807.03341
- কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
To the confusion in the term "generative models" I'd personally add the confusion in "Bayesian networks" which once meant "Belief Networks", i.e. a graphical model, and is now sometimes used to describe a neural network modeling uncertainty.
-
also, a Bayesian Network is a graphical description of conditional independence assumptions and does not imply the use of Bayesian inference. One can run MLE with fixed priors on a Bayesian network.
-
Agreed :) But that is more a 33 year old misunderstanding ...
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
So is this the kind of debate where someone else has an opposing position? Would love to read their paper. PS. Great sentence: "The apparent rapid progress in ML has at times engendered an attitude that strong results excuse weak arguments."
ধন্যবাদ। আপনার সময়রেখাকে আরো ভালো করে তুলতে টুইটার এটিকে ব্যবহার করবে। পূর্বাবস্থায়পূর্বাবস্থায়
-
-
-
Very helpful piece. The analogy with economics here is illuminating. cc
@LPEblogpic.twitter.com/jI7W7Tq5pM
ধন্যবাদ। আপনার সময়রেখাকে আরো ভালো করে তুলতে টুইটার এটিকে ব্যবহার করবে। পূর্বাবস্থায়পূর্বাবস্থায়
-
-
-
Love this so much, thanks for writing it. On “Overloading Technical Terminology” I’d add the terms ‘convolution’ (actually a cross-correlation) and ‘tensor’ (has a very specific meaning in geometry, why don’t we just say array?). Those two have bugged me for ages
-
Well why technically true for convolution, I‘d argue it‘s no problem. The difference is merely a flip anyway. However, it‘s also a flip of a learned kernel. So you don‘t even know whether the kernel is flipped in memory anyway and hence you don‘t know if it‘s cross-correlation.
-
Sure, for the standard 2D convolution/correlation that's true. But when you explore other esoteric variants such as complex-valued filters, group-convolutions, or convolutions on other algebraic structures convolutions and correlations start to diverge somewhat.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
লোড হতে বেশ কিছুক্ষণ সময় নিচ্ছে।
টুইটার তার ক্ষমতার বাইরে চলে গেছে বা কোনো সাময়িক সমস্যার সম্মুখীন হয়েছে আবার চেষ্টা করুন বা আরও তথ্যের জন্য টুইটারের স্থিতি দেখুন।
