|
David Sussillo ☝️🤓
@
SussilloDavid
Portola Valley, CA, USA
|
|
Computational neuroscience / deep learning. Scientist at Google Brain, adjunct prof at Stanford. All my tweet are belong to me. 😋🍑🍃
|
|
|
1.256
Tweetovi
|
426
Pratim
|
7.830
Osobe koje vas prate
|
| Tweetovi |
| David Sussillo ☝️🤓 proslijedio/la je tweet | ||
|
Timothy Verstynen
@tdverstynen
|
27. sij |
|
I think the real lesson of places like DeepMind, Janelia, and Allen Brain Inst. is the amazing things scientists can do when you actually give them the resources they need to do their work properly, instead of spending so much energy hustling for cash/resources.
|
||
|
|
||
| David Sussillo ☝️🤓 proslijedio/la je tweet | ||
|
Janelle Shane
@JanelleCShane
|
26. sij |
|
I am making these neural net horseradish brownies and I regret everything twitter.com/JanelleCShane/… pic.twitter.com/WNZrcixSYf
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
24. sij |
|
More broadly, I suspect that policing bad science on Twitter isn’t going to stop bad science.
For egregious cases I follow @RetractionWatch and also get a kick out of @Neuro_Skeptic but these accounts serve very different purposes than what was implied by Andrew’s original poll.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
24. sij |
|
Criticism of preprints on Twitter seems particularly pernicious because the work isn’t finished yet and comments that are ever-lasting & widely distributed could have disproportionately negative effects on publication.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
24. sij |
|
‘Criticize privately, praise publicly’ is my motto.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
24. sij |
|
Sure! If you know the authors well enough.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
24. sij |
|
Emphatically NOT twitter.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
23. sij |
|
Yes! We used a well known bag-of-words baseline that is interpretable.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
23. sij |
|
A+ tweet
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
22. sij |
|
Like linear regression but squishier.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
22. sij |
|
Today is the first day I have ever used logistic regression. 😱
|
||
|
|
||
| David Sussillo ☝️🤓 proslijedio/la je tweet | ||
|
eLife - the journal
@eLife
|
10. sij |
|
The unexpected discovery that the brain area responsible for hand and arm movements is also active during speech suggests a new way to help restore lost speech elifesciences.org/digests/46015/…
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
30. pro |
|
Fair. But it depends on what the deep net is doing. 10 years ago we didn’t even understand RNNs that did very simple things and we barely knew how to build them.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
30. pro |
|
I wonder if Larry wouldn’t significantly revise some of this, given the progress in model building over the past decade.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
28. pro |
|
I completely agree that dimensionality is the the current problem. We are beginning to contend with it though.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
27. pro |
|
As if I have a definition of understanding.
People design linear systems. There’s a whole subfield of engineering dedicated to filters, which iir jargon fir linear system design.
If building something isn’t understanding, nothing is.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
27. pro |
|
Wrong. While it is true that one can always linearize locally, it remains a surprising empirical fact for RNNs that the volume where the linearization is valid often sheds meaningful light on the nonlinear computation.
That was not necessarily going to be true a priori.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
27. pro |
|
I don’t regret “prerequisite” insofar as it forced the issue but I agree the word is too strong.
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
27. pro |
|
Me thinking about brains and artificial neural networks. pic.twitter.com/FXclTV1zCl
|
||
|
|
||
|
David Sussillo ☝️🤓
@SussilloDavid
|
27. pro |
|
The Falses have it by a 3:1 margin.
For my part I would have voted True. That is, we’ll have chewed through what DL has to offer conceptually long before understanding biological neural computation and the exercise will have been invaluable. twitter.com/SussilloDavid/…
|
||
|
|
||