robocrunch
Boaz Barak
@boazbaraktcs
Theoretical Computer Scientist. See also https://t.co/EXWR5jOrFW and https://t.co/JEkxX4JHSw
Tweets by Boaz Barak
5/8 The CMF has an issue with acceleration, but the papers they cite often have tiny effect sizes. Sometimes the papers come to opposite conclusion as with Burris et al (2006) that showed the benefits of "Algebra I for all 8th graders" policy.
Shared by
Boaz Barak
at
5/2/2022
1/8 Confused about the California Math Framework controversy? Don't feel like reading a 900 pages document? Brian Conrad (math prof at Stanford) did it for you at https://t.co/lvAFes6YG9 Some highlights (or lowlights) include:...
Shared by
Boaz Barak
at
5/2/2022
To be precise, the model has about 80K parameters (100x28x28 in first layer and 10x100 in second one). Square root of 80K is ~282
Shared by
Boaz Barak
at
4/26/2022
Model has about 10,000 parameters (2 layer MLP on MNIST with 100 hidden units). You expect a random direction in R^d to be ~sqrt(d) standard deviations worse than the optimal one.
Shared by
Boaz Barak
at
4/26/2022
Ignoring the visual recognition aspect, here is how GPT3 does on this challenge. (To be fair, this is the second generation - in the first attempt it said that the second sequence has a wide range of shapes). https://t.co/cNbO8YcLas
Shared by
Boaz Barak
at
4/8/2022
Should have it generate "Photo of a researcher on symbolic AI, explaining on a blackboard why neural networks cannot generate new data" 😀
Shared by
Boaz Barak
at
4/6/2022
This Friday in our ML foundations seminar, Eero Simoncelli will show us how to extract implicit image priors from neural networks. https://t.co/4LwTAfJsur https://t.co/HnCe3EjwB6
Shared by
Boaz Barak
at
4/4/2022
This can and should be done, but not via lowering standards (e.g., skipping algebra and calculus in favor of "data science") but by providing more training and support for both students and teachers.
Shared by
Boaz Barak
at
3/31/2022
You might also be interested in this article from 1993: "A Sociological History of the Neural Network Controversy" by Olazaran https://t.co/4aCKWO6vmP
Shared by
Boaz Barak
at
3/15/2022
One view of this: modern techniques such as cross-entropy loss, augmentation , dropout could optimize the 1989 "baby" network on the "baby dataset". Alternative view: could have discovered all the above techniques using only the "babies"
Shared by
Boaz Barak
at
3/15/2022
Hope the updated CMF fixed deep issues in first version, but this is wrong and misleading. Funneling high-school kids away from algebra and calculus and into "spreadsheet science" is a sure way to shut them out of careers in artificial intelligence or CS https://t.co/1G8FXSkXjM
Shared by
Boaz Barak
at
3/14/2022
The issue is not that calculus is a relic - in these days of gradient descent it's more relevant than ever. Challenge is how to ensure more students, in particular students of color and with less resources, actually get to calculus. This can be done, see https://t.co/UZr3n0Flyn
Shared by
Boaz Barak
at
3/14/2022
This analysis of ranking is worth reading, not because of anything specific to Columbia, but because it gives a quantitative evidence to how rankings can distort not just reporting of data but also incentives to provide good education. https://t.co/HUahN4fBBR
Shared by
Boaz Barak
at
3/9/2022
Maybe the formal statement of the negative should be "By January 1, 2027, models holding most state-of-the-art position in most benchmarked tasks in natural language processing will include a significant architectural innovation over the transformer architecture."
Shared by
Boaz Barak
at
3/2/2022
Am I understanding correctly that @srush_nlp's moral position is there will be some innovation in the next 5 years apart from just more compute & data that will be crucial to achieve SOTA on benchmarks? It's a good Q though whether the benchmarks will last 5 years...
Shared by
Boaz Barak
at
3/2/2022
I agree, which is why I include quantum in my undergraduate intro to computation https://t.co/bWj1GNdYa1 The beauty of TCS is that we use math to model computation *as it happens in the world*. If nature is quantum, we don't get to choose to ignore it.
Shared by
Boaz Barak
at
1/19/2022
1/2 Excited to hear from @drfeifei in machine learning theory seminar this Friday 3pm EST how vision is used for interaction with environment. As usual join mailing list on https://t.co/lW2xaXRUlq for announcements and Zoom link. Last talk of fall, but great plans for spring:
Shared by
Boaz Barak
at
12/6/2021
1/14 More than 150 scientists & educators signed open letter raising alarm on efforts to water down K-12 math education https://t.co/tuFcNGaUkH Signers include Fields, Nobel & Turing laurates, and also founders of HS STEM educational initiatives (eg @adrian_mims, @minilek).
Shared by
Boaz Barak
at
12/3/2021
On Friday in the ML Theory Seminar, Sho Yaida (Facebook) will tell us about theoretically analyzing finite-width deep nets. As usual, see https://t.co/lW2xaXRUlq to sign up for the mailing list and get the Zoom link.
Shared by
Boaz Barak
at
11/30/2021
Some good advice for (especially first-time) program committee members in theoretical computer science conferences. Note that "PC member" in TCS roughly corresponds to "area chair" in other fields. https://t.co/k1oKbsYY47
Shared by
Boaz Barak
at
11/23/2021
There is an N (natural) in code as well. See these copilot continuations of the semantically identical Python snippets def f(i): i += and def f(L): L +=
Shared by
Boaz Barak
at
11/23/2021
Optimal local classical algorithm for maximum cut in sparse high-girth graphs by @ahmed_el_alaoui, @Andrea__M, and @geoishard. In particular confirms conjecture with @kmarwahaha that local quantum algorithms offer no advantage for such graphs. https://t.co/YHIlzkj2wR
Shared by
Boaz Barak
at
11/19/2021
Even more opportunities posted now on https://t.co/fL9vDvhoLS including a joint position with Sasha Rakhlin at MIT. Sasha also has another postdoc opening in reinforcement learning.
Shared by
Boaz Barak
at
11/15/2021
Completely agree, but rather than trying to expand access to calculus (e.g. as done by https://t.co/0PtItsr66N ), new system makes it harder for less priveleged students to take it. Students can't do it without either taking a private course or doubling up / compressing.
Shared by
Boaz Barak
at
11/14/2021
Public Service Announcement: you can actually type math in powerpoint almost as easily as Latex: https://t.co/RNPY3CewKV With work you can even make all equations have different color https://t.co/j3VP0V0LKm https://t.co/090nbGUrWU
Shared by
Boaz Barak
at
11/12/2021
Professors keep insisting on teaching students sputnik-era useless topics such as calculus and algebra. Students want to learn about modern 21st century topics such as deep learning with gradient descent! Enough with teaching number theory - teach cryptocurrencies!
Shared by
Boaz Barak
at
11/10/2021
Playing around with deep learning gives me new appreciation for practitioners. It takes me ages fixing typos, wading through dimensions mismatches and cuda errors, until I can finally get to the point where my network doesn't work because I chose the wrong architecture..
Shared by
Boaz Barak
at
11/9/2021