Piyush Rai
@praiml
Followers
239
Following
60
Statuses
55
CS faculty at IIT Kanpur. Research interests in machine learning and AI. Views/opinions here are personal.
Kanpur
Joined July 2017
You can do statistics/ML/deep learning, or "generative AI" (or whatever you want to call it) using tree-based approaches too, as they basically do divide-and-conquer, i.e., breaking big, complex problems into smaller, simpler problems (and also provide better interpretability)
Statistics — just dot products. Generative AI — just dot products. Bridging the two is "traditional" machine learning which oddly thrives on tree-based models.
0
0
11
There was also a time, from an era long gone by, when anxious students would ask "Don't kernel SVMs which operate in ultra-high dimensional feature spaces overfit?", and the answer thankfully wasn't as speculative
What's the right answer in my Deep Learning class when anxious students say: "Doesn't that lead to overfitting!"
0
0
5
this approach is indeed more viable than Altman's "universal basic compute" :)
Grad Students: Networking is critical for success in academia. But the real question is WHO to network with. I’ve known grad students who will spend time at conferences drinking artisanal kombucha with grad students from programs with no GPUs. Is that 1/
0
0
3
tl;dr: if you can't beat them, join them (ps: it's quite likely the thread is a satire)
Grad Students: Networking is critical for success in academia. But the real question is WHO to network with. I’ve known grad students who will spend time at conferences drinking beers in the bars with grad students from unranked programs. Is that 1/
0
0
5
Well, it will soon do your laundry and dishes, too. The existential dread is real
0
0
3
The shamer is apparently too obsessed with conformal prediction as if everything else is BS. Someone please tell the person that the VAE paper just won the test of time award at ICLR, and that the loss function in diffusion models is akin to the ELBO in variational Bayes
Disagree all you want with Bayesian learning, but there are _far_ better ways to do it than screenshotting a PhD graduate's thesis and shaming an entire university. I didn't study Bayesian methods at Cambridge but I interacted with many Bayesians and learnt _so much_ from them.
1
1
24
@mnwsth @theindiamini Bhai.. 29 seconds to sirf chair adjust karne aur focus karne mein lag jaate hain :)
1
0
1
Why just high-school kids? Please include *all* kids (including toddlers who may potentially submit their creative and original ML ideas with the help of Neuralink chips). Leave no talent untapped in your attempts at making social impact.
NeurIPS 2024 will have a track for papers from high schoolers.
4
13
124
Patience.
Many academics come across tools or apps they can't live without during their PhDs. Do you have one you would recommend people check out?
1
0
13
Yearly reminder **to universities** that if you aren't admitting a PhD applicant, don't forget that you may be losing out on a star who might apply for faculty positions after 5 years with a stellar PhD from elsewhere and ignore your univ. So don't blow future colleagues off :)
Yearly reminder that if you are declining offers at Universities, don’t forget you may be applying to the same Universities for faculty positions in 5 years and so don’t blow future colleagues off.
1
0
23
Or maybe it is simulating a magic show where a magician duck makes humans disappear?
Confirmed! Sora is just a standard learnt latent space diffusion model! Not some reality simulator. The duck stomps over the human being that totally disappears after the stomping 🤣 Let’s tone down the hype…
0
0
3
This is nice and timely. You could also consider discussing some other complementary/competing approaches (like you do for conformal prediction) that estimate uncertainty using a single model, e.g., evidential deep learning, prior networks, distance awareness based methods, etc
1
0
3
Hmmm.. Fully support your (and anyone's who is considering) move but, FWIW, capturing point clouds efficiently and in cost-effective ways is also a machine learning (more specifically, sequential decision making) problem
I’m 25 and tomorrow is the first day of my new career. After an unnecessarily hectic year in tech/ML, I’m out. I’ve decided that the industry is on track to become something I don’t want to participate in. Labor will be squeezed between LLMs, increasing amounts of insider games, and a desperate labor force willing to work for cheaper and longer. There is no shortage of talent in tech, in fact it seems to be a black hole for talented people actually capable of changing the world for better. All of this on the backdrop of talent shortages in other industries. These industries that are neglected, present the highest ROI for applying tech. This is where applying machine learning yields actual ROI. I’d describe this as a breath of fresh air. It seems like tech has just become SaaS companies exchanging api keys and acting like billions of dollars of efficiency and value is being created. Fake! Ownership and mastery are important to me. I want to master a skill and build a business around it. Simple is better here. Machine learning seems to offer neither, so I must leave to pursue my new field of work. After long consideration I have decided to move into geomatics and surveying. At a glance - indoor and outdoor job - ancient - market can be cornered - cannot be outsourced or mechanized - basic skills can be mastered On top of job perks the licensing is the ultimate draw. Professional Land Surveyor (PLS) licenses are rare, required, and respected. The median age of these license holders is in their 60s. In other words at least half of the license holders are looking to retire soon. But it’s a very long process to get a license, even with a bachelors of applied math. Likely more school and years of work experience. The first step is in transitioning my data engineering skills from machine learning with point clouds to capturing the point clouds as a lidar tech. New job starts next week. Today is the first day of the rest of your like or some cheesy bullshit. PS. I’ll still be keeping up with ml but only as a hobby
0
0
1
1
0
1
Hmmm.. Saying that is quite "mean"
Quantile regression is not for the average.
0
0
2
Super impressive! And Alex is an undergrad, not an LLM
for the past 2 weeks I’ve been reading every abstract from #neurips2023 here are my notes about what I discovered: i think everyone can learn something new from this, and I hope this resource is useful!
0
0
8