Imtiaz Humayun

@imtiazprio

Research Scientist at Google. PhD @RiceUniversity with @rbaraniuk. Co-founded @bengali_ai. DL theory, interpretability and generative models 🇧🇩

New York, USA

imtiazhumayun.github.io

Joined August 2018

Pinned
Imtiaz Humayun
@imtiazprio
Mar 13, 2023
How are Deep Neural Networks black-boxes if you can visualize them in an 'exact' manner? Our new #CVPR23 paper, presents a fast and scalable PyTorch toolbox to visualize the linear regions, aka partition+decision boundary, of any DNN (red🔻)! bit.ly/splinecam 🧵 1/N
00:00
Imtiaz Humayun
@imtiazprio
Feb 12, 2025
Indeed, it is that simple! The wiggliness induced by each layer allows NNs to approximate non-linear functions. More layers -> more possible wiggle -> more non-linearity. A nice way of thinking about this is imagining NNs doing origami on an elastic piece of paper!
Jack Morris
@jxmnop
Feb 11, 2025
the simplest explanation for why NNs work so well
Imtiaz Humayun
@imtiazprio
Feb 9, 2025
Checkout buddy @randall_balestr discussing our work and more, why thinking of NNs as splines can provide strong intuitions of how they learn, generalize and grok! Also features a small cameo from yours truly!!
Machine Learning Street Talk
@MLStreetTalk
Feb 8, 2025
Professor @randall_balestr discussing some exciting research he has been working on recently, in particular around spline-based interpretability. • Why reconstruction learning can fail for perception • How deep nets partition space like a crystal • Spline theory new insights
00:00
Imtiaz Humayun
@imtiazprio
Nov 7, 2024
We have already achieved self-improvement in diffusion models, via its own synthetic data! Not that far away definitely!
01:22
Chubby♨️
@kimmonismus
Nov 7, 2024
Mustafa Suleyman, CEO of Microsoft AI, says in an interview surprisingly openly that self-improving AI will be achieved in 3-5 years at the latest, but certainly before 2030. That would be the point in time when we would slowly enter the singularity. The consequences after that
Imtiaz Humayun
@imtiazprio
Jun 18, 2024
New pre-print up with amazing co-authors, led by @Omek10!!
Omer Ronen
@Omek10
Jun 18, 2024
We developed a simple and computationally efficient method to mitigate over exploration in latent space optimization, code is available at github.com/OmerRonen/scal…. Super fun project with a great team! @randall_balestr @imtiazprio @rbaraniuk and @bbiinnyyuu