Assaf Shocher

Most of you probably heard about Invertible Neural Networks. But have you heard of *Pseudo*-Invertible Networks? And what does Pseudo-Inverse (PInv) of a non-linear function even mean? 👇

@geopavlakos And this is even without them knowing how good you are at getting agents to locate soccer balls! Well deserved, congrats!

I'm honored to receive the NSF CAREER award! 🎉 Thank you to my students, mentors, and colleagues. Very grateful to NSF for their support.

@YarivLior Very nice!

Why pay full compute for pixels you're not even looking at? In our new work, Foveated Diffusion, we introduce a new concept for efficient image and video generation, motivated by how the human visual system works. (See full thread below)

Video generation models are zero shot planners, but their pixel level trajectories are often not feasible. Our idea: Map the generated video into a differentiable world model latent space and optimize for the nearest feasible trajectory which respects the environment dynamics.

@nickinpractice @huskydogewoof x.com/AssafShocher/s…

@huskydogewoof I recall reading a paper called Idempotent Generative Network. The idea of training a model such that the data distribution is a fixed point isn't new...

“In contrast, our work presents a conceptually different paradigm and does not rely on SDE or ODE formulations as in diffusion or flow models.” Reading it gave me a very particular feeling: an equilibrium flavored sweet spot that drifts away from, yet stays spiritually adjacent to (ranked from farther to closer), diffusion and flow matching, GAN style one shot generation, and DEQ style fixed point thinking. Instead of defining sampling as an explicit time dynamics, they learn a drifting field that moves samples, while training itself evolves the pushforward distribution. The “equilibrium” is not a metaphor, it is literally the point where the model distribution matches the data distribution. The punchline is delightfully clean: it naturally supports one step inference when optimal parameters that can achieve the equilibrium (V, the measure of distance between data and network distribution, being zero) is reached after training; and the ImageNet 256 numbers are strong (FID 1.54 in latent space, 1.61 in pixel space). ❓Question: are we all secretly chasing the same fixed point, just in different coordinate systems?

@iamnopilot There you go: x.com/AssafShocher/s…

show me a generative model that is idempotent

I made a Claude Code skill that generates conference posters 🛠️ Instead of a static PDF, it outputs a single HTML file — drag to resize columns, swap sections, adjust fonts, then give your layout back to Claude. 🔁 🔗 Skill 👉 github.com/ethanweber/pos…

Considering submitting to our workshop How Do Vision Models Work @CVPR 2026! We have both a non-proceedings and proceedings track. More info at sites.google.com/view/how-cvpr-….

An interesting connection between Drifting models @Goodeat258 to Idempotent Generative Networks by @AssafShocher et al: start from drifting loss: L = E_z || f(z) − sg(f(z) + V(f(z))) ||² Set drift as the generator residual: V(x) = f(x) − x Recovering idempotence L = E_z || f(z) - sg(f(f(z)))|| . Eg, a sample which is already on the data manifold should not change!

@isosnovik @Yamitehr @NimrodBe Thank you for pointing out! This is true and we should have cited it, sorry. It is also similar structure to SurVAE. The difference is the use of the degree of freedom. Anyway, we will update and add a citation to PIE, thanks again.

@AssafShocher @Yamitehr @NimrodBe Nice perspective on nonlinear pseudo-inverses. The SPNN construction seems structurally similar to PIE arxiv.org/pdf/2111.00619

Most of you probably heard about Invertible Neural Networks. But have you heard of *Pseudo*-Invertible Networks? And what does Pseudo-Inverse (PInv) of a non-linear function even mean? 👇

This work was led by my talented student @Yamitehr and in collaboration with the great @NimrodBe. Project page: yamitehr.github.io/spnn-project-p…. ArXiv: arxiv.org/abs/2602.06042… Also thanks to Eliakim Hastings Moore and Sir. Roger Penrose. (Below is a snippet from Moore's 1920 paper).

Solving inverse problems with diffusion models is a hot topic (DDRM, DDNM, etc). We combine this with our PInv. We take a frozen diffusion model and apply our NLBP after every single timestep. The diffusion is responsible for the data prior, while our back-projection strictly forces the image to stay on the "manifold" of the target class (e.g., "Must be Smiling"). This is just one demo of what you can do with it and we have many plans for the future.

@LoosJonas The inputs to generative models certainly do.

@AssafShocher Then the question arises: Do neural networks operate basically on the surface of a hypersphere?

What do you see when you imagine a high-dimensional standard Gaussian? There is a known anecdote that I'm embarrassed to say I wasn't aware of. Apparently, my mental image was wrong for years and I just got my mind blown 🤯. Just in case I can save someone else the embarrassment: A high-dim Gaussian doesn't look like a blob, but like a spherical shell! The curse of dimensionality dictates an image different than our intuition. In the illustration below, I visualized a 2D slice of the distribution for increasing num of dims. The blue rings highlight the shell where 99.7% of the data actually lives (Mean ± 3σ). Notice the shift from a solid "blob" at N=2 to a hollow, razor-thin skin at N=100K. 👇

@MaziyarPanahi My guess: Loss is calculated by dividing by hard-coded batch size, but the last batch in the epoch is smaller.

repeat after me, it's ALWAYS: dataset, dataset, dataset!

@docmilanfar It depends on the objective. Data suggest that question titles get downloaded more but get less citations. (of course, you can challenge causality here). link.springer.com/article/10.100…

because a research paper is expected to provide answers - but more specifically: - hides the result, it’s inefficient - teases the reader - shows lack of authority - Betteridge’s law, answer: “No” Eg: “Does sleep affect memory?” vs: “The effects of sleep on memory”