(s)gd is all you need. prev: cs/soc @stanford, ml @oxfordstats. now: stats @harvard. always @zahra_thab.
cambridge, ma
Joined February 2014
- Tweets4.1K
- Following745
- Followers1.3K
- Likes11.5K
We wrote something about world models for biology! Taking a swing at discussing a new fad is tricky, but we gave it a shot. My perspective is that WMs in biology are really a series of complex missing data problems. This requires new forms of data and new methods to address.
Biology does not need models that just describe the world. It needs models that predict what happens when we intervene.
In @CellCellPress Focus on AI in Biology, we lay out how to build biomedical world models, from the data they need to how they should be trained and evaluated.
Nic Fishman retweeted
Biology does not need models that just describe the world. It needs models that predict what happens when we intervene.
In @CellCellPress Focus on AI in Biology, we lay out how to build biomedical world models, from the data they need to how they should be trained and evaluated.
Was so much fun to give this talk! You can find the associated paper threads here!
1. Generative distribution embeddings: xitter.cf/njwfish/status/1927027…
2. Distribution conditioned transport: xitter.cf/njwfish/status/2029931…
Next, we had @njwfish present "Generative Distribution Embeddings" presenting his work from:
(GDE): arxiv.org/abs/2505.18150
(Distribution-Conditioned Transport; DCT): arxiv.org/abs/2603.04736
@iclr_conf #ICLR2026 #AI4Bio #Genomics #Gen2Workshop
If you are in Rio would love to see you at the workshop! I am giving a keynote at 1:55 BST on my work on generative modeling on the space of distributions. It will be a good time!
1 week to go until our inaugural Gen² workshop at #ICLR2026 on April 27! We’re excited to share our updated schedule: genai-in-genomics.github.io/…
Nic Fishman retweeted
1 week to go until our inaugural Gen² workshop at #ICLR2026 on April 27! We’re excited to share our updated schedule: genai-in-genomics.github.io/…
Come say hi at poster Pavilion 3 P3-#1013!
Count Bridges: the discrete analog of diffusion for count data. Closed-form on ℤᵈ.
On count tasks, they beat CFM and DFM by an order of magnitude — and enable nucleotide-resolution single-cell RNA-seq and reference-free spatial deconvolution.
#ICLR2026 poster tomorrow 🧵
Nic Fishman retweeted
Count Bridges: the discrete analog of diffusion for count data. Closed-form on ℤᵈ.
On count tasks, they beat CFM and DFM by an order of magnitude — and enable nucleotide-resolution single-cell RNA-seq and reference-free spatial deconvolution.
#ICLR2026 poster tomorrow 🧵
Count Bridges: the discrete analog of diffusion for count data. Closed-form on ℤᵈ.
On count tasks, they beat CFM and DFM by an order of magnitude — and enable nucleotide-resolution single-cell RNA-seq and reference-free spatial deconvolution.
#ICLR2026 poster tomorrow 🧵
Our second application is magic: we aggregate MERFISH into Visium-like spots, and only train on those. Using an EM approach, we can learn to generate single-cell profiles using only nuclear-stain images as side information. This beats a SOTA canonical for spatial deconvolution.
Joint work with Gokul Gowri, Tanush Kumar, Jiaqi Lu, @ValentinDeBort1, @jgooten, and @omarabudayyeh.
Come by the poster tomorrow, Sat Apr 25, 3:15 PM – 5:45 PM BRT.
📄 openreview.net/forum?id=4nOZ…
💻 github.com/njwfish/count_bri…
Nic Fishman retweeted
Parallelizing nonlinear RNNs is gaining traction!
More efficient than transformers; more expressive than linear RNNs.
My PhD thesis provides an intro guide to the math (Newton's method) behind the parallelization.
Great as a quick-start if you want to explore this new field!
There's a growing worry that AI will break empirical social science -- that agents can p-hack until they find something that "works."
We think that worry deserves to be taken seriously. Our new paper shows that is true empirically and makes it precise: njw.fish/static/papers/agent…
We open-source the full pipeline: specification trees, agent prompts, mechanical verification, and all estimation code.
Paper: njw.fish/static/papers/agent…
Code: github.com/gsekeres/agentic-…
Joint work with the excellent @GabeSekeres!
These ideas are super relevant to some of the thoughts I have seen floating around from @ahall_research, @causalinf, @ben_golub, @akoustov, @ddkang.