Julian Tachella

The manuscript encapsulates 5 years of experience working in this rapidly emerging field and focuses on the core concepts and ideas behind self-supervised methods.

The manuscript illustrates some ideas on imaging inverse problems, but the concepts therein can be applied to other data modalities, such as time series or volumetric data.

Self-supervised learning methods for inverse problems allow for training neural networks without ground-truth references, and pave the way for AI models that can discover structures from raw, noisy and/or incomplete, measurement data across science and engineering.

Self-Supervised Learning from Noisy and Incomplete Data Many important problems in science and engineering involve inferring a signal from noisy and/or incomplete observations, where the observation process is known. Historically, this problem has been tac...

📖 We put together with Mike Davies a review of self-supervised learning for inverse problems, covering the main approaches in the literature with a unified notation and analysis.

arxiv.org/abs/2601.03244

You can now solve inverse problems with SOTA diffusion models with DeepInverse and the brand new integration of @hf.co HF diffusers!

Get started with an example: deepinv.github.io/deepinv/auto...

DeepInverse tutorial - computational imaging with AI YouTube video by DeepInverse

We created a 1-hour live-coding tutorial to get started in imaging problems with AI, using the deepinverse library

youtu.be/YRJRgmXV8_I?...

🫂 This is an important milestone for the deepinv maintainer team and contributors, who keep shipping new features!

✨ If you like the library, give us a star of support in GitHub: github.com/deepinv/deep...

Interested?
1- pip install deepinv
2- Get started with our 5-minute tutorial: deepinv.github.io/deepinv/auto...
3- Go deeper by checking examples and our extensive documentation: deepinv.github.io/deepinv/inde...

📸 DeepInverse provides state-of-the-art AI models for image reconstruction across many imaging modalities, such as computational photography, astronomical imaging, microscopy, and medical imaging.

💥 DeepInverse is now part of the official PyTorch Landscape💥

We are excited to join an ecosystem of great open-source AI libraries, including @hf.co diffusers, MONAI, einops, etc.

pytorch.org/blog/deepinv...

Read more about this new feature in the docs: deepinv.github.io/deepinv/user...

and check out a Jupyter notebook demo: deepinv.github.io/deepinv/auto...

This is a great example of the under-the-hood features that make large-scale training seamless with the library!

The new feature leverages closed-form formulas for computing gradients without storing intermediate steps, only requiring an additional call to the least squares solver.

DeepInverse provided access to state-of-the-art matrix-free least squares solvers (eg, conjugate gradient, BiCGStab, lsqr, etc), but backproping through their steps required significant memory and slowed down training.

Many reconstruction models rely on a differentiable least squares solver, such as unfolded networks with proximal steps or reconstruct-anything-model.

🔎 A focus on the new implicit backprop for least squares solvers (by Minh Hai Nguyen), which unlocks training in large-scale imaging settings:

🚀 New Features:

- Implicit backprop for least squares solvers

- noise statistics for SAR imaging

- Multi-coil MRI coil-map estimation acceleration via CuPy

- RicianNoise model

- Self-supervised super-resolution loss

💥 DeepInverse v0.3.5 is out! 💥

docs: deepinv.github.io/deepinv/

github: github.com/deepinv/deep...

Stay tuned for more hackathons to come!
🫂If you are interested in joining the community
- join the discord discord.gg/qBqY5jKw3p
- open an issue or a pull request deepinv.github.io
- drop us a message!

❔ Uncertainty Quantification
- coverage plots + conformal prediction
🏫 Trainer v2!

💥 Self-supervised learning
- Noise2Self
- MERLIN for SAR

📖 Datasets
- PETRIC challenge dataset
- BSD500 dataset
- Calgary MRI dataset

🚀 More optimizers
- implicit differentiation for linear solvers
- more noisy data fidelities for diffusion
- SPECT preconditioned reconstruction
- phase unwrapping reconstruction

🖼️ More image priors/regularizers
- latent diffusion models
- general restoration models for PnP
- complex wavelets prior
- weakly convex regularisers

🌌 Large-scale imaging:
- multi-GPU distributed denoising/reconstruction
- 3D TV/TGV denoising in example
- 3D DRUNet

🔬 New forward operators
- ultrasound physics
- MRI-NUFFT physics wrapper
- ++ ASTRA integration
- near & far field radar operators
- pytomography wrapper SPECT
- SAR noise models
- multiview operators
- single-pixel spyrit wrapper
- spatial unwrapping operators
- multiscale

☀️ Just wrapped up the DeepInverse Hackathon!

We had 30+ imaging scientists from all over the world coding during 3 days next to the beautiful Calanques in Marseille, France. It was a great moment to meet new people, discuss science, and code new imaging algorithms!

Also available on Google Colab colab.research.google.com/drive/11YKc_...

For every step, you can either i) use a preexisting {operator,model,dataset} or ii) define a custom one yourself.

The tutorial, built by @andrewwango.bsky.social, walks you through the main steps of a computational imaging pipeline:

📸 Defining your forward operator
💻 Defining your reconstruction model
🧾 Defining your dataset