Bastian Grossenbacher-Rieck
@pseudomanifold.topology.rocks
Dad · Geometry ∩ Topology ∩ Machine Learning
Professor at University of Fribourg
While geometry & topology may not save the world, they may well save something that is homotopy-equivalent to it.
🏠 https://bastian.rieck.me/
🏫 https://aidos.group
Professor at University of Fribourg
While geometry & topology may not save the world, they may well save something that is homotopy-equivalent to it.
🏠 https://bastian.rieck.me/
🏫 https://aidos.group
And it just works!
Our pooling methods perform well across tasks and…
🏆 …reach top classification and regression performance.
🔥 …retain this robust performance across pooling ratios.
✨ …preserve graph structure and spectral properties
🧵4/n
Our pooling methods perform well across tasks and…
🏆 …reach top classification and regression performance.
🔥 …retain this robust performance across pooling ratios.
✨ …preserve graph structure and spectral properties
🧵4/n
November 11, 2025 at 3:48 PM
And it just works!
Our pooling methods perform well across tasks and…
🏆 …reach top classification and regression performance.
🔥 …retain this robust performance across pooling ratios.
✨ …preserve graph structure and spectral properties
🧵4/n
Our pooling methods perform well across tasks and…
🏆 …reach top classification and regression performance.
🔥 …retain this robust performance across pooling ratios.
✨ …preserve graph structure and spectral properties
🧵4/n
But…how?
🔍 We contract the most redundant edges that are least relevant for the graph’s structural diversity as measured by the magnitude or spread of a graph.
🧵3/n
🔍 We contract the most redundant edges that are least relevant for the graph’s structural diversity as measured by the magnitude or spread of a graph.
🧵3/n
November 11, 2025 at 3:48 PM
But…how?
🔍 We contract the most redundant edges that are least relevant for the graph’s structural diversity as measured by the magnitude or spread of a graph.
🧵3/n
🔍 We contract the most redundant edges that are least relevant for the graph’s structural diversity as measured by the magnitude or spread of a graph.
🧵3/n
Why do we need structure-aware pooling, anyway?
🔮Our methods, MagEdgePool and SpreadEdgePool, faithfully preserve the original graphs’ geometry.
Alternative pooling layers destroy graph structure to varying extents.
🧵2/n
🔮Our methods, MagEdgePool and SpreadEdgePool, faithfully preserve the original graphs’ geometry.
Alternative pooling layers destroy graph structure to varying extents.
🧵2/n
November 11, 2025 at 3:48 PM
Why do we need structure-aware pooling, anyway?
🔮Our methods, MagEdgePool and SpreadEdgePool, faithfully preserve the original graphs’ geometry.
Alternative pooling layers destroy graph structure to varying extents.
🧵2/n
🔮Our methods, MagEdgePool and SpreadEdgePool, faithfully preserve the original graphs’ geometry.
Alternative pooling layers destroy graph structure to varying extents.
🧵2/n
Big graph, little compute?
In our #NeurIPS2025 paper we propose geometry-aware edge-contraction-based pooling methods for GNNs.
Our methods preserve graph structure, make interpretable pooling choices, and ensure robust performance at downstream tasks.
🧵1/n
In our #NeurIPS2025 paper we propose geometry-aware edge-contraction-based pooling methods for GNNs.
Our methods preserve graph structure, make interpretable pooling choices, and ensure robust performance at downstream tasks.
🧵1/n
November 11, 2025 at 3:48 PM
Big graph, little compute?
In our #NeurIPS2025 paper we propose geometry-aware edge-contraction-based pooling methods for GNNs.
Our methods preserve graph structure, make interpretable pooling choices, and ensure robust performance at downstream tasks.
🧵1/n
In our #NeurIPS2025 paper we propose geometry-aware edge-contraction-based pooling methods for GNNs.
Our methods preserve graph structure, make interpretable pooling choices, and ensure robust performance at downstream tasks.
🧵1/n
As a rule of thumb, we observe that when the mean LID drops and stabilizes, performance is typically improving. By contrast, when the mean LID drops and rises again, this is a warning for overfitting.
🧵3/n
🧵3/n
November 3, 2025 at 3:32 PM
As a rule of thumb, we observe that when the mean LID drops and stabilizes, performance is typically improving. By contrast, when the mean LID drops and rises again, this is a warning for overfitting.
🧵3/n
🧵3/n
It's epoch 50. Do you know whether your LLM is grokking?
In our #NeurIPS2025 paper "Less is More: Local Intrinsic Dimensions of Contextual Language Models," we study how the local intrinsic dimension (LID) of embeddings changes during training and fine-tuning.
#research
🧵1/n
In our #NeurIPS2025 paper "Less is More: Local Intrinsic Dimensions of Contextual Language Models," we study how the local intrinsic dimension (LID) of embeddings changes during training and fine-tuning.
#research
🧵1/n
November 3, 2025 at 3:32 PM
It's epoch 50. Do you know whether your LLM is grokking?
In our #NeurIPS2025 paper "Less is More: Local Intrinsic Dimensions of Contextual Language Models," we study how the local intrinsic dimension (LID) of embeddings changes during training and fine-tuning.
#research
🧵1/n
In our #NeurIPS2025 paper "Less is More: Local Intrinsic Dimensions of Contextual Language Models," we study how the local intrinsic dimension (LID) of embeddings changes during training and fine-tuning.
#research
🧵1/n
It's amazingly fast (< 𝟏𝐡 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐟𝐮𝐥𝐥 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞, 𝐭𝐫𝐚𝐢𝐧𝐞𝐝 𝐨𝐧 𝐚 𝐠𝐚𝐦𝐢𝐧𝐠 𝐥𝐚𝐩𝐭𝐨𝐩) and high-quality.
Another Topological Deep Learning success story, coming soon to #NeurIPS2025!
🖥️ github.com/aidos-lab/in...
📜 arxiv.org/pdf/2410.18987
4/n
Another Topological Deep Learning success story, coming soon to #NeurIPS2025!
🖥️ github.com/aidos-lab/in...
📜 arxiv.org/pdf/2410.18987
4/n
October 29, 2025 at 8:19 AM
It's amazingly fast (< 𝟏𝐡 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐟𝐮𝐥𝐥 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞, 𝐭𝐫𝐚𝐢𝐧𝐞𝐝 𝐨𝐧 𝐚 𝐠𝐚𝐦𝐢𝐧𝐠 𝐥𝐚𝐩𝐭𝐨𝐩) and high-quality.
Another Topological Deep Learning success story, coming soon to #NeurIPS2025!
🖥️ github.com/aidos-lab/in...
📜 arxiv.org/pdf/2410.18987
4/n
Another Topological Deep Learning success story, coming soon to #NeurIPS2025!
🖥️ github.com/aidos-lab/in...
📜 arxiv.org/pdf/2410.18987
4/n
What we gain is a 𝐬𝐭𝐚𝐛𝐥𝐞 and 𝐞𝐱𝐩𝐫𝐞𝐬𝐬𝐢𝐯𝐞 latent space that permits simple interpolation.
3/n
3/n
October 29, 2025 at 8:19 AM
What we gain is a 𝐬𝐭𝐚𝐛𝐥𝐞 and 𝐞𝐱𝐩𝐫𝐞𝐬𝐬𝐢𝐯𝐞 latent space that permits simple interpolation.
3/n
3/n
Want to generate point clouds? Our (w/ Ernst Röell) #NeurIPS2025 paper shows a topology-driven method that is fast, simple, and high-quality. 🔥
1/n
1/n
October 29, 2025 at 8:19 AM
Want to generate point clouds? Our (w/ Ernst Röell) #NeurIPS2025 paper shows a topology-driven method that is fast, simple, and high-quality. 🔥
1/n
1/n
New blog post: Consciousness Is Overrated Anyway (for AI)
In which I, somewhat coherently, try to contribute to the whole "Is AI Conscious" debate.
🔗 bastian.rieck.me/blog/2025/co...
#AI #MachineLearning
In which I, somewhat coherently, try to contribute to the whole "Is AI Conscious" debate.
🔗 bastian.rieck.me/blog/2025/co...
#AI #MachineLearning
August 24, 2025 at 8:50 PM
New blog post: Consciousness Is Overrated Anyway (for AI)
In which I, somewhat coherently, try to contribute to the whole "Is AI Conscious" debate.
🔗 bastian.rieck.me/blog/2025/co...
#AI #MachineLearning
In which I, somewhat coherently, try to contribute to the whole "Is AI Conscious" debate.
🔗 bastian.rieck.me/blog/2025/co...
#AI #MachineLearning
Geometry and topology help us see structure and shape, providing a principled way to imbue models with such concepts…
…and it's not a niche perspective, thanks to a lot of great work by the community!
🧵3/5
…and it's not a niche perspective, thanks to a lot of great work by the community!
🧵3/5
August 7, 2025 at 3:56 PM
Geometry and topology help us see structure and shape, providing a principled way to imbue models with such concepts…
…and it's not a niche perspective, thanks to a lot of great work by the community!
🧵3/5
…and it's not a niche perspective, thanks to a lot of great work by the community!
🧵3/5
We all know the dogma:
"Given enough data, the model will learn everything."
But as LLMs hit diminishing returns (bough with even more GPUs!), this belief is being challenged.
I'm not the only one who posits that inductive biases remain important.
🧵2/5
"Given enough data, the model will learn everything."
But as LLMs hit diminishing returns (bough with even more GPUs!), this belief is being challenged.
I'm not the only one who posits that inductive biases remain important.
🧵2/5
August 7, 2025 at 3:56 PM
We all know the dogma:
"Given enough data, the model will learn everything."
But as LLMs hit diminishing returns (bough with even more GPUs!), this belief is being challenged.
I'm not the only one who posits that inductive biases remain important.
🧵2/5
"Given enough data, the model will learn everything."
But as LLMs hit diminishing returns (bough with even more GPUs!), this belief is being challenged.
I'm not the only one who posits that inductive biases remain important.
🧵2/5
🎥 Shapes, Spaces, Simplices, and Structure: Geometry, Topology & Machine Learning
"What if the answer to some problems in graph learning is not more, but better structure?"
This is the central premise of my talk @logml.bsky.social and @unireps.bsky.social.
🧵1/5
"What if the answer to some problems in graph learning is not more, but better structure?"
This is the central premise of my talk @logml.bsky.social and @unireps.bsky.social.
🧵1/5
August 7, 2025 at 3:56 PM
🎥 Shapes, Spaces, Simplices, and Structure: Geometry, Topology & Machine Learning
"What if the answer to some problems in graph learning is not more, but better structure?"
This is the central premise of my talk @logml.bsky.social and @unireps.bsky.social.
🧵1/5
"What if the answer to some problems in graph learning is not more, but better structure?"
This is the central premise of my talk @logml.bsky.social and @unireps.bsky.social.
🧵1/5
"SAFE TY FIRST"
Kerning. Once you know about it, your enjoyment of public signage will be ruined forever.
en.wikipedia.org/wiki/Kerning
Kerning. Once you know about it, your enjoyment of public signage will be ruined forever.
en.wikipedia.org/wiki/Kerning
August 6, 2025 at 4:04 PM
"SAFE TY FIRST"
Kerning. Once you know about it, your enjoyment of public signage will be ruined forever.
en.wikipedia.org/wiki/Kerning
Kerning. Once you know about it, your enjoyment of public signage will be ruined forever.
en.wikipedia.org/wiki/Kerning
🇨🇭 Happy #SwissNationalDay! It's already a year since I started my new position @unifr.bsky.social and I'v had the good fortune to be living here since 2018. Thanks for everything, Switzerland! 🙏 Here's to the future! 🚀
August 1, 2025 at 12:38 PM
🇨🇭 Happy #SwissNationalDay! It's already a year since I started my new position @unifr.bsky.social and I'v had the good fortune to be living here since 2018. Thanks for everything, Switzerland! 🙏 Here's to the future! 🚀
#MachineLearning loves data. #Topology loves shape.
Let's bring them together!
👉The Euler Characteristic Transform (ECT) bridges both worlds, as I outline in a recent article in the Notices of the
@amermathsoc.bsky.social
Engage!
🧵1/n
Let's bring them together!
👉The Euler Characteristic Transform (ECT) bridges both worlds, as I outline in a recent article in the Notices of the
@amermathsoc.bsky.social
Engage!
🧵1/n
July 22, 2025 at 8:16 AM
#MachineLearning loves data. #Topology loves shape.
Let's bring them together!
👉The Euler Characteristic Transform (ECT) bridges both worlds, as I outline in a recent article in the Notices of the
@amermathsoc.bsky.social
Engage!
🧵1/n
Let's bring them together!
👉The Euler Characteristic Transform (ECT) bridges both worlds, as I outline in a recent article in the Notices of the
@amermathsoc.bsky.social
Engage!
🧵1/n
Thoughts on current events.
![It is [1978 | 2005 | 2025]. The pope just died. I am excited about a new Star Wars Movie.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:iyzbg5xvdcdtquruxwofbczn/bafkreietxehet66luygluksy3seve5xuogxcdlclwwi4zslc2v2zgl7n5m@jpeg)
April 29, 2025 at 5:35 PM
Thoughts on current events.
A candidate for the most audacious correction ever. Replacing a large chunk of the bibliography with "relevant" citations, plus an example of "vegetative electron microscopy." What's going on here?
#AcademicSky
Source: www.mdpi.com/1996-1944/17...
#AcademicSky
Source: www.mdpi.com/1996-1944/17...
April 10, 2025 at 7:31 AM
A candidate for the most audacious correction ever. Replacing a large chunk of the bibliography with "relevant" citations, plus an example of "vegetative electron microscopy." What's going on here?
#AcademicSky
Source: www.mdpi.com/1996-1944/17...
#AcademicSky
Source: www.mdpi.com/1996-1944/17...
Our AIDOS Lab website has been fully revamped, thanks to a massive team effort as part of a very fun retreat in #Bern:
aidos.group
Check out our research directions and some of the things we have been up to these days!
#Academia #MachineLearning
aidos.group
Check out our research directions and some of the things we have been up to these days!
#Academia #MachineLearning
March 1, 2025 at 5:37 PM
Our AIDOS Lab website has been fully revamped, thanks to a massive team effort as part of a very fun retreat in #Bern:
aidos.group
Check out our research directions and some of the things we have been up to these days!
#Academia #MachineLearning
aidos.group
Check out our research directions and some of the things we have been up to these days!
#Academia #MachineLearning
📣 ATMCS 11 will be in beautiful Bozeman, MT. We are still soliciting 3-page extended abstracts (with subsequent full-paper proceedings in La Matematica).
Still open for submissions about applied #topology (broadly interpreted)!
Check out the CfP here:
👉 comptag.github.io/atmcs11/call... 👈
Still open for submissions about applied #topology (broadly interpreted)!
Check out the CfP here:
👉 comptag.github.io/atmcs11/call... 👈
February 24, 2025 at 8:06 AM
📣 ATMCS 11 will be in beautiful Bozeman, MT. We are still soliciting 3-page extended abstracts (with subsequent full-paper proceedings in La Matematica).
Still open for submissions about applied #topology (broadly interpreted)!
Check out the CfP here:
👉 comptag.github.io/atmcs11/call... 👈
Still open for submissions about applied #topology (broadly interpreted)!
Check out the CfP here:
👉 comptag.github.io/atmcs11/call... 👈
🔥 Friends, I finally collected them all. As of last week, I am finally tenured. 🔥
Thanks to everyone who supported me, in particular my great team (aidos.group), my equally great collaborators, and everyone else who brought their offerings to the fickle fates of academia!
#Academia #Tenure
Thanks to everyone who supported me, in particular my great team (aidos.group), my equally great collaborators, and everyone else who brought their offerings to the fickle fates of academia!
#Academia #Tenure
February 7, 2025 at 4:29 PM
🔥 Friends, I finally collected them all. As of last week, I am finally tenured. 🔥
Thanks to everyone who supported me, in particular my great team (aidos.group), my equally great collaborators, and everyone else who brought their offerings to the fickle fates of academia!
#Academia #Tenure
Thanks to everyone who supported me, in particular my great team (aidos.group), my equally great collaborators, and everyone else who brought their offerings to the fickle fates of academia!
#Academia #Tenure
Back from my first JMM (@jointmath.bsky.social). I enjoyed so many interesting talks and discussions. One highlight was a working (!) Enigma exhibit. Plus: Seattle is a beaut!
January 13, 2025 at 9:14 PM
Back from my first JMM (@jointmath.bsky.social). I enjoyed so many interesting talks and discussions. One highlight was a working (!) Enigma exhibit. Plus: Seattle is a beaut!
This is what my OpenReview notifications look like as an AC. (I read through all the e-mails of course but have to acknowledge them in the web interface as well).
Happy Holidays and best of luck for your papers!
@iclr-conf.bsky.social
Happy Holidays and best of luck for your papers!
@iclr-conf.bsky.social
December 29, 2024 at 8:46 AM
This is what my OpenReview notifications look like as an AC. (I read through all the e-mails of course but have to acknowledge them in the web interface as well).
Happy Holidays and best of luck for your papers!
@iclr-conf.bsky.social
Happy Holidays and best of luck for your papers!
@iclr-conf.bsky.social
Hey friends! I am interested in #geometry, #topology, and #MachineLearning. Silent lurker for some time, but now with the recent influx, I'd love to engage more.
If you followed me from the 'other place' but I did not follow you back, let me know! The notifications here are sometimes a bit much...
If you followed me from the 'other place' but I did not follow you back, let me know! The notifications here are sometimes a bit much...
November 20, 2024 at 11:00 AM
Hey friends! I am interested in #geometry, #topology, and #MachineLearning. Silent lurker for some time, but now with the recent influx, I'd love to engage more.
If you followed me from the 'other place' but I did not follow you back, let me know! The notifications here are sometimes a bit much...
If you followed me from the 'other place' but I did not follow you back, let me know! The notifications here are sometimes a bit much...