Skip to content

Commit bbfe9fe

Browse files
committed
neurips25
1 parent 763e038 commit bbfe9fe

File tree

11 files changed

+318
-225
lines changed

11 files changed

+318
-225
lines changed

_bibliography/preprints.bib

Lines changed: 0 additions & 18 deletions
Original file line numberDiff line numberDiff line change
@@ -1,14 +1,5 @@
11
---
22
---
3-
@misc{zhou2025payattentionsmallweights,
4-
title={Pay Attention to Small Weights},
5-
author={Chao Zhou and Tom Jacobs and Advait Gadhikar and Rebekka Burkholz},
6-
year={2025},
7-
eprint={2506.21374},
8-
archivePrefix={arXiv},
9-
primaryClass={cs.LG},
10-
url={https://arxiv.org/abs/2506.21374},
11-
}
123
@misc{jacobs2025hamhyperbolicstepregulate,
134
title={HAM: A Hyperbolic Step to Regulate Implicit Bias},
145
author={Tom Jacobs and Advait Gadhikar and Celia Rubio-Madrigal and Rebekka Burkholz},
@@ -27,15 +18,6 @@ @misc{reddy2025shifthappensconfounding
2718
primaryClass={cs.LG},
2819
url={https://arxiv.org/abs/2505.21422},
2920
}
30-
@misc{gadhikar2025signinlotteryreparameterizingsparse,
31-
title={Sign-In to the Lottery: Reparameterizing Sparse Training From Scratch},
32-
author={Advait Gadhikar and Tom Jacobs and Chao Zhou and Rebekka Burkholz},
33-
year={2025},
34-
eprint={2504.12801},
35-
archivePrefix={arXiv},
36-
primaryClass={cs.LG},
37-
url={https://arxiv.org/abs/2504.12801},
38-
}
3921
@misc{gadhikar2024cyclicsparsetrainingenough,
4022
title={Cyclic Sparse Training: Is it Enough?},
4123
author={Advait Gadhikar and Sree Harsha Nelaturu and Rebekka Burkholz},

_bibliography/references.bib

Lines changed: 31 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,36 @@
11
---
22
---
3+
@inproceedings{ zhou2025payattentionsmallweights,
4+
title={Pay Attention to Small Weights},
5+
author={Chao Zhou and Tom Jacobs and Advait Gadhikar and Rebekka Burkholz},
6+
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
7+
year={2025},
8+
url={https://openreview.net/forum?id=XKnOA7MhCz},
9+
pdf={https://openreview.net/pdf?id=XKnOA7MhCz},
10+
abstract={Finetuning large pretrained neural networks is known to be resource-intensive, both in terms of memory and computational cost. To mitigate this, a common approach is to restrict training to a subset of the model parameters. By analyzing the relationship between gradients and weights during finetuning, we observe a notable pattern: large gradients are often associated with small-magnitude weights. This correlation is more pronounced in finetuning settings than in training from scratch. Motivated by this observation, we propose NANOADAM, which dynamically updates only the small-magnitude weights during finetuning and offers several practical advantages: first, this criterion is gradient-free -- the parameter subset can be determined without gradient computation; second, it preserves large-magnitude weights, which are likely to encode critical features learned during pretraining, thereby reducing the risk of catastrophic forgetting; thirdly, it permits the use of larger learning rates and consistently leads to better generalization performance in experiments. We demonstrate this for both NLP and vision tasks.},
11+
}
12+
13+
@inproceedings{ Gadhikar2025SignInTT,
14+
title={Sign-In to the Lottery: Reparameterizing Sparse Training},
15+
author={Advait Gadhikar and Tom Jacobs and Chao Zhou and Rebekka Burkholz},
16+
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
17+
year={2025},
18+
url={https://openreview.net/forum?id=iwKT7MEZZw},
19+
pdf={https://openreview.net/pdf?id=iwKT7MEZZw},
20+
abstract={The performance gap between training sparse neural networks from scratch (PaI) and dense-to-sparse training presents a major roadblock for efficient deep learning. According to the Lottery Ticket Hypothesis, PaI hinges on finding a problem specific parameter initialization. As we show, to this end, determining correct parameter signs is sufficient. Yet, they remain elusive to PaI. To address this issue, we propose Sign-In, which employs a dynamic reparameterization that provably induces sign flips. Such sign flips are complementary to the ones that dense-to-sparse training can accomplish, rendering Sign-In as an orthogonal method. While our experiments and theory suggest performance improvements of PaI, they also carve out the main open challenge to close the gap between PaI and dense-to-sparse training.},
21+
img={signin.jpg}
22+
}
23+
24+
@inproceedings{ pham2025the,
25+
title={The Graphon Limit Hypothesis: Understanding Neural Network Pruning via Infinite Width Analysis},
26+
author={Hoang Pham and The-Anh Ta and Tom Jacobs and Rebekka Burkholz and Long Tran-Thanh},
27+
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
28+
year={2025},
29+
url={https://openreview.net/forum?id=EEZLBhyer1},
30+
pdf={https://openreview.net/pdf?id=EEZLBhyer1},
31+
abstract={Sparse neural networks promise efficiency, yet training them effectively remains a fundamental challenge. Despite advances in pruning methods that create sparse architectures, understanding why some sparse structures are better trainable than others with the same level of sparsity remains poorly understood. Aiming to develop a systematic approach to this fundamental problem, we propose a novel theoretical framework based on the theory of graph limits, particularly graphons, that characterizes sparse neural networks in the infinite-width regime. Our key insight is that connectivity patterns of sparse neural networks induced by pruning methods converge to specific graphons as networks' width tends to infinity, which encodes implicit structural biases of different pruning methods. We postulate the Graphon Limit Hypothesis and provide empirical evidence to support it. Leveraging this graphon representation, we derive a Graphon Neural Tangent Kernel (Graphon NTK) to study the training dynamics of sparse networks in the infinite width limit. Graphon NTK provides a general framework for the theoretical analysis of sparse networks. We empirically show that the spectral analysis of Graphon NTK correlates with observed training dynamics of sparse networks, explaining the varying convergence behaviours of different pruning methods. Our framework provides theoretical insights into the impact of connectivity patterns on the trainability of various sparse network architectures.},
32+
}
33+
334
@inproceedings{jacobs2025mirror,
435
title={Mirror, Mirror of the Flow: How Does Regularization Shape Implicit Bias?},
536
author={Tom Jacobs and Chao Zhou and Rebekka Burkholz},

_data/alumni_members.yml

Lines changed: 3 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -14,15 +14,15 @@
1414
end_date: Oct 24
1515
email: adarsh.jamadandi@cispa.de
1616
url: https://adarshmj.github.io
17-
next: PhD at IRISA, Université de Rennes
17+
next: PhD candidate at IRISA Rennes
1818

1919
- name: Harsha Nelaturu
2020
last_name: Nelaturu
2121
photo: c02hane.jpg
2222
start_date: Aug 23
2323
end_date: Jul 24
2424
url: https://nelaturuharsha.github.io/
25-
next: Applied Scientist Intern at Amazon
25+
next: PhD candidate at Zuse Institute Berlin
2626

2727
- name: Advait Athreya
2828
last_name: Athreya
@@ -45,7 +45,7 @@
4545
members:
4646
- name: Mohammed Adnan
4747
last_name: Adnan
48-
from: University of Calgary
48+
from: University of Calgary and Vector Institute
4949
start_date: Oct 25
5050
end_date: Nov 25
5151
url: https://adnan1306.github.io

_data/news.yml

Lines changed: 12 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,3 +1,13 @@
1+
- date: 10. November 2025
2+
headline: "Celia and Tom are presenting at the Workshop on Geometry, Topology, and Machine Learning ([GTML](https://www.mis.mpg.de/events/series/workshop-on-geometry-topology-and-machine-learning-gtml-2025)) in Leipzig."
3+
4+
- date: 18. September 2025
5+
headline: "Three papers
6+
[(1)](https://openreview.net/forum?id=XKnOA7MhCz)
7+
[(2)](https://openreview.net/forum?id=iwKT7MEZZw)
8+
[(3)](https://openreview.net/forum?id=EEZLBhyer1)
9+
have been accepted at NeurIPS 2025, and [one](https://sites.google.com/wimlworkshop.org/wimlworkshopneurips2025/program) at the WiML Workshop."
10+
111
- date: 15. September 2025
212
headline: "Rebekka and Celia are presenting at the Workshop on Mining and Learning with Graphs ([MLG](https://mlg-europe.github.io/2025/)) in Porto with a keynote and two posters, respectively."
313

@@ -38,7 +48,7 @@
3848
headline: "Welcome to Gowtham and Nik!"
3949

4050
- date: 25. December 2024
41-
headline: "Celia, Adarsh, Rebekka, and Nimrah are presenting their work at the [ELLIS Pre-NeurIPS](https://lacoco-lab.github.io/events/PreNeurIPS2024/) Session in Saarbrücken."
51+
headline: "Celia, Adarsh, Nimrah, and Rebekka are presenting their work at the [ELLIS Pre-NeurIPS](https://lacoco-lab.github.io/events/PreNeurIPS2024/) Session in Saarbrücken."
4252

4353
- date: 25. September 2024
4454
headline: "Three papers
@@ -51,7 +61,7 @@
5161
headline: "Welcome to Chao, Rahul, and Dong!"
5262

5363
- date: 14. June 2024
54-
headline: "Celia, Advait, and Adarsh are presenting at the Helmholtz AI Conference: AI for Science ([HAICON](https://eventclass.it/haic2024/scientific/external-program/session?s=S-05a)) in Düsseldorf ([video](/outreach#celia-rubio-madrigal--haicon-jun-14-2024))."
64+
headline: "Advait, Celia, and Adarsh are presenting at the Helmholtz AI Conference: AI for Science ([HAICON](https://eventclass.it/haic2024/scientific/external-program/session?s=S-05a)) in Düsseldorf ([video](/outreach#celia-rubio-madrigal--haicon-jun-14-2024))."
5565

5666
- date: 1. May 2024
5767
headline: "Our paper on [improving GATs](https://openreview.net/forum?id=Sjv5RcqfuH) has been accepted at ICML 2024."

_pages/publications.md

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -16,12 +16,12 @@ Explore our latest research by browsing **[pre-prints](#pre-prints)** and **[acc
1616
{% endfor %}{% endfor %}</select>
1717

1818

19-
## Pre-prints
19+
## Accepted papers
2020

21-
{% bibliography --file preprints %}
21+
{% bibliography %}
2222

2323
<hr>
2424

25-
## Accepted papers
25+
## Pre-prints
2626

27-
{% bibliography %}
27+
{% bibliography --file preprints %}
29.5 KB
Loading

_site/index.html

Lines changed: 10 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -130,28 +130,28 @@ <h1 id="relational-machine-learning-lab">Relational Machine Learning Lab</h1>
130130
<h3>News</h3>
131131
<div class="well">
132132

133-
<b>15 Sep 2025</b>
134-
<p>Rebekka and Celia are presenting at the Workshop on Mining and Learning with Graphs (<a href="https://mlg-europe.github.io/2025/">MLG</a>) in Porto with a keynote and two posters, respectively.</p>
133+
<b>10 Nov 2025</b>
134+
<p>Celia and Tom are presenting at the Workshop on Geometry, Topology, and Machine Learning (<a href="https://www.mis.mpg.de/events/series/workshop-on-geometry-topology-and-machine-learning-gtml-2025">GTML</a>) in Leipzig.</p>
135135

136136
<hr/>
137137

138-
<b>14 Aug 2025</b>
139-
<p>Tom is <a href="https://cohere.com/events/Cohere-Labs-Tom-Jacobs-2025">presenting</a> his work on implicit regularization at Cohere Labs: Open Science Community (<a href="/outreach#tom-jacobs--cohere-labs-aug-14-2025">video</a>).</p>
138+
<b>18 Sep 2025</b>
139+
<p>Three papers <a href="https://openreview.net/forum?id=XKnOA7MhCz">(1)</a> <a href="https://openreview.net/forum?id=iwKT7MEZZw">(2)</a> <a href="https://openreview.net/forum?id=EEZLBhyer1">(3)</a> have been accepted at NeurIPS 2025, and <a href="https://sites.google.com/wimlworkshop.org/wimlworkshopneurips2025/program">one</a> at the WiML Workshop.</p>
140140

141141
<hr/>
142142

143-
<b>12 Jun 2025</b>
144-
<p>Tom is attending the AI &amp; Mathematics Workshop (<a href="https://aimath.nl/index.php/2025/03/13/4th-aim-cluster-event-tilburg/">AIM</a>) at Tilburg University.</p>
143+
<b>15 Sep 2025</b>
144+
<p>Rebekka and Celia are presenting at the Workshop on Mining and Learning with Graphs (<a href="https://mlg-europe.github.io/2025/">MLG</a>) in Porto with a keynote and two posters, respectively.</p>
145145

146146
<hr/>
147147

148-
<b>02 Jun 2025</b>
149-
<p>Rebekka and Celia are presenting at <a href="https://netsci2025.github.io/">NetSci</a> in Maastricht with a <a href="/outreach#netbiomed-2025-keynote-jun-2-2025">satellite keynote</a> and a lightning talk.</p>
148+
<b>14 Aug 2025</b>
149+
<p>Tom is <a href="https://cohere.com/events/Cohere-Labs-Tom-Jacobs-2025">presenting</a> his work on implicit regularization at Cohere Labs: Open Science Community (<a href="/outreach#tom-jacobs--cohere-labs-aug-14-2025">video</a>).</p>
150150

151151
<hr/>
152152

153-
<b>27 May 2025</b>
154-
<p>Celia is visiting TU Wien and presenting at the ML Research Unit’s <a href="https://ml-tuw.github.io/ruml_seminar/">seminar</a>.</p>
153+
<b>12 Jun 2025</b>
154+
<p>Tom is attending the AI &amp; Mathematics Workshop (<a href="https://aimath.nl/index.php/2025/03/13/4th-aim-cluster-event-tilburg/">AIM</a>) at Tilburg University.</p>
155155

156156
<hr/>
157157

_site/news.html

Lines changed: 12 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -77,6 +77,16 @@ <h1 id="news">News</h1>
7777

7878
<hr />
7979

80+
<h4>10 November 2025</h4>
81+
<p>Celia and Tom are presenting at the Workshop on Geometry, Topology, and Machine Learning (<a href="https://www.mis.mpg.de/events/series/workshop-on-geometry-topology-and-machine-learning-gtml-2025">GTML</a>) in Leipzig.</p>
82+
83+
<hr />
84+
85+
<h4>18 September 2025</h4>
86+
<p>Three papers <a href="https://openreview.net/forum?id=XKnOA7MhCz">(1)</a> <a href="https://openreview.net/forum?id=iwKT7MEZZw">(2)</a> <a href="https://openreview.net/forum?id=EEZLBhyer1">(3)</a> have been accepted at NeurIPS 2025, and <a href="https://sites.google.com/wimlworkshop.org/wimlworkshopneurips2025/program">one</a> at the WiML Workshop.</p>
87+
88+
<hr />
89+
8090
<h4>15 September 2025</h4>
8191
<p>Rebekka and Celia are presenting at the Workshop on Mining and Learning with Graphs (<a href="https://mlg-europe.github.io/2025/">MLG</a>) in Porto with a keynote and two posters, respectively.</p>
8292

@@ -138,7 +148,7 @@ <h4>01 December 2024</h4>
138148
<hr />
139149

140150
<h4>25 December 2024</h4>
141-
<p>Celia, Adarsh, Rebekka, and Nimrah are presenting their work at the <a href="https://lacoco-lab.github.io/events/PreNeurIPS2024/">ELLIS Pre-NeurIPS</a> Session in Saarbrücken.</p>
151+
<p>Celia, Adarsh, Nimrah, and Rebekka are presenting their work at the <a href="https://lacoco-lab.github.io/events/PreNeurIPS2024/">ELLIS Pre-NeurIPS</a> Session in Saarbrücken.</p>
142152

143153
<hr />
144154

@@ -153,7 +163,7 @@ <h4>01 July 2024</h4>
153163
<hr />
154164

155165
<h4>14 June 2024</h4>
156-
<p>Celia, Advait, and Adarsh are presenting at the Helmholtz AI Conference: AI for Science (<a href="https://eventclass.it/haic2024/scientific/external-program/session?s=S-05a">HAICON</a>) in Düsseldorf (<a href="/outreach#celia-rubio-madrigal--haicon-jun-14-2024">video</a>).</p>
166+
<p>Advait, Celia, and Adarsh are presenting at the Helmholtz AI Conference: AI for Science (<a href="https://eventclass.it/haic2024/scientific/external-program/session?s=S-05a">HAICON</a>) in Düsseldorf (<a href="/outreach#celia-rubio-madrigal--haicon-jun-14-2024">video</a>).</p>
157167

158168
<hr />
159169

0 commit comments

Comments
 (0)