2026-09-19T21:03:48Zhttps://dspace.mit.edu/server/oai/requestoai:dspace.mit.edu:1721.1/1516602023-08-01T03:01:44Zcom_1721.1_7582com_1721.1_7581col_1721.1_131023
Poggio, Tomaso
Nasimov, Umarbek
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
2023-07-31T19:57:08Z
2023-07-31T19:57:08Z
2023-06
2023-06-06T16:35:02.790Z
https://hdl.handle.net/1721.1/151660
There is a recurring observation in deep learning that neural networks can be combined simply with arithmetic averages over their parameters. This observation has led to many new research directions in model ensembling, meta-learning, federated learning, and optimization. We investigate the evolution of this phenomenon during the training trajectory of neural network models initialized from a common set of parameters (parent). Surprisingly, the benefit of averaging the parameters persists over long child trajectories from parent parameters with minimal training. Furthermore, we find that the parent can be merged with a single child with significant improvement in both training and test loss. Through analysis of the loss landscape, we find that the loss becomes sufficiently convex early on in training, and, as a consequence, models obtained by averaging multiple children often outperform any individual child.
M.Eng.
Massachusetts Institute of Technology
In Copyright - Educational Use Permitted
Copyright retained by author(s)
https://rightsstatements.org/page/InC-EDU/1.0/
How early can we average Neural Networks?
Thesis
application/pdf
Master
Master of Engineering in Electrical Engineering and Computer Science
Publication
unknown
unknown
<Publication xmlns="https://www.openaire.eu/cerif-profile/1.1/" id="f9bc9ff4-d62c-4d62-98f5-be05f444c5e4">
<Type xmlns="https://www.openaire.eu/cerif-profile/vocab/COAR_Publication_Types">http://purl.org/coar/resource_type/c_1843</Type>
<Title>How early can we average Neural Networks?</Title>
<PublishedIn>
<Publication>
</Publication>
</PublishedIn>
<PublicationDate>2023-06</PublicationDate>
<Authors>
<Author>
<DisplayName>Nasimov, Umarbek</DisplayName>
<Affiliation>
<OrgUnit>
</OrgUnit>
</Affiliation>
</Author>
</Authors>
<Editors>
</Editors>
<Publishers>
<Publisher>
<DisplayName>Massachusetts Institute of Technology</DisplayName>
<OrgUnit />
</Publisher>
</Publishers>
<License>https://rightsstatements.org/page/InC-EDU/1.0/</License>
<Abstract>There is a recurring observation in deep learning that neural networks can be combined simply with arithmetic averages over their parameters. This observation has led to many new research directions in model ensembling, meta-learning, federated learning, and optimization. We investigate the evolution of this phenomenon during the training trajectory of neural network models initialized from a common set of parameters (parent). Surprisingly, the benefit of averaging the parameters persists over long child trajectories from parent parameters with minimal training. Furthermore, we find that the parent can be merged with a single child with significant improvement in both training and test loss. Through analysis of the loss landscape, we find that the loss becomes sufficiently convex early on in training, and, as a consequence, models obtained by averaging multiple children often outperform any individual child.</Abstract>
<Access xmlns="http://purl.org/coar/access_right"
>
</Access>
</Publication>