Warning: mkdir(): No space left on device in /var/www/group-telegram/post.php on line 37

Warning: file_put_contents(aCache/aDaily/post/natural_origin/--): Failed to open stream: No such file or directory in /var/www/group-telegram/post.php on line 50
∅ | Telegram Webview: natural_origin/1436 -
Telegram Group & Telegram Channel
The Geometry of Concepts: Sparse Autoencoder Feature Structure

Abstract:

Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: 1) The "atomic" small-scale structure contains "crystals" whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man-woman-king-queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently done with linear discriminant analysis. 2) The "brain" intermediate-scale structure has significant spatial modularity; for example, math and code features form a "lobe" akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. 3) The "galaxy" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.

https://arxiv.org/abs/2410.19750



group-telegram.com/natural_origin/1436
Create:
Last Update:

The Geometry of Concepts: Sparse Autoencoder Feature Structure

Abstract:

Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models. We find that this concept universe has interesting structure at three levels: 1) The "atomic" small-scale structure contains "crystals" whose faces are parallelograms or trapezoids, generalizing well-known examples such as (man-woman-king-queen). We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently done with linear discriminant analysis. 2) The "brain" intermediate-scale structure has significant spatial modularity; for example, math and code features form a "lobe" akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random. 3) The "galaxy" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. We also quantify how the clustering entropy depends on the layer.

https://arxiv.org/abs/2410.19750

BY ∅




Share with your friend now:
group-telegram.com/natural_origin/1436

View MORE
Open in Telegram


Telegram | DID YOU KNOW?

Date: |

"He has to start being more proactive and to find a real solution to this situation, not stay in standby without interfering. It's a very irresponsible position from the owner of Telegram," she said. On Feb. 27, however, he admitted from his Russian-language account that "Telegram channels are increasingly becoming a source of unverified information related to Ukrainian events." During the operations, Sebi officials seized various records and documents, including 34 mobile phones, six laptops, four desktops, four tablets, two hard drive disks and one pen drive from the custody of these persons. Either way, Durov says that he withdrew his resignation but that he was ousted from his company anyway. Subsequently, control of the company was reportedly handed to oligarchs Alisher Usmanov and Igor Sechin, both allegedly close associates of Russian leader Vladimir Putin. "Like the bombing of the maternity ward in Mariupol," he said, "Even before it hits the news, you see the videos on the Telegram channels."
from us


Telegram
FROM American