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 -
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.
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.
"We as Ukrainians believe that the truth is on our side, whether it's truth that you're proclaiming about the war and everything else, why would you want to hide it?," he said. The news also helped traders look past another report showing decades-high inflation and shake off some of the volatility from recent sessions. The Bureau of Labor Statistics' February Consumer Price Index (CPI) this week showed another surge in prices even before Russia escalated its attacks in Ukraine. The headline CPI — soaring 7.9% over last year — underscored the sticky inflationary pressures reverberating across the U.S. economy, with everything from groceries to rents and airline fares getting more expensive for everyday consumers. Telegram has become more interventionist over time, and has steadily increased its efforts to shut down these accounts. But this has also meant that the company has also engaged with lawmakers more generally, although it maintains that it doesn’t do so willingly. For instance, in September 2021, Telegram reportedly blocked a chat bot in support of (Putin critic) Alexei Navalny during Russia’s most recent parliamentary elections. Pavel Durov was quoted at the time saying that the company was obliged to follow a “legitimate” law of the land. He added that as Apple and Google both follow the law, to violate it would give both platforms a reason to boot the messenger from its stores. The Dow Jones Industrial Average fell 230 points, or 0.7%. Meanwhile, the S&P 500 and the Nasdaq Composite dropped 1.3% and 2.2%, respectively. All three indexes began the day with gains before selling off. At this point, however, Durov had already been working on Telegram with his brother, and further planned a mobile-first social network with an explicit focus on anti-censorship. Later in April, he told TechCrunch that he had left Russia and had “no plans to go back,” saying that the nation was currently “incompatible with internet business at the moment.” He added later that he was looking for a country that matched his libertarian ideals to base his next startup.
from sa
