Why are Sensitive Functions Hard for Transformers?
Empirical studies have identified a range of learnability biases and limitations of transformers, such as a persistent difficulty in learning to compute simple formal languages such as PARITY, and a bias towards low-degree functions. However, theoretical understanding remains limited, with existing expressiveness theory either overpredicting or underpredicting realistic learning abilities. We prove that, under the transformer architecture, the loss landscape is constrained by the input-space sensitivity: Transformers whose output is sensitive to many parts of the input string inhabit isolated points in parameter space, leading to a low-sensitivity bias in generalization. We show theoretically and empirically that this theory unifies a broad array of empirical observations about the learning abilities and biases of transformers, such as their generalization bias towards low sensitivity and low degree, and difficulty in length generalization for PARITY. This shows that understanding transformers' inductive biases requires studying not just their in-principle expressivity, but also their loss landscape..
| Medienart: |
|
|---|
Preprint| Erscheinungsjahr: |
2024 |
|---|---|
| Erschienen: |
2024 |
| Enthalten in: |
arXiv.org - (2024) vom: 15. Feb. Zur Gesamtaufnahme - year:2024 |
|---|
| Sprache: |
Englisch |
|---|
| Beteiligte Personen: |
Hahn, Michael [VerfasserIn] |
|---|
| Links: |
Volltext [kostenfrei] |
|---|
| Themen: |
|---|
| Förderinstitution / Projekttitel: |
|
|---|
| PPN (Katalog-ID): |
XCH042765692 |
|---|
| LEADER | 01000naa a22002652 4500 | ||
|---|---|---|---|
| 001 | XCH042765692 | ||
| 003 | DE-627 | ||
| 005 | 20240306114449.0 | ||
| 007 | cr uuu---uuuuu | ||
| 008 | 240306s2024 xx |||||o 00| ||eng c | ||
| 035 | |a (DE-627)XCH042765692 | ||
| 035 | |a (chemrXiv)2402.09963 | ||
| 040 | |a DE-627 |b ger |c DE-627 |e rakwb | ||
| 041 | |a eng | ||
| 100 | 1 | |a Hahn, Michael |e verfasserin |4 aut | |
| 245 | 1 | 0 | |a Why are Sensitive Functions Hard for Transformers? |
| 264 | 1 | |c 2024 | |
| 336 | |a Text |b txt |2 rdacontent | ||
| 337 | |a Computermedien |b c |2 rdamedia | ||
| 338 | |a Online-Ressource |b cr |2 rdacarrier | ||
| 520 | |a Empirical studies have identified a range of learnability biases and limitations of transformers, such as a persistent difficulty in learning to compute simple formal languages such as PARITY, and a bias towards low-degree functions. However, theoretical understanding remains limited, with existing expressiveness theory either overpredicting or underpredicting realistic learning abilities. We prove that, under the transformer architecture, the loss landscape is constrained by the input-space sensitivity: Transformers whose output is sensitive to many parts of the input string inhabit isolated points in parameter space, leading to a low-sensitivity bias in generalization. We show theoretically and empirically that this theory unifies a broad array of empirical observations about the learning abilities and biases of transformers, such as their generalization bias towards low sensitivity and low degree, and difficulty in length generalization for PARITY. This shows that understanding transformers' inductive biases requires studying not just their in-principle expressivity, but also their loss landscape. | ||
| 650 | 4 | |a Computer Science - Machine Learning |7 (dpeaa)DE-84 | |
| 650 | 4 | |a 000 |7 (dpeaa)DE-84 | |
| 700 | 1 | |a Rofin, Mark |4 aut | |
| 773 | 0 | 8 | |i Enthalten in |t arXiv.org |g (2024) vom: 15. Feb. |
| 773 | 1 | 8 | |g year:2024 |g day:15 |g month:02 |
| 856 | 4 | 0 | |u https://arxiv.org/abs/2402.09963 |z kostenfrei |3 Volltext |
| 912 | |a GBV_XCH | ||
| 951 | |a AR | ||
| 952 | |j 2024 |b 15 |c 02 | ||