A sparse null code emerges in deep neural networks

Brian S Robinson, Nathan Drenkow, Colin Conwell, Michael Bonner
Proceedings of UniReps: the First Workshop on Unifying Representations in Neural Models, PMLR 243:302-314, 2024.

Abstract

The internal representations of deep vision models are often assumed to encode specific image features, such as contours, textures, and object parts. However, it is possible for deep networks to learn highly abstract representations that may not be linked to any specific image feature. Here we present evidence for one such abstract representation in transformers and modern convolutional architectures that appears to serve as a null code, indicating image regions that are non-diagnostic for the object class. These null codes are both statistically and qualitatively distinct from the more commonly reported feature-related codes of vision models. Specifically, these null codes have several distinct characteristics: they are highly sparse, they have a single unique activation pattern for each network, they emerge abruptly at intermediate network depths, and they are activated in a feature-independent manner by weakly informative image regions, such as backgrounds. Together, these findings reveal a new class of highly abstract representations in deep vision models: sparse null codes that seem to indicate the absence of relevant features.

Cite this Paper


BibTeX
@InProceedings{pmlr-v243-robinson24a, title = {A sparse null code emerges in deep neural networks}, author = {Robinson, Brian S and Drenkow, Nathan and Conwell, Colin and Bonner, Michael}, booktitle = {Proceedings of UniReps: the First Workshop on Unifying Representations in Neural Models}, pages = {302--314}, year = {2024}, editor = {Fumero, Marco and Rodolá, Emanuele and Domine, Clementine and Locatello, Francesco and Dziugaite, Karolina and Mathilde, Caron}, volume = {243}, series = {Proceedings of Machine Learning Research}, month = {15 Dec}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v243/robinson24a/robinson24a.pdf}, url = {https://proceedings.mlr.press/v243/robinson24a.html}, abstract = {The internal representations of deep vision models are often assumed to encode specific image features, such as contours, textures, and object parts. However, it is possible for deep networks to learn highly abstract representations that may not be linked to any specific image feature. Here we present evidence for one such abstract representation in transformers and modern convolutional architectures that appears to serve as a null code, indicating image regions that are non-diagnostic for the object class. These null codes are both statistically and qualitatively distinct from the more commonly reported feature-related codes of vision models. Specifically, these null codes have several distinct characteristics: they are highly sparse, they have a single unique activation pattern for each network, they emerge abruptly at intermediate network depths, and they are activated in a feature-independent manner by weakly informative image regions, such as backgrounds. Together, these findings reveal a new class of highly abstract representations in deep vision models: sparse null codes that seem to indicate the absence of relevant features.} }
Endnote
%0 Conference Paper %T A sparse null code emerges in deep neural networks %A Brian S Robinson %A Nathan Drenkow %A Colin Conwell %A Michael Bonner %B Proceedings of UniReps: the First Workshop on Unifying Representations in Neural Models %C Proceedings of Machine Learning Research %D 2024 %E Marco Fumero %E Emanuele Rodolá %E Clementine Domine %E Francesco Locatello %E Karolina Dziugaite %E Caron Mathilde %F pmlr-v243-robinson24a %I PMLR %P 302--314 %U https://proceedings.mlr.press/v243/robinson24a.html %V 243 %X The internal representations of deep vision models are often assumed to encode specific image features, such as contours, textures, and object parts. However, it is possible for deep networks to learn highly abstract representations that may not be linked to any specific image feature. Here we present evidence for one such abstract representation in transformers and modern convolutional architectures that appears to serve as a null code, indicating image regions that are non-diagnostic for the object class. These null codes are both statistically and qualitatively distinct from the more commonly reported feature-related codes of vision models. Specifically, these null codes have several distinct characteristics: they are highly sparse, they have a single unique activation pattern for each network, they emerge abruptly at intermediate network depths, and they are activated in a feature-independent manner by weakly informative image regions, such as backgrounds. Together, these findings reveal a new class of highly abstract representations in deep vision models: sparse null codes that seem to indicate the absence of relevant features.
APA
Robinson, B.S., Drenkow, N., Conwell, C. & Bonner, M.. (2024). A sparse null code emerges in deep neural networks. Proceedings of UniReps: the First Workshop on Unifying Representations in Neural Models, in Proceedings of Machine Learning Research 243:302-314 Available from https://proceedings.mlr.press/v243/robinson24a.html.

Related Material