Repository navigation
Expand file tree
/
Copy pathCITATION.cff
More file actions
119 lines (119 loc) · 4.04 KB
/
Copy pathCITATION.cff
File metadata and controls
119 lines (119 loc) · 4.04 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
cff-version: 1.2.0
message: >-
If you use RecSAEI, please cite the associated paper using the preferred citation below.
type: software
title: RecSAEI
abstract: >-
Sparse autoencoders (SAEs) decompose visual representations into individual features,
offering a promising basis for explaining model behavior. Interpreting these features
typically relies on images or pixel regions that most strongly activate them.
However, this evidence becomes ambiguous when multiple features with different
semantics are frequently activated in the same locations, such as texture and
color of the same patterns. We propose Reconstructive SAE Intervention (ResSAEI),
a framework that shifts feature explanation from passive observation to active visual
reconstruction. By intervening on representations along SAE decoder directions and
reconstructing the visual output, we provide direct evidence of a feature's semantic
meaning. We demonstrate that our method successfully disambiguates feature meanings
across multiple datasets and vision encoders. Furthermore, our framework grounds
[CLS] SAE interventions in a sparse set of patch SAE decoder directions.
Beyond stable target response, this attribution also exposes the compositional
relationship between elementary visual patterns and high-level semantics.
Our results advocate for intervention reconstruction as an essential complementary
signal for explaining neural networks.
authors:
- family-names: Gu
given-names: Jianyang
- family-names: Beattie
given-names: Jacob
- family-names: Khurana
given-names: Mridul
- family-names: Chae
given-names: Julia
- family-names: Xiao
given-names: Xi
- family-names: Zhang
given-names: Cheng
- family-names: Karpatne
given-names: Anuj
- family-names: Chao
given-names: Wei-Lun
- family-names: Beery
given-names: Sara
- family-names: Berger-Wolf
given-names: Tanya
repository-code: https://github.com/Imageomics/RecSAEI
url: https://github.com/Imageomics/RecSAEI
version: 1.0.0
license: MIT
keywords:
- sparse autoencoders
- computer vision
- interpretability
- representation autoencoders
- image editing
preferred-citation:
type: article
title: 'From Activation to Intervention: Interpreting Sparse Visual Features through Reconstruction'
journal: 'arXiv preprint arXiv:2611.05522'
year: 2026
url: https://arxiv.org/abs/2611.05522
identifiers:
- type: other
description: arXiv identifier
value: '2611.05522'
authors:
- family-names: Gu
given-names: Jianyang
- family-names: Beattie
given-names: Jacob
- family-names: Khurana
given-names: Mridul
- family-names: Chae
given-names: Julia
- family-names: Xiao
given-names: Xi
- family-names: Zhang
given-names: Cheng
- family-names: Karpatne
given-names: Anuj
- family-names: Chao
given-names: Wei-Lun
- family-names: Beery
given-names: Sara
- family-names: Berger-Wolf
given-names: Tanya
references:
# Software citation from the upstream saev CITATION.cff.
- type: software
title: saev
authors:
- family-names: Stevens
given-names: Samuel
orcid: https://orcid.org/0009-0000-9493-7766
repository-code: https://github.com/Imageomics/saev
url: https://imageomics.github.io/saev/
doi: 10.5281/zenodo.20753477
# Paper citation from the upstream RAEv2 README.
- type: article
title: Improved Baselines with Representation Autoencoders
authors:
- family-names: Singh
given-names: Jaskirat
- family-names: Zheng
given-names: Boyang
- family-names: Wu
given-names: Zongze
- family-names: Zhang
given-names: Richard
- family-names: Shechtman
given-names: Eli
- family-names: Xie
given-names: Saining
journal: arXiv
year: 2026
url: https://arxiv.org/abs/2605.18324
repository-code: https://github.com/nanovisionx/RAEv2
identifiers:
- type: other
description: arXiv identifier
value: '2605.18324'