thesis/index.html at main · DetoxAI/thesis · GitHub

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
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
<!DOCTYPE html>
<html>
  <head>
    <meta charset="utf-8" />
    <meta name="description" content="DetoxAI is a Python package for debiasing neural networks in image classification tasks." />
    <meta property="og:title" content="DetoxAI: Improving fairness in neural networks via concept unlearning" />
    <meta property="og:description" content="DetoxAI is a Python package for debiasing neural networks in image classification tasks." />
    <meta property="og:url" content="https://github.com/DetoxAI/detoxai" />

    <meta name="twitter:title" content="DetoxAI: Improving fairness in neural networks via concept unlearning" />
    <meta name="twitter:description" content="DetoxAI is a Python package for debiasing neural networks in image classification tasks." />
    <meta name="twitter:card" content="summary" />
    <meta name="keywords" content="artificial intelligence, machine learning, neural networks, image classification, fairness, debiasing" />
    <meta name="viewport" content="width=device-width, initial-scale=1" />

    <title>DetoxAI: Improving fairness in neural networks via concept unlearning</title>
    <link rel="icon" type="image/x-icon" href="static/images/favicon.ico" />
    <link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet" />

    <link rel="stylesheet" href="static/css/bulma.min.css" />
    <link rel="stylesheet" href="static/css/bulma-carousel.min.css" />
    <link rel="stylesheet" href="static/css/bulma-slider.min.css" />
    <link rel="stylesheet" href="static/css/fontawesome.all.min.css" />
    <link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css" />
    <link rel="stylesheet" href="static/css/index.css" />
    <link rel="manifest" href="site.webmanifest" />

    <script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
    <script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
    <script defer src="static/js/fontawesome.all.min.js"></script>
    <script src="static/js/bulma-carousel.min.js"></script>
    <script src="static/js/bulma-slider.min.js"></script>
    <script src="static/js/index.js"></script>
  </head>
  <body>
    <section class="hero">
      <div class="hero-body">
        <div class="container is-max-desktop">
          <div class="columns is-centered">
            <div class="column has-text-centered">
              <div class="publication-title">
                <h1 class="title is-1" style="font-size: 5rem;">
                  <img src="static/images/android-chrome-512x512.png" alt="DetoxAI Logo" style="height: 1.2em; margin-right: 10px; vertical-align: middle; padding: 0; display: inline-block; position: relative; top: -0.1em; border-radius: 10px;">
                  DetoxAI
                </h1>
                <h2 class="subtitle is-3" style="font-weight: normal; margin-top: 0;">
                  Improving fairness in neural networks via concept unlearning
                </h2>
              </div>
              <div class="is-size-5 publication-authors" style="margin-top: 1rem;">
                <span class="author-block">
                  <a href="https://ignacystepka.com" target="_blank">Ignacy Stepka</a><sup>*</sup>
                </span>
                <span class="author-block">
                  <a href="https://lukaszsztukiewicz.com" target="_blank">Lukasz Sztukiewicz</a><sup>*</sup>
                </span>
                <span class="author-block">
                  <a href="https://michal-wilinski.com" target="_blank">Michal Wilinski</a><sup>*</sup>
                </span>
              </div>

              <div class="is-size-5 publication-authors">
                <span class="author-block">
                  Supervisor: <a href="https://www.cs.put.poznan.pl/jstefanowski/" target="_blank">Jerzy Stefanowski</a>
                </span>
              </div>

              <div class="is-size-5 publication-authors">
                <span class="author-block">
                  <span style="display: inline-flex; align-items: center; justify-content: center; margin-bottom: 8px;">
                    <img src="static/images/put_logo.png" alt="PUT Logo" style="height: 1.5em; margin-right: 10px;">
                    <span style="padding-top: 3px; padding-bottom: 3px;">Poznan University of Technology</span>
                  </span>
                  <br>
                </span>
              </div>
              <span class="eql-cntrb is-size-6">
                <small><sup>*</sup>Indicates Equal Contribution</small>
              </span>

              <div style="margin-top: 0.5rem;">
                <span class="is-size-6"><em>Bachelor Thesis submitted in partial fulfillment of the requirements for the degree of Bachelor of Science</em></span>
              </div>

              <div class="publication-links" style="margin-top: 1rem; margin-bottom: 1rem;">
                  <span class="link-block">
                    <a href="static/pdfs/thesis_document.pdf" target="_blank" class="external-link button is-normal is-rounded is-dark">
                      <span class="icon">
                        <i class="fas fa-file-pdf"></i>
                      </span>
                      <span>Thesis</span>
                    </a>
                  </span>

                  <span class="link-block">
                    <a href="https://github.com/DetoxAI/detoxai" target="_blank" class="external-link button is-normal is-rounded is-dark">
                      <span class="icon">
                        <i class="fab fa-github"></i>
                      </span>
                      <span>Code</span>
                    </a>
                  </span>

                  <span class="link-block">
                    <a href="https://detoxai.readthedocs.io" target="_blank" class="external-link button is-normal is-rounded is-dark">
                      <span class="icon">
                        <i class="fas fa-book-open"></i>
                      </span>
                      <span>Documentation</span>
                    </a>
                  </span>
              </div>
            </div>
          </div>
        </div>
      </div>
    </section>

    <section class="hero teaser" style="margin-top: -1.5rem;">
      <div class="container is-max-desktop">
        <div class="hero-body">
          <div class="code-snippet">
            <div class="code-header" style="background-color: #1e1e1e; color: #d4d4d4; padding: 8px 15px; border-top-left-radius: 6px; border-top-right-radius: 6px; display: flex; align-items: center;">
              <span style="margin-right: 10px;"><i class="fas fa-code"></i></span>
              <span>Code Example</span>
            </div>
            <pre class="language-python" style="margin-top: 0; border-top-left-radius: 0; border-top-right-radius: 0;"><code class="language-python"><span style="color: #88846f;"># DetoxAI is available on PyPI, and can be installed by running the following command:</span>
<span style="color: #005eff;">pip install detoxai</span>

<span style="color: #88846f;"># DetoxAI Quickstart Example</span>
<span style="color: #ff3535;">import</span> <span style="color: #005eff;">detoxai</span>

<span style="color: #88846f;"># Load your pretrained model</span>
model = ...

<span style="color: #88846f;"># Prepare a dataloader yielding (input, label, protected attribute)</span>
dataloader = ...

<span style="color: #88846f;"># Debias the model using DetoxAI</span>
corrected = detoxai.<span style="color: #ff3535;">debias</span>(model, dataloader)

<span style="color: #88846f;"># Access metrics from the SAVANIAFT debiasing method</span>
metrics = corrected[<span style="color: #005eff;">"SAVANIAFT"</span>].<span style="color: #ff3535;">get_all_metrics</span>()

<span style="color: #88846f;"># Retrieve the debiased model</span>
model = corrected[<span style="color: #005eff;">"SAVANIAFT"</span>].<span style="color: #ff3535;">get_model</span>()
            </code></pre>
          </div>
          <h2 class="subtitle has-text-centered" style="margin-top: 1.5rem;">
            DetoxAI provides a simple API for debiasing neural networks in image classification tasks.
          </h2>
        </div>
      </div>
    </section>

    <section class="section hero is-light">
      <div class="container is-max-desktop">
        <div class="columns is-centered has-text-centered">
          <div class="column is-four-fifths">
            <h2 class="title is-3">Abstract</h2>
            <div class="content has-text-justified">
              <p>
                Machine learning (ML) systems have achieved remarkable success across numerous domains due to their predictive capabilities. However, their widespread adoption gives rise to many ethical concerns, particularly in light of recent regulations that posit fairness as a legal requirement. While foundational research on fairness has produced effective toolkits for tabular data, these approaches have become prohibitively expensive when adapted to deep learning systems dealing with high-dimensional data such as images. The absence of specialized debiasing methods and proper software frameworks leaves a critical gap in addressing fairness issues for these systems. To address these challenges, we introduce DetoxAI, a Python-based software framework for post-hoc debiasing of neural networks in image classification tasks. Designed with deep learning in mind, DetoxAI integrates state-of-the-art interventions, evaluation metrics, and visualization tools into a unified, production-ready ecosystem. Our approach applies post-training adaptation, allowing users to mitigate bias while maintaining model performance. Focusing on high-level semantic representations, DetoxAI addresses the unique challenges posed by vision data, where protected attributes such as race or gender are not explicitly encoded. This toolkit provides a modular interface for bias mitigation, making it accessible and adaptable for real-world applications. Through experimental studies, we quantitatively demonstrate that DetoxAI reliably improves upon baseline vanilla models on a fairness-performance trade-off. Moreover, with attribution maps, we qualitatively show that DetoxAI's methods can shift the model's focus away from protected attributes.
              </p>
            </div>
          </div>
        </div>
      </div>
    </section>

    <section class="hero is-small">
      <div class="hero-body">
        <div class="container">
          <div class="columns is-centered">
            <div class="column">
              <figure class="image">
                <img src="static/images/side-by-side.webp" alt="Side by side comparison" />
                <h2 class="subtitle has-text-centered">
                  Saliency maps obtained with Layer-wise Relevance Propagation (LRP) showing where the model focuses when classifying images.
                </h2>
              </figure>
            </div>
            <div class="column">
              <figure class="image">
                <img src="static/images/aggregate.webp" alt="Aggregate results" />
                <h2 class="subtitle has-text-centered">
                  Aggregate visualizations (average photos) that we used in our work to show the effect of debiasing.
                </h2>
              </figure>
            </div>
          </div>
        </div>
      </div>
    </section>

    <section class="section" id="BibTeX">
      <div class="container is-max-desktop content">
        <h2 class="title">BibTeX</h2>
        <pre><code>
          @misc{detoxai,
            authors={Ignacy Stepka and Lukasz Sztukiewicz and Michal Wilinski and Jerzy Stefanowski},
            title={DetoxAI: a Python Package for Debiasing Neural Networks},
            year={2025},
          }
        </code></pre>
      </div>
    </section>

    <footer class="footer">
      <div class="container">
        <div class="columns is-centered">
          <div class="column is-8">
            <div class="content">
              <p>
                This page was built using the
                <a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a>
                which was adopted from the <a href="https://nerfies.github.io" target="_blank">Nerfies</a> project page.
                <br />
                This website is licensed under a
                <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">Creative Commons Attribution-ShareAlike 4.0 International License</a>.
              </p>
            </div>
          </div>
        </div>
      </div>
    </footer>

  </body>
</html>