Spaces:

aah94
/

JEPAMatch

Running

Ali Aghababaei Claude Sonnet 5 commited on 1 day ago

Commit

f5c17be

1 Parent(s): 86def9b

Add paper/code landing page

Static page for JEPAMatch (arXiv:2604.21046): abstract, architecture diagram,
full results tables, convergence/pseudo-label-quality figures, quickstart,
and citation. Links to the code at github.com/aah94/JEPAMatch.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Files changed (3) hide show

README.md +3 -2
index.html +147 -17
style.css +95 -18

README.md CHANGED Viewed

@@ -5,7 +5,8 @@ colorFrom: red
 colorTo: purple
 sdk: static
 pinned: false
-short_description: Geometric Representation Shaping for  SSL
 ---
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

 colorTo: purple
 sdk: static
 pinned: false
+short_description: Geometric Representation Shaping for SSL
 ---
+Paper + code landing page for JEPAMatch. See [index.html](index.html) for the page source, or
+check out the [configuration reference](https://huggingface.co/docs/hub/spaces-config-reference).

index.html CHANGED Viewed

@@ -1,19 +1,149 @@
 <!doctype html>
-<html>
-	<head>
-		<meta charset="utf-8" />
-		<meta name="viewport" content="width=device-width" />
-		<title>My static Space</title>
-		<link rel="stylesheet" href="style.css" />
-	</head>
-	<body>
-		<div class="card">
-			<h1>Welcome to your static Space!</h1>
-			<p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
-			<p>
-				Also don't forget to check the
-				<a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
-			</p>
-		</div>
-	</body>
 </html>

 <!doctype html>
+<html lang="en">
+<head>
+  <meta charset="utf-8" />
+  <meta name="viewport" content="width=device-width, initial-scale=1" />
+  <title>JEPAMatch</title>
+  <meta name="description" content="JEPAMatch: Geometric Representation Shaping for Semi-Supervised Learning" />
+  <link rel="stylesheet" href="style.css" />
+</head>
+<body>
+<main>
+  <header>
+    <h1>JEPAMatch: Geometric Representation Shaping<br>for Semi-Supervised Learning</h1>
+    <div class="authors">Ali Aghababaei-Harandi &middot; Aude Sportisse &middot; Massih-Reza Amini</div>
+    <div class="affil">Universit&eacute; Grenoble Alpes, CNRS, Computer Science Laboratory LIG, Grenoble, France</div>
+    <div class="badges">
+      <a href="https://github.com/aah94/JEPAMatch">Code (GitHub)</a>
+      <a href="https://arxiv.org/abs/2604.21046">Paper (arXiv:2604.21046)</a>
+      <a href="https://github.com/aah94/JEPAMatch/blob/main/LICENSE.txt">License: MIT</a>
+    </div>
+  </header>
+  <p>
+    JEPAMatch combines <strong>FlexMatch</strong>-style adaptive pseudo-labeling with a
+    <strong>LeJEPA</strong>-inspired representation-level objective, so that instead of only
+    thresholding softmax outputs, the model also explicitly shapes its latent space into
+    well-separated, isotropic per-class clusters. This fixes two long-standing FixMatch-family
+    bottlenecks &mdash; majority-class dominance in pseudo-labeling, and the ~2&sup2;&#8304; iteration
+    convergence tax &mdash; cutting the iteration budget by <strong>8&times;</strong> while matching
+    or beating prior SSL methods on CIFAR-100, STL-10, and Tiny-ImageNet.
+  </p>
+  <figure>
+    <img src="https://raw.githubusercontent.com/aah94/JEPAMatch/main/assets/architecture.png" alt="JEPAMatch architecture diagram">
+    <figcaption>A shared backbone feeds two levels: the <strong>Curriculum Level</strong> (top) does
+      FlexMatch-style adaptive pseudo-labeling on weak/strong views; the <strong>Representation
+      Level</strong> (bottom) aligns global and local crops via a JEPA prediction loss, regularized
+      by <strong>Adaptive Class-wise SIGReg</strong> &mdash; per-class isotropic Gaussians instead of
+      one global one, kept apart by an active repulsion term.</figcaption>
+  </figure>
+  <h2>Highlights</h2>
+  <ul class="highlights">
+    <li><strong>8&times; fewer iterations to converge.</strong> JEPAMatch reaches FlexMatch's final
+      CIFAR-100 (400-label) accuracy roughly 50k steps earlier, and finishes training at 2&sup1;&#8311;
+      iterations vs. the standard 2&sup2;&#8304;.</li>
+    <li><strong>State-of-the-art or competitive results</strong> on CIFAR-100, STL-10, and
+      Tiny-ImageNet against FixMatch, FlexMatch, FreeMatch, SoftMatch, CrMatch, SimMatch, Suave,
+      and RegMixMatch.</li>
+    <li><strong>More robust under class imbalance</strong>, and <strong>works as a drop-in
+      module</strong> on top of FlexMatch, FreeMatch, or SoftMatch's curriculum.</li>
+  </ul>
+  <h2>Results</h2>
+  <p>Error rate (%), lower is better, 3 seeds.</p>
+  <div class="table-wrap">
+  <table>
+    <caption>CIFAR-100 &amp; STL-10</caption>
+    <thead>
+      <tr><th>Method</th><th>Iter.</th><th>CIFAR-100 (400)</th><th>CIFAR-100 (2.5K)</th><th>CIFAR-100 (10K)</th><th>STL-10 (40)</th><th>STL-10 (1K)</th></tr>
+    </thead>
+    <tbody>
+      <tr><td>FixMatch</td><td>2&sup2;&#8304;</td><td>46.42 &plusmn; 0.82</td><td>28.03 &plusmn; 0.16</td><td>22.20 &plusmn; 0.12</td><td>35.97 &plusmn; 4.14</td><td>6.25 &plusmn; 0.33</td></tr>
+      <tr><td>FlexMatch</td><td>2&sup2;&#8304;</td><td>39.94 &plusmn; 1.62</td><td>26.49 &plusmn; 0.20</td><td>21.90 &plusmn; 0.15</td><td>29.15 &plusmn; 4.16</td><td>5.77 &plusmn; 0.18</td></tr>
+      <tr><td>FreeMatch</td><td>2&sup2;&#8304;</td><td>37.98 &plusmn; 0.42</td><td>26.47 &plusmn; 0.20</td><td>21.68 &plusmn; 0.03</td><td>15.56 &plusmn; 0.55</td><td>5.63 &plusmn; 0.15</td></tr>
+      <tr><td>SoftMatch</td><td>2&sup2;&#8304;</td><td>37.10 &plusmn; 0.77</td><td>26.66 &plusmn; 0.25</td><td>22.03 &plusmn; 0.03</td><td>21.42 &plusmn; 3.48</td><td>5.73 &plusmn; 0.24</td></tr>
+      <tr><td>CrMatch</td><td>2&sup2;&#8304;</td><td>39.45 &plusmn; 1.69</td><td>25.43 &plusmn; 0.14</td><td>20.40 &plusmn; 0.08</td><td>&ndash;</td><td>4.89 &plusmn; 0.17</td></tr>
+      <tr><td>SimMatch</td><td>2&sup2;&#8304;</td><td>37.81 &plusmn; 2.21</td><td>25.07 &plusmn; 0.32</td><td>20.58 &plusmn; 0.11</td><td>&ndash;</td><td>&ndash;</td></tr>
+      <tr><td>FlatMatch</td><td>2&sup2;&#8304;</td><td>38.76 &plusmn; 1.62</td><td>25.38 &plusmn; 0.85</td><td>19.01 &plusmn; 0.43</td><td>16.20 &plusmn; 4.34</td><td>4.82 &plusmn; 1.21</td></tr>
+      <tr><td>Suave*</td><td>2&sup2;&#8304;</td><td>35.40</td><td>23.00</td><td><strong>18.40</strong></td><td>&ndash;</td><td>&ndash;</td></tr>
+      <tr><td>RegMixMatch*</td><td>2&sup2;&#8304;</td><td>35.27</td><td>23.78</td><td>19.41</td><td><strong>11.74</strong></td><td>4.66</td></tr>
+      <tr class="ours"><td>JEPAMatch (Ours)</td><td>2&sup1;&#8311;</td><td>34.25 &plusmn; 1.97</td><td>22.59 &plusmn; 1.17</td><td>18.55 &plusmn; 0.85</td><td>13.44 &plusmn; 3.2</td><td>4.28 &plusmn; 1.43</td></tr>
+    </tbody>
+  </table>
+  </div>
+  <p style="color:var(--muted); font-size:0.85rem;">*standard deviation not reported in the original paper. Full baseline table (PseudoLabel, MeanTeacher, MixMatch, ReMixMatch, UDA) in the paper.</p>
+  <div class="tables-row">
+    <div class="table-wrap">
+      <table>
+        <caption>Tiny-ImageNet</caption>
+        <thead><tr><th>Method</th><th>Iter.</th><th>1K labels</th><th>10K labels</th></tr></thead>
+        <tbody>
+          <tr><td>FlexMatch</td><td>2&sup1;&#8312;</td><td>41.73</td><td>27.89</td></tr>
+          <tr><td>SoftMatch</td><td>2&sup1;&#8312;</td><td>40.09</td><td>25.92</td></tr>
+          <tr class="ours"><td>JEPAMatch</td><td>2&sup1;&#8312;</td><td>38.82</td><td>24.50</td></tr>
+        </tbody>
+      </table>
+    </div>
+    <div class="table-wrap">
+      <table>
+        <caption>Drop-in module on other curricula (CIFAR-100, 400)</caption>
+        <thead><tr><th>Method</th><th>Iter.</th><th>Error</th></tr></thead>
+        <tbody>
+          <tr><td>FlexMatch</td><td>2&sup2;&#8304;</td><td>50.15 &plusmn; 1.51</td></tr>
+          <tr><td>FreeMatch</td><td>2&sup2;&#8304;</td><td>49.64 &plusmn; 1.46</td></tr>
+          <tr><td>SoftMatch</td><td>2&sup2;&#8304;</td><td>49.24 &plusmn; 2.16</td></tr>
+          <tr class="ours"><td>JEPAMatch (Flex)</td><td>2&sup1;&#8311;</td><td>45.77 &plusmn; 2.77</td></tr>
+          <tr class="ours"><td>JEPAMatch (Free)</td><td>2&sup1;&#8311;</td><td>45.12 &plusmn; 1.98</td></tr>
+          <tr class="ours"><td>JEPAMatch (Soft)</td><td>2&sup1;&#8311;</td><td>44.65 &plusmn; 2.14</td></tr>
+        </tbody>
+      </table>
+    </div>
+  </div>
+  <h2>Convergence speed &amp; pseudo-labeling quality</h2>
+  <figure>
+    <img src="https://raw.githubusercontent.com/aah94/JEPAMatch/main/assets/convergence_speed.png" alt="Convergence speed vs FlexMatch" style="max-width:480px;">
+    <figcaption>JEPAMatch reaches FlexMatch's peak accuracy (CIFAR-100, 4 labels/class) roughly
+      50k iterations earlier, and keeps climbing to a substantially higher final accuracy.</figcaption>
+  </figure>
+  <div class="two-figs">
+    <figure>
+      <img src="https://raw.githubusercontent.com/aah94/JEPAMatch/main/assets/data_utilization.png" alt="Data utilization vs FlexMatch">
+      <figcaption>JEPAMatch keeps more pseudo-labels above the confidence threshold, at higher correctness, than FlexMatch.</figcaption>
+    </figure>
+    <figure>
+      <img src="https://raw.githubusercontent.com/aah94/JEPAMatch/main/assets/max_class_count.png" alt="Class dominance comparison">
+      <figcaption>FlexMatch's majority class can claim up to a quarter of all pseudo-labels in a batch; JEPAMatch keeps this far more balanced.</figcaption>
+    </figure>
+  </div>
+  <h2>Code</h2>
+  <p>Training code, all configs, and the full paper are on
+    <a href="https://github.com/aah94/JEPAMatch">GitHub: aah94/JEPAMatch</a>.</p>
+  <pre><code>git clone https://github.com/aah94/JEPAMatch.git
+cd JEPAMatch
+pip install -r requirements.txt
+python train.py --c config/classic_cv/jepamatch/jepamatch_cifar100_400_0.yaml</code></pre>
+  <h2>Citation</h2>
+  <pre><code>@article{aghababaeiharandi2026jepamatch,
+  title   = {JEPAMatch: Geometric Representation Shaping for Semi-Supervised Learning},
+  author  = {Aghababaei-Harandi, Ali and Sportisse, Aude and Amini, Massih-Reza},
+  journal = {arXiv preprint arXiv:2604.21046},
+  year    = {2026}
+}</code></pre>
+  <footer>
+    Built on the <a href="https://github.com/microsoft/Semi-supervised-learning">USB</a>
+    semi-supervised learning benchmark. This Space is a paper/code landing page &mdash; an
+    interactive inference demo will be added once a released checkpoint is available.
+  </footer>
+</main>
+</body>
 </html>

style.css CHANGED Viewed

@@ -1,28 +1,105 @@
 body {
-	padding: 2rem;
-	font-family: -apple-system, BlinkMacSystemFont, "Arial", sans-serif;
 }
-h1 {
-	font-size: 16px;
-	margin-top: 0;
 }
-p {
-	color: rgb(107, 114, 128);
-	font-size: 15px;
-	margin-bottom: 10px;
-	margin-top: 5px;
 }
-.card {
-	max-width: 620px;
-	margin: 0 auto;
-	padding: 16px;
-	border: 1px solid lightgray;
-	border-radius: 16px;
 }
-.card p:last-child {
-	margin-bottom: 0;
 }

+:root {
+  --ink: #1a1a1a;
+  --muted: #5a5a5a;
+  --line: #e3e3e3;
+  --accent: #7c3aed;
+  --accent-bg: #f5f0ff;
+  --bg: #ffffff;
+  --code-bg: #f6f6f8;
+  --max-width: 880px;
+}
+* { box-sizing: border-box; }
 body {
+  margin: 0;
+  padding: 0 1.25rem 4rem;
+  font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Helvetica, Arial, sans-serif;
+  color: var(--ink);
+  background: var(--bg);
+  line-height: 1.6;
+}
+main { max-width: var(--max-width); margin: 0 auto; }
+header {
+  text-align: center;
+  padding: 3rem 0 1.5rem;
+}
+header h1 {
+  font-size: 2rem;
+  margin: 0 0 0.75rem;
+  line-height: 1.25;
 }
+.authors { font-size: 1.05rem; margin-bottom: 0.25rem; }
+.affil { color: var(--muted); font-size: 0.95rem; margin-bottom: 1.25rem; }
+.badges { display: flex; gap: 0.6rem; justify-content: center; flex-wrap: wrap; margin-bottom: 0.5rem; }
+.badges a {
+  display: inline-block;
+  padding: 0.35rem 0.9rem;
+  border-radius: 999px;
+  background: var(--accent-bg);
+  color: var(--accent);
+  text-decoration: none;
+  font-size: 0.85rem;
+  font-weight: 600;
+  border: 1px solid #e6dbfa;
 }
+.badges a:hover { background: #ece2ff; }
+h2 {
+  font-size: 1.35rem;
+  margin-top: 2.75rem;
+  margin-bottom: 1rem;
+  padding-bottom: 0.4rem;
+  border-bottom: 1px solid var(--line);
 }
+p { margin: 0.75rem 0; }
+figure { margin: 1.5rem 0; text-align: center; }
+figure img { max-width: 100%; border-radius: 8px; }
+figcaption { color: var(--muted); font-size: 0.9rem; margin-top: 0.6rem; text-align: left; }
+.highlights { padding-left: 1.25rem; }
+.highlights li { margin-bottom: 0.5rem; }
+table { border-collapse: collapse; width: 100%; margin: 1rem 0; font-size: 0.92rem; }
+caption { caption-side: top; text-align: left; color: var(--muted); font-size: 0.9rem; margin-bottom: 0.5rem; }
+th, td { padding: 0.5rem 0.7rem; border-bottom: 1px solid var(--line); text-align: center; }
+th:first-child, td:first-child { text-align: left; }
+thead th { border-bottom: 2px solid var(--ink); font-weight: 600; }
+tr.ours td { font-weight: 700; background: var(--accent-bg); }
+.table-wrap { overflow-x: auto; }
+.tables-row { display: flex; gap: 2rem; flex-wrap: wrap; }
+.tables-row > div { flex: 1 1 320px; }
+.two-figs { display: flex; gap: 1rem; flex-wrap: wrap; }
+.two-figs figure { flex: 1 1 320px; margin: 1rem 0; }
+pre {
+  background: var(--code-bg);
+  border: 1px solid var(--line);
+  border-radius: 8px;
+  padding: 1rem 1.2rem;
+  overflow-x: auto;
+  font-size: 0.88rem;
 }
+code { font-family: "SFMono-Regular", Consolas, "Liberation Mono", Menlo, monospace; }
+p code, li code { background: var(--code-bg); padding: 0.1rem 0.35rem; border-radius: 4px; }
+footer {
+  text-align: center;
+  color: var(--muted);
+  font-size: 0.85rem;
+  margin-top: 3.5rem;
+  padding-top: 1.5rem;
+  border-top: 1px solid var(--line);
 }
+footer a { color: var(--accent); }
+a { color: var(--accent); }