🐛 Fix description

Author: o-laurent

index.html

@@ -7,10 +7,10 @@
   <!-- Replace the content tag with appropriate information -->
   <meta name="description" content="Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It. Accepted to ICLR 2025.">
   <meta property="og:title" content="Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It" />
-  <meta property="og:description" content="Don't forget about majority voting when you evaluate your TTA method :)" />
+  <meta property="og:description" content="A lead to explain the disastrous effects of label-smoothing on the selective classification abilities of your model." />
   <meta property="og:url" content="https://ensta-u2is-ai.github.io/Understanding-Label-smoothing-Selective-classification/" />
   <!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
-  <meta property="og:image" content="static/images/teaser.png" />
+  <!-- <meta property="og:image" content="static/images/teaser.png" /> -->
   <meta property="og:image:width" content="1200" />
   <meta property="og:image:height" content="630" />
 
@@ -211,97 +211,6 @@ <h2 class="title is-3">Abstract</h2>
   </section>
   <!-- End paper abstract  -->
 
-  <!-- <section class="hero teaser">
-    <div class="container is-max-desktop">
-      <div class="hero-body has-text-centered">
-        <h2 class="title is-3">Takeaways</h2>
-        <div style="display: flex; justify-content: center;">
-          <img src="static/images/teaser.png" alt="Banner Image" height="100%" width="75%" style="margin-bottom: 55px;">
-        </div> -->
-        <!-- <h2 class="subtitle has-text-justified">
-          <br><b>Background on Marginal Entropy Minimization.</b> Test-Time Adaptation aims at adapting a model to a single image at inference time.
-          Ideally, no form of prior or external knowledge should be employed in doing so.
-          An established paradigm for TTA is <b>M</b>arginal <b>E</b>ntropy <b>M</b>inimization, which works by augmenting the image N times, 
-          computing the so-called "marginal probability distribution" (i.e., the average probability distribution over the views), and 
-          minimizing the entropy of this distribution.<br><br>
-
-          <b>Findings.</b> We find that the argmax of the marginal distribution is invariant to <b>MEM</b> most of the time (and can be guaranteed to be so under certain conditions), 
-          and that this marginal distribution itself is reasonably better than standard inference, under the assumption that the model is well-calibrated.
-          <br><br>Empirical evidence for these findings is shown below (left: invariance, right: ensemble verification).
-        </h2>
-        <img src="static/images/I_binned_ent_vs_invariance.png" alt="Banner Image" style="height: 18em; width: auto; padding-inline: 2rem;">
-        <img src="static/images/ensemble_verification_over_datasets.png" alt="Banner Image" style="height: 18em; width: auto; padding-inline: 2rem">
-
-        <h2 class="subtitle has-text-justified">
-          <b>Problem.</b> Calibration is missing on augmented data, but we largely observe that CLIP models are still pretty accurate in this regime.
-          For example, here is what the reliability plots of CLIP-ViT-B-16 look like.
-        </h2>
-        <img src="static/images/rpI.png" alt="Banner Image" style="height: 20em; width: auto;">
-
-        <h2 class="subtitle has-text-justified">
-          <b>TTA with "zero" temperature</b> is a direct consequence of these observations: since confidence information is unreliable, 
-          simply compute the marginal distribution <i>after</i> the temperature has been zeroed-out! By only adapting this parameter, we are effectively marginalizing
-          across one-hot encoded vectors... does this remind you of something?
-        </h2>
-
-      </div>
-    </div>
-  </section>  -->
-  <!-- End paper abstract -->
-
-  <!-- Method overview-->
-  <!-- <section class="section hero is-light">
-    <div class="container is-max-desktop">
-      <h2 class="title is-3 has-text-centered">Implementation</h2>
-      <h2 class="subtitle has-text-centered" style="padding: 0px; margin: 0px">ZERO is implemented in a few lines of code. You can find a PyTorch-like implementation right here :)</h2>
-      <pre class="has-text-justified" style="width: 80rem; overflow-x: auto; padding: 0px; margin: 0px">
-      <code class="python" style="padding: 0px; margin: 0px">
-        def zero(image, z_txt, N, gamma, temp):
-          """
-          :param z_txt: pre-computed text embeddings (C,hdim)
-          :param temp: model’s original temperature
-          :param augment: takes (C,H,W) and returns (N,C,H,W)
-          :param gamma: filtering percentile (e.g., 0.3)
-          """
-          views = augment(image, num_views=N) # generate augmented views
-          l = model.image_encoder(views) @ z_txt.t() # predict (unscaled logits)
-          l_filt = confidence_filter(l, temp, top=gamma) # retain most confident preds
-          zero_temp = torch.finfo(l_filt.dtype).eps # zero temperature
-          p_bar = (l_filt / zero_temp).softmax(dim=1).sum(dim=0) # marginalize
-          return p_bar.argmax()
-      </code>
-      </pre>
-    </div>
-  </section> -->
-  <!-- End method overview -->
-
-    <!-- Results -->
-    <!-- <section class="hero">
-      <div class="container is-max-desktop">
-        <div class="hero-body has-text-centered">
-          <h2 class="title is-3">Results</h2>
-          <h2 class="subtitle has-text-justified">
-            <br>We evaluate ZERO on the standard TTA benchmarks, including robustness to Natural Distribution Shifts and Fine-grained Classification.
-            The results below report CLIP-ViT-B-16 from OpenAI, and compare ZERO to TPT, PromptAlign and RLCF.
-          </h2>
-  
-          <p><b>Robustness to Natural Distribution Shifts</b></p>
-          <img src="static/images/nds.png" alt="Banner Image" style="height: auto; width: 100em;">
-          <br><br>
-
-          <p><b>Fine-grained Classification</b></p>
-          <img src="static/images/fg.png" alt="Banner Image" style="height: auto; width: 100em;">
-
-          <h2 class="subtitle has-text-justified"><br>
-            We find that ZERO, in all its simplicity, establishes a new <b>state-of-the-art</b> in TTA! 
-            Don't forget about majority voting when you evaluate your TTA method!! :)
-          </h2>
-
-        </div>
-      </div>
-    </section> -->
-    <!-- Results -->
-
     <!-- Acknowledgements -->
     <section class="hero is-light">
       <div class="container is-max-desktop">