new post: tut

Author: f14-bertolotti

index.html

@@ -24,6 +24,23 @@ <h1>
             Hello and welcome to my blog! As a postdoctoral researcher in machine learning and artificial intelligence, I’m here to share my journey, insights, and notes with you. I strive to be as clear and grounded as possible, and I hope you’ll find something interesting and valuable in my posts.
         </p>
 
+        <div onclick="location.href='posts/06-03-25-tut/index.html';" style="cursor: pointer;">
+        <article class="light-post-theme post-block"> 
+            <div class="post-inner">
+                <div class="post-title"> 
+                    <img src="icons/pin.svg" width=16px height=16px/> Mid-Training Untying: A Fix That Barely Fixes Anything 
+                </div>
+                <div class="post-desc"> 
+                    In one of the previous posts, we have seen how tying embeddings can be destabilize the training if the the data do not satisfy certain assumptions (see <a href="https://openreview.net/forum?id=yyYMAprcAR">here</a>). In this post, we will explore a simple idea to get the best of both worlds: early training boost with tied embeddings and late training stability with untied one. This was a research idea that I had in mind however it did not work as well as expected so I decided to share it here. 
+                </div>
+                <div class="post-foot"> 
+                    <img src="icons/clock.svg" width=12px height=12px/> Date: 06 March, 2025  |  Estimated Reading Time: ~5 min 
+                </div>
+            </div>
+        </article>
+        </div>
+        <br>
+
         <div onclick="location.href='posts/29-12-24-chaos-II/index.html';" style="cursor: pointer;">
         <article class="light-post-theme post-block"> 
             <div class="post-inner">