https://youvenz.github.io/tags/masked-autoencoder/Recent content in Masked-Autoencoder on Rachid Youven ZeghlacheHugoen-usThu, 01 Jan 2026 00:00:00 +0000https://youvenz.github.io/research/self-supervised-learning/Thu, 01 Jan 2026 00:00:00 +0000https://youvenz.github.io/research/self-supervised-learning/<h2 id="overview">Overview</h2> <p>Supervised deep learning requires large amounts of labelled data. In medical imaging, expert annotations are expensive, time-consuming, and often in short supply. Self-supervised learning (SSL) sidesteps this bottleneck by learning rich representations from <em>unlabelled</em> data through pretext tasks — the labels emerge from the data itself.</p> <h2 id="methods">Methods</h2> <h3 id="masked-autoencoders-mae">Masked Autoencoders (MAE)</h3> <p>Inspired by BERT in NLP, MAE randomly masks a high proportion (75%) of image patches and trains an encoder-decoder to reconstruct the missing regions. The encoder learns spatially rich features without any labels. My L-MAE extends this to temporal sequences of medical images, masking across both space and time.</p>