Important Dates

  • Proposal deadline: 1399/1/15
  • Progress report deadline: 1399/3/6
  • Project presentation session: 1399/5/26
  • Final report deadline: 1399/5/28
Please read the Final Project Guidelines carefully, We'll guide you through all details you need.

We have made our students' projects available; you can use them for inspiration or even contact the authors if you are interested.

Featured Projects

  • The Convexity of BERT: From Cause to Solution
    A. Modarressi, H. Mohebbi
    Abstract: Many research has been done on probing and analyzing how BERT works and what are its characteristics. A similar feature within the layerwise results is a convexity in performance, which means the last layers in BERT are showing degradation in probing results. We investigate this fact remains in fine-tuned models by conducting probing tasks....
  • Single Image Super Resolution
    T. Samavati
    Abstract: Single Image Super Resolution (SISR) is an important task of computer vision, in which we seek to improve resolution of images. In recent years, due to the increase in computing power, significant progress has been made in the field of deep learning. Convolutional neural networks form the basis of many machine vision algorithms including SISR....
  • Isotropic ContextualWord Representation in BERT
    S. Rajaee, M. sheikhi
    Abstract: Recent advance in representation learning shows that isotropic (i.e., unit-variance and uncorrelated) embeddings can significantly improve performance on downstream tasks with faster convergence and better generalization. In this project, we proposed an algorithm to apply in both the pre-trained and fine-tuning phase....
  • Toxic comment classification
    K. Rezaee
    Abstract: Traditional methods of learning static cross lingual word embeddings have been relying on various sources of supervision such as bilingual dictionaries, parallel corpus or online Google Translate. To learn contextualized cross-lingual word embeddings, however, we require supervision at context-level rather than token-level....

Other Projects