CS 688

Machine Learning

Extra Credit

Details:

• Each worth up to 10 points (out of 100 total in the class). Note: the total extra credit awarded will be reflective of the quality of the project. Not all extra credit is guaranteed.
• Only one person/team can do each project.
• A team (of up to two people) will have to combine a presentation with a replication project, unless a project explicitly allows a 2-people team.
• To claim a project, email the instructor and the TA, with a ranking of the projects in terms of preference. You will be assigned the highest ranked project that is still available. The process is first-come first-serve, but students that are on-track for an A grade might be given lower priority if there are a lot of interested students.
• If you claim a project but you don't deliver, there will be a penalty of 3 points out of 100 in the class (because you're not allowing another student to claim that extra credit).
• The projects will be presented in class in the three last lectures of the semester.

Available Projects (0)

• [10 points, potentially a 2-person project] Create an interactive demo that simulates active learning, on a task of your choice (in consultation with the instructor). The demo should implement uncertainty and diversity sampling over a given dataset, and it should show (a) examples as they get selected, (b) the model accuracy on a separate development set as we add examples.
  Deliverables: a Github repository with code, in python, that replicates their experiments, and a notebook with an example to run in-class.
• [8 points] Present the findings of this paper: A new learning paradigm: Learning using privileged information.
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.
• [8 points] Present the findings of this paper: Gradients without Backpropagation.
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.
• [8 points] Present the findings of this paper: On Calibration of Modern Neural Networks.
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.
• [10 points] Replicate the findings of this paper: Understanding Deep Learning Requires Re-Thinking Generalization.
  Deliverables: a Github repository with code, in python, that replicates their experiments, and a notebook with an example to run in-class.
• [8 points] Present the findings of this paper: Understanding Deep Learning Requires Re-Thinking Generalization (same as above).
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.
• [10 points] Replicate the findings of this paper: Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets.
  Deliverables: a Github repository with code, in python, that replicates their experiments, and a notebook with an example to run in-class.
• [8 points] Present the findings of this paper: Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets.
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.
• [8 points] Present the findings of this paper: Visualizing the Loss Landscape of Neural Nets.
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.
• [8 points] Present the findings of this paper: Deep Double Descent: Where Bigger Models and More Data Hurt and this blogpost.
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.
• [8 points] Present the findings of this paper: Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time.
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.
• [8 points] Present the findings of this paper: The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks.
  Deliverables: a video presenting the paper and an in-class presentation (8 minutes + 2 for Q&A). Video will be due 3 days before the in-class presentation, in order to ensure that it is appropriate for class.