Lex Fridman and Yann LeCun discuss the peer review process

TL;DR

Researchers discuss the rejection of their co-authored paper and the benefits and limitations of the review process in computer science conferences.

Transcript

you posted the paper you co-authored was rejected from europe's as you said proudly in quotes rejected can you joke yeah i know can you describe this paper and like what was the idea in it and also maybe this is a good opportunity to ask what are the pros and cons what works and what doesn't about the review process yeah let me talk about the paper... Read More

Key Insights

• 👻 Joint embedding architecture allows for the production of multiple predictions in supervised learning models.
• 🏛️ Non-contrastive joint embedding methods provide a promising approach for building predictive models and learning hierarchical representations.
• 🖤 The peer review process in computer science conferences is plagued by biases, a focus on finding flaws, and a lack of appreciation for new ideas.
• 🤗 Alternative systems, such as open reviews and collective recommender systems, may enhance the reviewing process and incentivize researchers to provide detailed evaluations.
• 👨‍🔬 The current system prioritizes fairness and accurate credit allocation but may impede the progress of scientific research.
• 💡 Shifting to a more innovative and progressive review process is crucial for the advancement of science and effective communication of ideas.
• 🤗 Open access repositories like arXiv and platforms like Twitter and archive sanity already enable continuous conferences and decentralized reviewing.
• ❓ Implementing reputation systems for reviewers and evaluating the predictive success of their evaluations could incentivize thorough and thoughtful reviews.

Questions & Answers

Q: Can you explain the concept of joint embedding architecture and its role in supervised learning?

Joint embedding architecture involves training a system to predict the continuation of video clips by using a hidden vector or latent variable. This approach enables the system to produce multiple plausible predictions, creating informative representations of the video clips.

Q: What are the pros and cons of the two methods for handling multi-modality in predictions?

The first method utilizes latent variables and predicts pixels directly, while the second method predicts abstract representations of pixels. The second method eliminates irrelevant details and focuses on preserving meaningful information about the input. However, it is still a matter of debate and research to determine the exact reasons for the effectiveness of the second method.

Q: What is the purpose of the V-Craig paper, and how does it differ from Barlow Twins?

V-Craig builds upon the Barlow Twins paper, addressing some of its shortcomings and making improvements. While some reviewers criticized V-Craig for not being different enough from Barlow Twins, the researchers emphasize that V-Craig is a practical method that will likely be used by researchers in the field.

Q: What are the flaws associated with the current peer review process in computer science conferences?

The peer review process in computer science conferences can be biased, driven by self-interest, and focused on finding flaws in papers rather than appreciating new ideas. Additionally, the field's exponential growth has led to a significant number of inexperienced and junior reviewers. These factors hinder the acceptance of papers that introduce new concepts and advances.

Summary & Key Takeaways

• The co-authored paper, titled "V-Craig," focuses on variance covariance regularization and its application in joint embedding architecture for non-contrastive learning techniques.

• The paper explores the handling of uncertainty in predictive models using latent variables and representations of video clips.

• The researchers discuss their shift in preference towards non-contrastive joint embedding methods and their potential for developing predictive models and hierarchical representations.