An Evolutionary Approach to Dynamic Introduction of Tasks in Large-scale Multitask Learning Systems — mentioned by avinassh

Sun, 29 May 2022 04:34:06 GMT

Jeff Dean responded to OP: (The paper mentioned by OP is https://arxiv.org/abs/2205.12755, and I am one of the two authors, along with Andrea Gesmundo, who did the bulk of the work). The goal of the work was not to get a high quality cifar10 model. Rather, it was to explore a setting where one can dynamically introduce new tasks into a running system and successfully get a high quality model for the new task that reuses representations from the existing model and introduces new parameters somewhat sparingly, while avoiding many of the issues that often plague multi-task systems, such as…

Paper and discussion history. A mention is not an endorsement.

An Evolutionary Approach to Dynamic Introduction of Tasks in Large-scale Multitask Learning Systems — HN submission

Thu, 26 May 2022 22:34:30 GMT

Evolutionary Approach to Introduction of Tasks in Multitask Learning Systems

2 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

An Evolutionary Approach to Dynamic Introduction of Tasks in Large-scale Multitask Learning Systems · HackerPapers discussions

An Evolutionary Approach to Dynamic Introduction of Tasks in Large-scale Multitask Learning Systems — mentioned by avinassh

An Evolutionary Approach to Dynamic Introduction of Tasks in Large-scale Multitask Learning Systems — HN submission