Here is the link to the paper: Caruana, Rich. "Multitask learning." Machine learning 28.1 (1997): 41-75. https://t.co/y9gq7dAnEh
key insight #MultitaskLearning: "Often features become available after predictions are made. These features cant be used as inputs as they aren't available at runtime. But they can be collected and used as tasks to provide extra info during training." htt
In multi task learning (MTL) one model is trained for multiple tasks with same input, instead of training separate model for each task . This paper contains very good explanations of MTL with 3 layer networks. Our brain probably l…https://t.co/HSK2iwKGan
@sleepinyourhat This is a good source in general, although it's not specifically a study on MTL for different task dataset sizes https://t.co/JD2CQr30V7 ...I personally prefer to weight each loss in the batch