Intuitive understanding of multitask learning Can anybody explain what multitask learning is in an intuitive way?  
From Wiki: 

Multitask learning aims to improve the performance of learning algorithms by learning classifiers for multiple tasks jointly. 

What is "multiple tasks"? Can anybody provide an example? How is this related to the structure inside the training data?
 A: I think of MTL as learning two [or more] tasks in parallel, with the constraint that the input data must first pass through a shared information bottleneck before being mapped to prediction values.  This means an MTL algorithm is responsible for learning two things:


*

*a mapping V from the raw input data to the intermediate "bottleneck" representation.

*a mapping W from the intermediate representation to target values for multiple tasks.


The idea is to learn both at the same time (V and W might not even be distinguishable).
The key concept is that V will be forced to learn about what makes the two tasks similar in nature in a way that W can use.  For example, given an image of a face, predicting the location of the eyes can be helpful for predicting the location of the mouth (and vice versa).  Or, given sensor readings from an object at time t-1 and t, predicting the sensor reading from the object at time t+1 can be helpful for predicting the sensor reading from the object at time t+2.
Readings: 


*

*Caruana's "Multitask Learning" is very accessible.

*Langford et al.'s "Learning Nonlinear Dynamical Models" is accessible and gives the flavor of an application.

*Read up on 'Reduced-Rank Regression' (which has been reinvented a few times) which requires some background in linear algebra.


PS - Wiki is not a good place to learn about these concepts from the first time for a variety of reasons; you're much better off Googling for "multitask learning tutorial", "multitask learning introduction" or something like that.
