DLS Course 3 Week-2 Difference between multi-task learning versus multi-class classification

Hi DLS learners and mentors,

I have a question from multi-task learning video lectures. The cost function looks like a multi-class logistic regression classification happening at the output layer.
In logistic regression classification, we take the largest y_predicted value and classify the input belonging to that class.
For example, for 4 neurons in the output layer L, y_predicted[L] = [0.6 0.2 0.1 0.1]
I would have classified it as belonging to Class-A if it were a logistic regression classification.

But how do we interpret y_predicted in case of multi-task learning, given that cost function for both are the same ?


Hey @manojkumarg, welcome to the DLS! :grinning:

Multi-task learning makes sense when multiple classes might be presented in a single training example. In the course video, a single picture may include a pedestrian, a car, a stop sign, or a traffic light. We apply a loss function to each dimension of the output vector to get probabilities for each class being presented on the picture.

If our task were to predict a single category, meaning only a pedestrian, a car, a stop sign, or a traffic light might be presented simultaneously in a training example. It would make more sense just to use a single loss function (such as softmax). We wouldn’t call such a task a multi-task learning task though.

Thanks @manifest for clarifying

1 Like