View in #questions-forum on Slack
@Michael_Milstead: Hey! Reading through the paper on “Progressive Neural Networks” https://arxiv.org/abs/1606.04671, and I’m confused about something:
When the network learns a new task, a new “Column” of nodes is added to the network, including a new input layer for the new task’s input data.
Later, when the network is fed a new sample for inference, how does it determine which input to send the data to? Or is this not actually a goal of the network, and when you want to use the network for a previously learned task, you need to manually use the correct input?
arXiv.org: Progressive Neural Networks
@andcos: Hi Michalel. Apart from the growth in complexity due to the new column, you just highlighted the major disadvantage of Progressive networks. At test time, Progressive network requires test input + task label in order to feed the input to the column associated to the task. This is a very strong (and also quite unrealistic) assumption for CL.
@Michael_Milstead: Ah makes sense, thanks for the reply!
So just to make sure I understand, if the input format was always the same (maybe we want a progressive network that can just learn new image classifications), this wouldn’t be an issue since we could use the same input for all tasks, correct? Just use the same input for all tasks?
Oh nevermind… I just realized that the inputs for the different tasks are all images. For some reason I was imagining that the model could take in multiple types of data (image, audio, etc.). So any image can be sent into the network to get the output of all tasks, and the real problem is determining which output is relevant? Or no?