This was a really interesting paper. Thank you for presenting!
Do you expect XdG and synaptic stabilization to help alleviate catastrophic forgetting even in much larger tasks (larger than the 500 you tested with 90.7% accuracy)? When you test with larger numbers, do you gate a larger number of units to keep the number of unchanged units low? In your task using 500 of the MNIST dataset, what percentage of units did you gate?
This was a really interesting paper. Thank you for presenting!
Do you expect XdG and synaptic stabilization to help alleviate catastrophic forgetting even in much larger tasks (larger than the 500 you tested with 90.7% accuracy)? When you test with larger numbers, do you gate a larger number of units to keep the number of unchanged units low? In your task using 500 of the MNIST dataset, what percentage of units did you gate?