Machine learning
When does a larger learning rate stop helping?
Hold the training budget fixed and discover when a larger gradient-descent step converges, oscillates, or grows.
What breaks when backpropagation forgets a path?
Omit one backward contribution, compare against a numerical gradient check, and design a case that could hide the bug.
What changes when we remove the hidden activation?
Remove a hidden activation while keeping the weights fixed, then explain the changed XOR representations.
Which shifts does pooling actually ignore?
Move a response within and across pooling windows to test which shifts preserve the output.
Can a model become more confident without getting more answers right?
Change softmax temperature on a frozen batch and compare accuracy, confidence, and cross-entropy.
Why can eight nonzero updates leave a value unchanged?
Compare float32 and float64 updates with an exact total to discover when nonzero increments disappear.