Why does this neutralize() give different results?

Course Q&A Deep Learning Specialization Sequence Models
1

In neutralize() implementation, the denominator of e_biascomponent is the square of the norm of g:

np.square(np.sqrt(np.sum(np.square(g))))

This is based on the original formula. However, this gets different results than the testing answer. The expression should be equal:

np.sum(np.square(g)))

This simplified version passed the test. Why does the two generate different results? For example:

np.square(np.sqrt(8)) = 8

Which basically np.square() cancels out np.sqrt(). But why do the two above get different results?

2

There’s room for small variations due to floating point operations by the CPU. That said, you can use np.isclose(norm_square(x,2), YOUR_CODE) and see that the simplified representation of np.sum(np.square(x)) is valid.

3

How much different are the results? Can you provide some values?

4

I just tried to reproduce the result by re-running the notebook, but it doesn’t output the prints after it goes to the ‘Debiasing word vectors’ section, so I can’t compare the diff of the two values now.