BIZNIQUE / LINEAR ALGEBRA / 15
Account for preprocessing cost
Account for preprocessing cost
Estimate the storage and passes required before adding another transformed dataset.
Embedded Files
THE IDEA
Read the example.
Read the example.
Computing feature statistics and applying a simple scaling rule each require work proportional to the number of visited values. Keeping both original and transformed dense arrays can approximately double value storage.
One dense pass touches n × d values
Use it well
Use it well
Centring a sparse matrix usually turns implicit zeros into nonzero values. Check sparse-compatible operations and peak memory, rather than assuming that preprocessing is free.
TRY IT YOURSELF
A quick check.
A quick check.
What happens to a stored zero after subtracting a nonzero feature mean?
Check your reasoning
Check your reasoning
It becomes −μ. Applying this to implicit zeros can destroy sparsity.
Page updated
Google Sites
Report abuse