Took this class at Berkeley last semester. Hard but very good, as long as you have the right mathematical background. Think graduate level math/stats and not "I took the ML course from Coursera".
If you don't already understand them, probably a good idea to skip the detailed derivations and look for the big picture. The course is really quite hard to follow closely if it's your first exposure to the material.
Skimming the slide set #1 (Systems) is highly useful, even for people that don't do machine learning.
For example, he covers the frequency of hardware failure, and also gives latencies for different operations (L1 cache read, disk read, etc.)
Slide 25 lists many different types of data on the web, categorized. This jumped out at me because, reading the list in one big picture got the gears in my head turning about potential data sources, and what could be done with them.
This is an awesome series of lectures. They were my regular evening listening for a time. Note some of the earlier lectures don't have sound. That makes them a bit hard to follow. :-) Also, the later lectures were missing last time I looked.
The ML/ML conflict actually forces me to backtrack reading some sentences simply because I always assume the person is talking about the language.