Mainly useful for real-time applications, with the key benefit being that your data storage capacity can be much smaller than the amount of data being fed through the machine.
@whuber I believe I've hit the reputation cap every day so far this week, so if all future days are like this one, I should have the gold badge before too long.
Some background: I'm an undergrad student doing research in psychology and learning / decision making. One way to view certain psych experiments is that subjects are performing an online estimation algorithm: take a sample, condense it to memory, output a guess, repeat.
Most questions I see take the form of "what's an online algorithm to compute ___?" whereas I guess I'm more interested in the theoretical limits... when can/can't sufficiently-efficient online algorithms exist, and what's the theoretical tradeoff between memory usage and accuracy?
@MatthewDrury Thank you! Most of the stuff that I wrote is about my 1 year sojourn that was trying to get a generative language model to do well. Do you like being a data scientist in Seattle? I'm wondering if our nation's capital is the best place for a data scientist to have a career...
@Sycorax: It would be interesting to hear your definition of "better" (and "best")! Having collaborated with people from Seattle and D.C. I think one would gain -and lose- different things from working in different environments (I was based in California at the time).
@usεr11852 I just meant that data science in DC tends to be focused on the local industry, i.e. government. I've worked on gov't projects before; not sure that I'd like to do that as my next job.