5 ms·

Not for a useful interview. If you tell me you founded and sold a machine learning company but can't explain the pros and cons of using variance versus standard

Not for a useful interview. If you tell me you founded and sold a machine learning company but can't explain the pros and cons of using variance versus standard deviation then you didn't sell a machine learning company.

thaumasiotes 10y ago
Hm. I have not founded or sold a machine learning company, or even studied the field. But either of those quantities fully determines the other, so I'd really hope that, if they're part of the model, they'd be exactly equivalent to a learning system. Do you mean for measuring goodness of fit between a prediction and the known truth in training data? If you want to assess the presence of knowledge, give a test rather than an interview. The interview is just a way to make it easier for the con artist to snow you.
solipsism 10y ago
One (stddev) is in the same units as the quantity being measured, while the other isn't. I know nothing about machine learning and suck at math and statistics. Did I pass?
- caub 10y ago
  if I was recruiter, math knowledge would make the difference, it helps even unconsciously for problem solving, logic reasoning, algorithms
  - solipsism 10y ago
    math... helps even unconsciously for problem solving, logic reasoning, algorithms What makes you think this? Anything aside from intuition?
    - caub 10y ago
      at least one point that helps math minds is that a lot of math notions are present in programming languages
- 3pt14159 10y ago
  Yeah, I don't know why I was downvoted, it's a completely legitimate question to ask a data scientist. But anyway, you'd get enough for a follow up question, but not really enough to pass. The key is that the sum of uncorrelated distributions has a variance that is the sum of the variance of those distributions. This isn't true for standard deviation. This is useful so we use variance when we think we're dealing with random variables. Standard deviation isn't as useful for intermediary processes. Another follow up question - actually the first question in my head, but by a mental slip-up I typed out the wrong thing: Why do we use square error when absolute error would do? The answer to this is actually quite illuminating because there are multiple valid but contradictory positions on it, much like the Bayes vs Frequentist debate.
BigJeffeRonaldo 10y ago
I'd be surprised if there weren't somewhere out there who didn't understand additivity of variance (and linear combination with the covariance terms, or that it is a more natural parameter for the normal distrovution, or some other reason) but had founded and sold a machine learning company. Further, this seems a bit like asking someone with several years of development experience if he knows what a compiler is.

thaumasiotes 10y ago

Hm. I have not founded or sold a machine learning company, or even studied the field. But either of those quantities fully determines the other, so I'd really hope that, if they're part of the model, they'd be exactly equivalent to a learning system. Do you mean for measuring goodness of fit between a prediction and the known truth in training data? If you want to assess the presence of knowledge, give a test rather than an interview. The interview is just a way to make it easier for the con artist to snow you.

solipsism 10y ago

One (stddev) is in the same units as the quantity being measured, while the other isn't. I know nothing about machine learning and suck at math and statistics. Did I pass?

caub 10y ago
if I was recruiter, math knowledge would make the difference, it helps even unconsciously for problem solving, logic reasoning, algorithms
- solipsism 10y ago
  math... helps even unconsciously for problem solving, logic reasoning, algorithms What makes you think this? Anything aside from intuition?
  - caub 10y ago
    at least one point that helps math minds is that a lot of math notions are present in programming languages
3pt14159 10y ago
Yeah, I don't know why I was downvoted, it's a completely legitimate question to ask a data scientist. But anyway, you'd get enough for a follow up question, but not really enough to pass. The key is that the sum of uncorrelated distributions has a variance that is the sum of the variance of those distributions. This isn't true for standard deviation. This is useful so we use variance when we think we're dealing with random variables. Standard deviation isn't as useful for intermediary processes. Another follow up question - actually the first question in my head, but by a mental slip-up I typed out the wrong thing: Why do we use square error when absolute error would do? The answer to this is actually quite illuminating because there are multiple valid but contradictory positions on it, much like the Bayes vs Frequentist debate.

BigJeffeRonaldo 10y ago

I'd be surprised if there weren't somewhere out there who didn't understand additivity of variance (and linear combination with the covariance terms, or that it is a more natural parameter for the normal distrovution, or some other reason) but had founded and sold a machine learning company. Further, this seems a bit like asking someone with several years of development experience if he knows what a compiler is.