1,320 karma · joined August 2, 2013
Yes, I agree that attempting to schedule things in this manner is better than not, should you find yourself able to do so. But, I don't agree with the parent comment that it is "the ultimate and most powerful interviewing technique". Scheduling by preference is moot if (1) you don't have a resume/background that lands interviews in the first place and (2) you don't have the chops to be successful in technical rounds.
I'll piggy-back off this to remind people that when someone shares something, they are not necessarily asking for your advice. I know HN is a crowd of problem solvers, but it's OK (preferable!) to simply respond with empathy. If it's not clear, consider asking the person if they'd like your advice before immediately jumping in to try to fix it for them.
Same. It's a shame because it's often added to some blends that I would otherwise enjoy and would not expect licorice root to be added.
I took this approach when I was in grad school because I find it much more enjoyable to read and take notes directly on a physical copy.
The problem with this is that you end up with giant stacks of papers that take up a good amount of physical space. Every physical copy becomes a subsequent judgement of whether you need to keep it or recycle it. It's easy to accumulate a large stack of papers: (1) that you _intend_ to read, or (2) mediocre papers that create a sunk cost situation because you've invested time in printing out and taking notes on it.
On the face of it, a tablet reader seems like it would solve this problem, but I've never had the desire to invest $300-500 in one.
However, there's also times when I don't have the luxury of using an IDE or when I just need to make quick edits to a single file. In these circumstances, I can just fire up vim and quickly do what I need to do because vim is ubiquitous and editing with it is efficient.
1. The geometric mean is the minimizer of:
L(x) = Σ [ log(x_i) - log(x) ]²
2. The so-called Jaccard median is the minimizer of:
L(x) = Σ [ 1-J(x_i, x) ]
where J is the Jaccard overlap score (aka, IoU = Intersection over Union). The second is interesting because the Jaccard score is often used to describe how close two segmentations are in image processing. The Jaccard median can be used to consolidate multiple segmentations of the same object in such a way that is best under the Jaccard distance.
> What kind of CNN architectures for classification do you know?
Is rated as an expert level question, but reads more like "what are some of your favorite pokemon?".
We have laws to protect consumers against the latter. Why should advertising in politics be held to different standards?
ish: https://www.dictionary.com/browse/ish
So, having an aesthetic similar to rhythm and blues.
Well, that and a decision tree isn't any more interpretable if its nodes are all filled with `Pixel[i,j] > Threshold` -- or is the idea that you would somehow extract logical predicates by tracing paths through the tree or glean interpretation from those predicates?
Just to be clear, additive loss doesn't imply stochastic gradient estimate. Rather, because the loss function is additive, then stochastic gradient estimates of the loss are now possible. But, this of course does not mean one has to use stochastic gradient estimates.
It's just that it's easier to update and monitor progress this way, rather than computing the gradient term for every single example in the training set and then taking a descent step. The surprising thing is that stochastic gradient descent convergences quickly in practice relative to proper gradient descent. All of the justification and whatnot for SGD for ML is largely post-hoc because it works so unreasonably well and is so intuitive to anyone having taken calculus.
The other aspect (with respect to the context of optimization in machine learning) is that this optimization is performed over a loss over a training dataset for which you really don't even want convergence to an exact minima over the training loss. What you really care about is the expected generalization loss. Convergence to the exact minima over training loss doesn't necessarily guarantee the best generalization loss. I mention this because it contributes to the general aloofness towards optimization convergence rates in ML.
> I believe many have investigated quasi-Newton methods based on estimate gradients but I haven’t investigated that thoroughly.
Until semi-recently, quasi-newton was not explored in the stochastic setting because of the question of how to extend the Wolfe conditions to this arena. There's been a bit of work on this [1], but I don't think it's caught on outside of the optimization community (not that it necessarily should considering the points above).