mirker

16 ms·

121.
by mirker 5y ago
There isn’t much theory on why it works in the nonconvex case. You can similarly make discrete heuristic algorithms for sorting, etc. with no guarantees. Anyway, as nonconvex models are usually stacked convex models, you can find works that
122.
by mirker 5y ago
Even the tokenization is wonky. Imagine if you had no concept of math characters and instead has a lookup table of common-ngrams (BPE encoding). For example, the binary addition function “a+b” may be tokenized as a unary “3+b” because “3+b”
123.
by mirker 5y ago
The greedy approach has an 1-1/e type of approximatation. You basically prove that the problem is in the class and then solve it greedily. It’s not much different than proving the problem is convex and then using gradient descent.
124.
by mirker 5y ago
There are discrete forms of “gradients” which generalize classical CS concepts. Submodular optimization, for example, covers many algorithms that search for optimal configurations of discrete sets, and it does so by arguments that are analo
125.
by mirker 5y ago
One factor for why is that bringing bad news may poorly reflect on the organization, and therefore the person’s career.
126.
by mirker 5y ago
That’s fine until you have to store the whole world’s cat videos on Youtube at minimal cost.
127.
by mirker 5y ago
In between those, VGG was parameterized by layers, though severely limited in scope (16-19 layers).
128.
by mirker 5y ago
I think this is more of a “reset jetlag” observation than one about nutrition. Food timings can affect circadian rhythms. For example, this article discusses fasting to reset jet lag https://www.health.harvard.edu/blog/
129.
by mirker 5y ago
One guess: if the reliability of Intel is considered a national security issue, they may get funding specifically to be more reliable to offset these costs.
130.
by mirker 5y ago
FOSS eliminates more uncertainty than price “supply” uncertainty. Since it’s free, you don’t have to know the “demand” value it provides to make a good purchase. Furthermore, you can always add features you need. All of this uncertainty is
131.
by mirker 5y ago
Yes, and the professors are also part of the government and financial success of a research-based school. Undergraduate students are perhaps only 50% of the school’s priorities.
132.
by mirker 5y ago
There are a lot of older (and otherwise susceptible) people on campus. Regardless, no they can’t do that because if something did happen, they’d get blasted. It’s the same with anything safety-related, really. Driving a car is super unsafe,
133.
by mirker 5y ago
You missed the most fun option: pair them up with a problem that leads to nowhere!
134.
by mirker 5y ago
Ah well that requires a lot of manual inspection. Usually there isn’t enough attention span to do that in practice.
135.
by mirker 5y ago
What if there are many alternatives? Each implementation would be a challenge to optimize fairly. You’d be forced to have some open competition to make it convincing.
136.
by mirker 5y ago
I agree with you, but at the same time, I think it’s easy to dismiss this whole line of thought as quacky. It’s very hard to argue with tables and plots. Economics has a lot of mathematical work for similar “hard science” reasons.
137.
by mirker 5y ago
One of the quagmires with systems research is that it makes “stuff” (performance or usability) easier for the user. Research is supposed to be “scientific”, so you need ample evidence of those claims. This is difficult to obtain for very we
138.
by mirker 5y ago
You basically have to pick between spending 5 years doing software “residency” and finishing the MD. Your drive in either will partially determine success. I agree with other commenters that you should find a way to do both, because realist
139.
by mirker 5y ago
The part where their salary wasn’t very good in the salary table; you have to play the market, which is currently favorable toward software engineering. I’m not sure what people are expecting. Eating rice and beans on $100k is still less th
140.
by mirker 5y ago
Yeah. Even modern CV methods are hacky insofar as picking the “right” way to apply linear algebra. Convolution layers are hacked up matrix multiplications that are “inspired” by human vision. Of course, the real reason for the hacks is that
141.
by mirker 5y ago
It’s this issue: https://github.com/ValveSoftware/steam-for-linux/issues/5778
142.
by mirker 5y ago
One big pain point is driver compatibility. Nvidia dropped 32-bit OpenGL, which is required for steam. CUDA packages in Ubuntu sometimes don’t pull in these dependencies, creating “dependency hell” for users.
143.
by mirker 5y ago
I don’t agree with the definition of “human research” you are using. For example, Cornell IRB defines it as 1) extracting data from humans OR 2) collecting private information. The email collects information from a human, therefore, by 1) i
144.
by mirker 5y ago
It is unethical because you are lying to people in an effort to manipulate their natural state. As a researcher, you have to balance the good in a research study versus the bad, and you can’t always predict how bad the outcome will be. Ther
145.
by mirker 5y ago
I agree you can make software rigorous like in ME. The part which is hard is that debugging or proving properties about a program is much more difficult than writing the program. These costs are currently hard to amortize over multiple proj
146.
by mirker 5y ago
Having read some Jax high-performance code, I do like Jax, but it does feel a bit too abstract and low level sometimes. Maybe there aren’t good coding conventions or performance trumped them? Definitely needs improvement on error messages,
147.
by mirker 5y ago
I’ve had palpitations before from having too much caffeine regularly (yes, while stressed).
148.
by mirker 5y ago
Are dependency chains complex functions or a few instructions? For example, I see people using ternary expressions “a > b : c ? d”. I’m trying to understand if this is an algorithmic problem or a microarchitecture problem. Are “c” and “d
149.
by mirker 5y ago
Well, given the “flaw” is a machine learning model, who was the developer and how do they relate to the training data?
150.
by mirker 5y ago
The precursor to cloud provisioning, I see.

More ›