trott

11 ms·

121.
by trott 6y ago
Genetic testing used to be very expensive: $3B to sequence the human genome. So Ancestry/23/etc. tried to guess certain properties based on very sparse features of your genome, which is hard. But today, whole genome sequencing cos
122.
by trott 6y ago
> But you still lose something, e.g. if you use half precision on V100 you get virtually double speed, if you do on a 1080 / 2080 you get... nothing because it's not supported. That's not true. FP16 is supported and can be
123.
by trott 6y ago
EA developers weren't happy with STL, so they created their own version, EASTL. One of their complaints had to do with custom allocators: STL allocators combine allocation and initialization. Reading their critique of STL design decisi
124.
by trott 6y ago
> https://threadreaderapp.com/thread/1325592112428163072.html Do you know why the scatter plots for SD and CA are showing a few fractional counts? Shouldn't they all be integers?
125.
by trott 6y ago
If you can use DDD (which is an interface to GDB ), it shows the disassembly.
126.
by trott 6y ago
> Why is it surprising that a CFG can approximate a subset of English grammar? "Colorless green ideas sleep furiously" is a famous example of a sentence that is grammatical, but meaningless. The goal of SHRDLU was far more ambi
127.
by trott 6y ago
Looking at the diagram, this appears to be just an LSTM slapped on top of a CNN. If so, I'm failing to see any novelty in this approach. RNNs on top of CNNs have been used before, including for deepfake detection. See for example: htt
128.
by trott 6y ago
> Vitamin D has been endlessly promoted as a treatment for many diseases due to evidence from observational studies, but hundreds of millions of dollars in randomized controlled trials have never demonstrated it to be effective for any d
129.
by trott 6y ago
But there is no need to use a particularly small NN before using a slightly larger one still in the "ascent" region. (Overfitting is relative)
130.
by trott 6y ago
> Actually, there's compelling evidence that overfitting is a necessary step for achieving state-of-the-art performance with DNNs! I think the term you are looking for is "overparametrization". Overfitting is, by definitio
131.
by trott 6y ago
Thanks for sharing your experiences! I did well in Kaggle's highest-prized competition[1], so I wondered if I should explore this kind of consulting myself. The competition I was in was effectively limited to US residents though, but a
132.
by trott 6y ago
It seems that the authors misunderstood the Boston study they are citing: > In a Boston homeless shelter, 100% of 147 COVID-19 positive subjects were asymptomatic [3]. That study actually excluded N=15 symptomatic people: > The origin
133.
Human body-on-chip platform could speed up drug testing(europeanscientist.com)
84 points by trott 6y ago|38 comments
134.
by trott 6y ago
> From the article it should be clear that the observers were there to learn from their Chinese colleagues, not to conduct independent research on human-to-human transmission. The Taiwanese experts saw clusters of disease, such as a wife
135.
by trott 6y ago
> Taiwan didn't even have a confirmed case until January 21. They couldn't possibly have reported anything but unconfirmed rumors from China. Actually, Taiwan sent its own experts to Wuhan: https://www.cdc.gov.tw&#x2
136.
by trott 7y ago
> If we had neurons that were 1,000x faster, could we use fewer of them for the same task? Absolutely. In fact, we do: in batched processing (pervasive in deep learning today), CNNs and other weight sharing schemes (e.g. transformers) th
137.
by trott 7y ago
I took part in TSA/Kaggle's competition a year ago. https://www.kaggle.com/c/passenger-screening-algorithm-chall... Let me first say that how well these things work is not a property of the scanners alone, bu
138.
by trott 8y ago
Thanks for the kind words!
139.
by trott 8y ago
My point was that, the way the language is actually used by its speakers, "blue" is more general than either of its two Russian translations.
140.
by trott 8y ago
> Are we in a situation now where, if I have a bad anti-botic resistant infection I can just order these molecules on the off chance that they will help me? Very far from it. I wrote a docking program myself (the most popular one, curren
141.
by trott 8y ago
The sky is blue. The flag is red, white and blue. To a Russian speaker, calling these colors by the same name seems absurd.
142.
by trott 8y ago
Right. I'm just saying that this myopia is not a fundamental property of the recurrence any more than of convolution. Clockwork RNNs subsample, BTW, so they are more analogous stride=2 in CNNs than to dilation.
143.
by trott 8y ago
In an RNN, you could connect each hidden state at time step t, h(t) to h(t-N), instead of, or in addition to, h(t-1), making it analogous to dilated convolutions, but with hidden-to-hidden connections at the same layer. So I don't thin
144.
by trott 8y ago
Should developers also teach juggling to professional jugglers?
145.
by trott 8y ago
JFT has 17K classes. I'm assuming that they used specialized experts to tell them apart (dog breeds, plant and animal species, etc.)
146.
by trott 8y ago
> For commercially useful applications, collecting the data, labeling it, etc, costs orders of magnitude more than a team of PhDs building the models. I don't think it's typical. For example, JFT has 350e6 images, and it probab
147.
by trott 8y ago
I think this kind of resource is in a chicken-and-egg situation: for people to bother to contribute to it, it needs to be seen by others as an authoritative reference, and vice versa. I agree that it's badly needed. There were other pr
148.
by trott 8y ago
Commendable effort, but badly outdated already. For example, the best result on CIFAR-10 for density models that I know of is 2.85 bits per dim (PixelSNAIL) rather than 3.35 (Glow), and this is a year old. PixelCNN++ is well known, and got
149.
by trott 9y ago
Thanks. The spherical wave assumption used in the calculation is not compatible with the focusing effect, and could be wildly wrong in this case.
150.
by trott 9y ago
> This predicts 138 dB of sound. I doubt that one can calculate the sound amplitude accurately without doing experiments or complex simulations here. It's possible that the toilet bowl or lid focused the sound like a satellite dish,

More ›