Skip to main content

Posts

When you use noise layer instead of dropout layer in neural network

Google has patent for dropout layer. So if you don't want to violate it, you might as well just use random noise layer. Results pretty similar for MNIST.
convolution,1,28,28,200,8,8,0.5,-0.001 max,200,21,21,7,7 noise,1800,0.5 matrix,1800,130,0.5,-0.001 sigmoid,130 matrix,130,10,0.5,-0.001 softmax,10
Tested with WideOpenThoughts

When all you have to do is bring smiling face and a bottle of wine

When you downgrade GPU

WideOpenThoughts works even on GTX 1060 for laptops

Splashdown

Return of NASA's astronauts

On 48 hours testing

Bill Gates says most coronavirus tests are a 'complete waste' because the results come back too slow
From article:
"The simplest thing, which has to do with such insanity, is you should not reimburse somebody for getting a test that it takes more than 48 hours to get the result back," Gates said on Tuesday. "That test is a complete waste. And to all these numbers about how much we test, the majority is just complete waste," he added, calling it "insane" to have to pay for test results that could take more than three days and up to a full week.


My opinion:
I can't say I completely agree. First, if you test yourself, be responsible and until you have results do not meet with anyone. It's all about self responsibility. Also, even if your test comes negative, you want to meet with as few people as possible during these times.
And most importantly, results of these tests might be useful for "aftermath" of pandemic, when trying to better und…

When you realize you need only 64 random convolutions

when searching for 8 by 8 convolutions, as with 64 free parameters for every one of them you can create output of any convolution by linear combination of outputs from 64 random convolutions (which are not trained). 
Config:
convolution,1,28,28,64,8,8,0.5,0 convolution,1,64,441,200,64,1,0.5,-0.001 max,200,21,21,7,7 matrix,1800,130,0.5,-0.001 sigmoid,130 dropout,130,0.5 matrix,130,10,0.5,-0.001 softmax,10
First convolution is not trained, and you need only 64 of them (8x8 parameters = 64). Then with next convolution you find convolutions which you need by combining outputs of previous convolutions.
Results (above 99%):