Lesson 4 In-Class Discussion ✅

https://arxiv.org/abs/1607.06520

@rachel can you please update ?

Even sadder part is that the parent company Bytedance (字节跳动)recently became the world’s most valuable startup (valued at $75 billion)

5 Likes

sounds good

1 Like

I’d recommend following Rachel on twitter

2 Likes

Andrew Ng mentions this paper in his course - https://arxiv.org/abs/1607.06520

New stream: https://www.youtube.com/watch?v=GK1XhPM3K0g

5 Likes

I’m not sure if its even that. If historically more men applied for jobs at Amazon and as a result more men got hired, the model will make the unwanted correlation that maleness is a factor in the hiring decision. It is not that Amazon had bias, but that historical data is not representative of the current reality.

6 Likes

I don’t think is unintentionally. I recently read this article that shows this scenario over and over again.

1 Like

can someone give me the youtube link, all i have is the old recorded stream

This is a question about LRFinder. I hope it is not too late to ask it!

The LRFinder tries a sequence of learning rates and this lets us shows a plot of learning rate (LR) versus Loss. But the documentation for it says that it starts with the low learning rate and changes the learning rate “at each mini-batch”.

This seems to mean that, at every iteration, the LRFinder is changing both the learning rate it tries and the mini-batch of data being used to assess that learning rate.

Doesn’t that mean that it’s producing distorted comparisons of the learning rates, because each learning rate sees a different mini-batch and also benefits from the learning done on the earlier learning rates?

Are these distortions trivial? Or am I misunderstanding what it does?

8 Likes

The same one works too.

5 Likes

https://www.youtube.com/watch?v=F90C0A6UmVI

1 Like

Great example. Machine is not creating any bias on its own; it is simply learning from our historical biases. In some sense, it is a conservative, not a progressive. We call the result biased because we have an idea of what the right thing should be.

Hey, maybe we can use ML/DL to uncover the current biases?

1 Like

If the distortions are significant you will see it reflected as lots of noise in the learning rate finder plot.

  • livestream is working.
  • livestream is not working.

0 voters

is there a new link?

1 Like

Jeremy posted the new link (a few posts above)

Yes look above to Jeremy post

all the link does is play to the end when it stopped