[0:05] Plan Day
[0:48] Think about product distribution case for semi-adversarial setting
[0:44] Write up product distribution case
[0:19] Improve de-serialization speed for risk estimation code
[0:48] Set up 2 more datasets for risk estimation project
[0:34] Understand where estimates are off and why
Possible reasons I discovered:
-very narrow set of words get classified as anchors (causes distributional skew, often leads to under-estimate of risk)
-very easy and common examples end up not getting classified as anchors (leads to over-estimate of risk)
-removing the anchor from the prediction can be a large hit to accuracy if classification is based on only a few features (leads to over-estimate of risk)
[0:10] Think about how to improve the estimates
[0:15] Look over Uri / Frederik’s papers on representation learning for counterfactual estimation
[0:42] More reading about non-negative matrix factorization
[1:05] Meeting about code translation project
[1:00] Climbing


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s