New Task! Visual Commonsense Reasoning


#1

A new task and large-scale dataset for cognition-level visual understanding:

http://visualcommonsense.com/


#2

Code and data coming soon!

Looking forward to taking a look at those :slight_smile:

I’m actually pretty happy to see Commonsense reasoning being brought back as topic for research.

I got interested in AI/ML when my university Logic professor introduced me to these papers from John McCarthy and Erik Mueller.
Following these papers I got to know more about the work of Erik and after sending an email and actually getting a response(!!!) he introduced me to the work of the late Push Singh at the MIT Media Lab on Commonsense Reasoning, like ConceptNet and EM-ONE.

Those days people were still discussing Statistical Methods vs Logic approaches, big data crunching wasn’t a thing and Neural Nets were still buried in the past, so now that we’ve actually had so much progress I’m really excited to see what’s gonna come out of it.

I only took a quick look at the first 2 pages but I’m definitely gonna be taking some time this week to go over this article. Thanks for sharing!


#3

Thank you for sharing those papers! they look pretty interesting :smiley:

Also, I think not only the Commonsense reasoning dataset is cool, but also the method used for creating it: Adversarial matching. With it you can create different kinds of multiple-choice datasets, and control how difficult they are for humans vs. how difficult they are for machines, among other things.

The same author created the SWAG dataset earlier this year by using a similar adversarial technique: Adversarial Filtering [paper].

I find it awesome that adversarial techniques can not only be used for GANs or adversarial attacks, but also for creating datasets!


#4

Thank you so much for sharing this. Quite excited to read the paper!


#5

Just noticed the data is out for download!