CVPR Daily - Wednesday

In the end, the team found that models exploit linguistic biases, so a lot of the gain of accuracy is not actually improving visual understanding, it’s just toying with the dataset better. In terms of the metrics for specifically measuring compositional reasoning, they found some that models did okay at, but some that they did very poorly at. Madeleine hopes to use this dataset to think about ways of more explicitly reasoning over symbolic representations of videos or questions and using that to take the annotations that they have in their dataset to explore further. Moving away from the science for a moment, we want to talk to Madeleine about another one of her passions: competitive gymnastics! “ That makes me happy to think about because it’s been a few years, ” she smiles. “ It’s always good to work towards a goal and to challenge yourself. When you break down a skill like a tumbling pass it has a bunch of individual parts that you have to learn and then you put it all together and you’re able to train your body to do something that you couldn’t do before. I liked it because it wasn’t just being able to do something faster or longer, it was taking something that was out of reach and learning it. I also like performing, so it was nice to go to competitions, win trophies, have a good time, and get to show off everything that you’ve done! ” We have to ask: How does everything link together? Someone who codes, someone who tries to understand human minds and processes, and someone who likes to make incredible jumps on a beam! What do they all have in common? “ An overachiever! ” she laughs. “ Don’t put that in! But seriously, breaking down 12 DAILY CVPR Wednesday Presentation