Computer Vision News - January 2019

16 Challenge: ABCD Computer Vision News Challenge it deep learning? Susan can talk about the whole scientific rationale looking at the relationship between brain structure and neurocognitive performance. Susan: I’m very excited to do this challenge so we can find some interesting ways to see how neurocognition relates to brain structure in kids. One thing that I think is very important is that we’re very focused at ABCD and looking at how the brain developed over time and the factors that influenced how that brain development might be different in some kids than in others. A really important consideration is: does it make any difference? The brain is used for many things, but one is for thinking and memory tasks. If we pull in the neurocognitive performance data collected in ABCD, we can help to snap on the meaning of brain developmental differences. Do you have any tips that you want to give to the participating teams? Susan: I guess a little bit of literature review would be a really good idea! Think through what we know about brain anatomy and cognition. I think that would be a helpful hint. Are there any mistakes to avoid? Susan: I think if there’s a finding that’s very different from what’s seen in the literature, you’d want to be more cautious. Wes: We’re trying to have people avoid making trivial predictions based on things that are not really relevant. One of the things that we residualize on is brain volume or head volume. We don’t want to key in on how big the head is or how big the brain is. Kilian can explain that better. We’re also residualizing socio-demographic factors because we don’t want people to focus on things that might be correlated with the brain structure, but not intrinsic to the performance of the brain. We’re also residualizing on data collection site because it is a multisite study with 21 sites. We don’t want people to key in again on site differences that would be irrelevant for making external inferences about what the relationships are that people are finding. We’re trying to make sure, as much as we can, that people don’t come up with algorithms that are senseless in the sense that they don’t generalize or say anything mechanistically about how the brain is related to neurocognition. Having said that, I agree with Susan completely that a good tip might be to review the literature and see what people know about the relationship between brain and neurocognition. How could it be built into a machine algorithm to improve its performance for its interpretability Susan Tapert is Professor of Psychiatry at UC San Diego