🎯 So what can we do with this? The cool thing is that we can use these models to predict changes that were not part of training. For example, how does category training affect IT information about an object's size? We found that IT-like models effectively predicted changes for many variables! [5/7]

🦾We only tested models that performed as well as the monkeys after training. Notably, around 28% of all such trained models using gradient descent showed differences indistinguishable from those in IT across all tested metrics. Most models (~88%) matched the changes for at least one metric. [4/7]

🤖 To understand HOW this happens, we used pre-trained ANNs, mapped them to task-naïve primate IT & mimicked the monkey's training using various optimization approaches. This enabled us to simulate various scenarios of how IT might change in service of the behavioral performance gains [3/7]

📊 Our neural data analyses showed that responses in IT cortex were more categorical with category training, although only to a moderate extent across various metrics. Notably, these neural differences were dwarfed by those in behavior. [2/7]