Accuracy [% ] ElasticTransform 1 2 3 4 5 0 20

However, co-training on predicted neural responses improves the robustness of the models even more. We believe,thisisbecause theMTL-Monkeymodel usesthesameimages, i.e. tinyImageNet images amounting to100k images,for both tasks, namely neural prediction and image classification. In order to achieve this, we built a new model, starting with a random VGG, and fine-tuned the early layers (up toconv-3-1) to predict the actual neuronal data. Wenoticedthatthelowest resolution values are associated with the highest classification accuracies and vice versa. Stimulus presentation The stimuli were presented on a16:9 HDLCD monitor, with arefresh rate of 100 Hz, with a resolution of 1920x1080 pixels.