Fig. 2: Multitask training on MNIST composites (Part 1/2).
From: Modeling attention and binding in the brain through bidirectional recurrent gating
Results for a single model trained on seven tasks simultaneously. The figure includes input and output signals, as well as the target signals (i.e., the desired outputs of the model). If the target signal is used during training, it is marked by “training” subscript (e.g., attention maps for the top-down search task), otherwise marked by “evaluation” and framed by dashed outlines (e.g., class label for top-down search). Here we present the results for: a object recognition, b perceptual grouping using spatial cue, c orienting via symbolic cue, d pop-out saliency, and e top-down visual search. Note: colors for the input sequence in task (a) have been inverted to improve visualization. Digit images are adapted and modified from the MNIST dataset ©LeCun, Cortes, and Burges79, available under a CC BY-SA 3.0 license.
