Fig. 6. The multi-tasking model produces abstract representations from image inputs.

a Examples from the 2D shape dataset⁴³ (top) and chair image dataset⁴² (bottom). The 2D shapes dataset is from: Matthey, L., Higgins, I., Hassabis, D. & Lerchner, A. dsprites: Disentanglement testing sprites dataset. https://github.com/deepmind/dsprites-dataset/(2017). b Schematic of modified model. The multi-tasking model now begins with a networked pre-trained on the ImageNet challenge, followed by a few additional layers of processing before learning binary tasks as before (see “Pre-processing using a pre-trained network” in Methods). c The classifier (left) and regression (right) generalization performance when applied to the shape image pixels (top) or ImageNet representations (bottom). d The classifier (left) and regression (right) generalization performance of the multi-tasking model on the shape images. e The same as c but for the chair images. f The same as d but for the chair images.