NeuroAI researchers formalize the notion of contravariance by demonstrating that for any two minimal Deep Neural Network (DNN) solutions to a sufficiently hard task, weak alignment based on affine mappings guarantees strong alignment of privileged axes. This alignment zippers up the network hierarchy, causing privileged axes to emerge from end-to-end task optimization.
- Weak alignment of network representations guarantees strong alignment of privileged axes.
- Alignment propagates up the network hierarchy during optimization.
- Convergent evolution between artificial and real brain networks is likely inevitable with sufficiently strong tasks.
These results suggest that when tasks are sufficiently hard, the choice of metric for inter-network comparison becomes less sensitive.