Beyond gradients: Noise correlations control Hebbian plasticity to shape credit assignment

Two key problems that span biological and industrial neural network research are how networks can be trained to generalize well and to minimize destructive interference between tasks. Both hinge on credit assignment, the targeting of specific network weights for change. In artificial networks, credit assignment is typically governed by gradient descent. Biological learning is thus often analyzed as a means to approximate gradients. We take the complementary perspective that biological learning rules likely confer advantages when they aren't gradient approximations. Further, we hypothesized that noise correlations, often considered detrimental, could usefully shape this learning. Indeed, we show that noise and three-factor plasticity interact to compute directional derivatives of reward, which can improve generalization, robustness to interference, and multi-task learning. This interaction also provides a method for routing learning quasi-independently of activity and connectivity, and demonstrates how biologically inspired inductive biases can be fruitfully embedded in learning algorithms.