Beyond prediction error: 25 years of modeling the associations formed in the insect mushroom body

Abstract

The insect mushroom body has gained increasing attention as a system in which the computational basis of neural learning circuits can be unraveled. We now understand in detail the key locations in this circuit where synaptic associations are formed between sensory patterns and values leading to actions. However, the actual learning rule (or rules) implemented by neural activity and leading to synaptic change is still an open question. Here, I survey the diversity of answers that have been offered in computational models of this system over the past decades, including the recurring assumption-in line with top-down theories of associative learning-that the core function is to reduce prediction error. However, I will argue, a more bottom-up approach may ultimately reveal a richer algorithmic capacity in this still enigmatic brain neuropil.

Figure 1.

Key features of the mushroom body (MB) and proposed learning rules. (Top) Sensory input is carried by projection neurons (PNs) which make sparse connections to MB Kenyon cells (KCs). KC axons bifurcate and connect to a small number of MB output neurons (MBONs). Reward and punishment signals are carried by dopaminergic neurons (DANs), which influence the plastic KC–MBON connections. MBONs also connect (selectively) to each other and in feedback loops (direct and indirect) to DANs; these connections can be excitatory or inhibitory. Additional features not illustrated include neurons conveying global inhibitory feedback from the lobes to the calyx, connections between KCs, and direct connections from KCs to DANs and from DANs to MBONs. (Bottom) Four broad classes of learning mechanisms have been proposed in the models discussed in the text (see also Table 1): Hebbian, in which simultaneous KC and MBON activation strengthens the synapse between them (this might assume that the MBON's initial activation is caused by direct input from a DAN); Kandelian, in which DAN activation releases a neuromodulator that strengthens the active KC synapses onto MBONs; three-factor, in which Hebbian learning is gated by DAN activation; and prediction error, in which Kandelian learning is modulated by negative feedback from MBON to DAN (note that if coincidence of KC–DAN activity is assumed to depress the synapse, then “negative feedback” would be implemented by an excitatory connection from MBON to DAN).