Abstract: The cortex learns to make associations between stimuli and spiking activity which supports behaviour. It does this by adjusting synaptic weights. The complexity of these transformations implies that synapses have to change without access to the full error information, a problem typically referred to as "credit-assignment". However, it remains unknown how the cortex solves this problem. We propose that a combination of plasticity rules—Hebbian, acetylcholine-dependent, noradrenaline-dependent, and inhibitory plasticity—effectively solves the credit assignment problem. Our work suggests that detailed cortical excitatory-inhibitory (E/I) balance reduces the dimensionality of the problem of associating inputs with outputs, thereby allowing imperfect "supervision" by neuromodulatory systems to guide learning effectively.
The Biological Credit Assignment Problem
In machine learning, the credit assignment problem—determining which specific weights are responsible for an error—is typically solved using backpropagation. However, biological synapses are "local"; they generally only have access to the activity of the two neurons they connect, not a global error signal from a distant output. This paper explores how a single neuron can learn complex associations using only local activity and broad, "imperfect" neuromodulatory signals like Acetylcholine (ACh) and Noradrenaline (NE).
A Multi-Rule Framework for Learning
The authors developed a mathematical model of a neuromodulated binary perceptron that integrates four distinct plasticity mechanisms. Rather than relying on a single complex rule, the model leverages the interaction between unsupervised, supervised, and homeostatic processes.
As shown in Figure 1, the model incorporates:
- Hebbian Plasticity: Standard unsupervised learning based on pre- and post-synaptic activity.
- Cholinergic (ACh) Gating: Excitatory plasticity gated by the target output, handling "expected uncertainty."
- Noradrenergic (NE) Gating: Non-specific excitatory plasticity acting as a "world model" update for "unexpected uncertainty."
- Inhibitory Plasticity: A rule that adjusts inhibitory weights to maintain a specific E/I balance.
The Geometry of Learning: The Tilted Balance Line
A key insight of this work is the role of inhibitory plasticity in shaping the learning landscape. By projecting high-dimensional synaptic changes into a 2D space defined by total excitatory (E) and inhibitory (I) currents, the researchers analyzed the dynamics of learning.
Traditional models often assume E/I balance is parallel to the spiking threshold. However, the analysis reveals that for robust learning, the balance line must be "tilted"—intersecting the threshold (Figure 2b). This tilt effectively reduces the dimensionality of the input space. Instead of navigating a complex high-dimensional error surface, the neuron only needs to be "nudged" to the correct side of the threshold along this 1D balance line. This allows neuromodulators to guide the neuron toward its theoretical storage capacity with very little information.
Key Results and Experimental Validation
The proposed multi-rule algorithm allows a neuron to learn a number of associations reaching 80-90% of its theoretical capacity (the Gardner limit). Remarkably, it does so even when the information content of the neuromodulatory supervision is 50 times lower than what the standard Delta-rule requires.
To test these theoretical predictions, the authors reanalyzed electrophysiological data from the mouse auditory cortex. The data confirmed two critical predictions:
- Tilted Balance: Inhibitory plasticity indeed drives synapses toward a "tilted" balance line with a slope less than one ($a < 1$), as predicted for robust learning.
- LTD Bias: ACh-dependent plasticity is biased toward long-term depression (LTD) for non-paired stimuli, a counterintuitive but necessary condition for stable learning in this framework.
Conclusion
By combining detailed E/I balance with broad neuromodulatory signals, the cortex can approximate complex learning rules using only local information. This suggests that the brain's architecture of balance is not just for stability, but is a fundamental component of its learning algorithm, allowing imperfect biological signals to achieve near-optimal performance.
Reference: Aljadeff, J., D’amour, J., Field, R. E., Froemke, R. C., & Clopath, C. (2019). Cortical credit assignment by Hebbian, neuromodulatory and inhibitory plasticity. arXiv:1911.00307.
Summary updated