Learning cortical hierarchies with temporal Hebbian updates.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Mammalian intelligence relies on hierarchical sensory processing, from low-level features to high-level object recognition.
• Artificial neural networks (ANNs) exhibit similar hierarchical structures but often use biologically implausible training algorithms like backpropagation.
• Biologically plausible alternatives exist, but a clear mechanism for neuronal comparison of signals for error calculation remains elusive.

Purpose of the Study:

• To propose a biologically plausible mechanism for neuronal error calculation and weight updates in deep learning.
• To demonstrate how local error signals can be computed by comparing neuronal compartmental activities.
• To show that this mechanism supports supervised hierarchical learning.

Main Methods:

• Introduced a novel learning rule combining apical feedback's effect on postsynaptic firing rate with differential Hebbian updates.
• Formulated and proved the equivalence of weight updates to minimizing inference latency and top-down feedback.
• Validated the differential Hebbian updates in established feedback-based deep learning frameworks like Predictive Coding and Equilibrium Propagation.

Main Results:

• Demonstrated that the proposed differential Hebbian update rule minimizes inference latency and top-down feedback requirements.
• Showed that this learning mechanism is effective within other biologically plausible deep learning frameworks.
• Established a link between temporal Hebbian learning and supervised hierarchical learning in ANNs.

Conclusions:

• This work removes a key requirement for biologically plausible deep learning models.
• Proposes a novel learning mechanism that bridges temporal Hebbian learning rules with supervised hierarchical learning.
• Offers a potential explanation for how biological neural networks achieve complex learning and intelligence.