核心概念
Biologically-plausible credit assignment algorithms offer solutions to the limitations of backpropagation, enabling more robust, energy-efficient, and hardware-compatible machine learning models.
摘要
This paper reviews several prominent biologically-inspired credit assignment algorithms that aim to address the limitations of backpropagation, a widely used but biologically implausible learning method in artificial neural networks.
The key issues with backpropagation that these algorithms seek to resolve include:
- Weight Transport (WT): The use of the same weights for forward and backward passes is biologically implausible.
- Forward Locking (FL) and Backward Locking (BL): The sequential dependencies in the forward and backward passes are at odds with the parallel, distributed nature of computation in biological neural systems.
- Forward-Backward Differentiation (FBD): The divergence in computation between the forward and backward passes is seen as implausible.
The reviewed algorithms include:
- Predictive Coding (PC): Minimizes prediction error by having neurons predict their inputs and propagating discrepancies upwards through the hierarchy.
- Contrastive Hebbian Learning (CHL) and Equilibrium Propagation (EP): Utilize an energy-based model that relaxes to solutions through iterative computation phases.
- Forward-Only Learning (FO): Avoids feedback pathways and relies solely on the forward inference process for credit assignment.
- Other emerging approaches like Direct Feedback Alignment (DFA), Target Propagation (TP), and Local Representation Alignment (LRA).
Each algorithm is described in terms of its underlying energy functional and learning dynamics. The paper also discusses the potential of these biologically-inspired algorithms for neuromorphic hardware implementations, as well as future research directions to improve their performance and scalability.