Abstract
We present a global algorithm for training multilayer neural networks in this Letter. The algorithm is focused on controlling the local fields of neurons induced by the input of samples by random adaptations of the synaptic weights. Unlike the backpropagation algorithm, the networks may have discrete-state weights, and may apply either differentiable or nondifferentiable neural transfer functions. A two-layer network is trained as an example to separate a linearly inseparable set of samples into two categories, and its powerful generalization capacity is emphasized. The extension to more general cases is straightforward.
Introduction
Since the 1980s, the Backpropagation (BP) algorithm has been the dominant method for training multilayer neural networks. However, BP faces several fundamental "bottlenecks." First, it requires smooth, differentiable transfer functions to calculate gradients, making it incompatible with step functions. Second, it produces continuous weights, whereas biological systems and digital hardware often favor discrete states. Finally, BP is a local update rule, adjusting weights based on individual samples or mini-batches, which can lead to suboptimal global solutions. This paper introduces the Monte Carlo Adaptation (MCA) algorithm, a global training method designed to overcome these limitations.
The Monte Carlo Adaptation (MCA) Algorithm
The MCA algorithm replaces gradient descent with a stochastic search process. Instead of local updates, it employs a global evaluation strategy: for every weight change, the algorithm assesses performance across the entire training dataset.
The core innovation lies in the control of Local Fields of Neurons Induced by the Input of Samples (LFNIIS). The local field is the weighted sum of inputs to a neuron before the activation function is applied. The training loop proceeds by randomly flipping a discrete weight state and accepting the change only if it improves global classification accuracy and increases the average absolute value of the local fields in the hidden layer.
Robustness through Local Field Control
By maximizing the absolute value of local fields, the algorithm pushes neural activity away from the zero-threshold. This creates a "gap" in the distribution of internal signals, making the network robust against input noise.
As shown in Figure 1, the hidden layer fields under MCA (green line) form a distinct two-peak structure far from zero. In contrast, traditional gradient-based methods often result in fields clustered near the threshold, where small input perturbations can easily flip the neuron's state.
Performance and Generalization
The authors tested the MCA algorithm on synthetic datasets of $N$-dimensional binary vectors ($N=1000$). They focused on linearly inseparable sets where the number of samples $M$ exceeded the capacity of a single-layer perceptron.
The results demonstrate that MCA-trained networks possess superior generalization capacity. When tested on "unlearned" patterns—training samples with randomly flipped bits—the MCA network maintained a significantly higher rate of correct classification compared to BP-trained networks.
Beyond generalization, the MCA algorithm proved highly efficient. For a two-layer network, MCA reached training goals in approximately one hour on a standard PC, while BP required nearly six hours. Furthermore, because MCA works natively with discrete weights (e.g., binary $\pm 1$), it is ideally suited for implementation on digital hardware like FPGAs.
Conclusion
The Monte Carlo Adaptation algorithm offers a powerful global alternative to backpropagation. By focusing on the statistical distribution of local fields rather than local gradients, it achieves better generalization and hardware compatibility. While future work is needed to explore its scalability to deep architectures and massive datasets, MCA represents a significant step toward more robust and biologically plausible neural training methods.
Summary updated