Persistency of Excitation for Robustness of Neural Networks
This work addresses robustness and training stability in neural networks, particularly for classification, by identifying a fundamental limitation in multi-layer architectures, which is incremental but provides insights into adversarial vulnerabilities.
The paper tackles the problem of ensuring persistent excitation of neural network weights during training to guarantee correct parameter estimation, showing that intermediate layers become low-dimensional and fail to remain persistently excited in multi-layer networks for classification tasks. It proposes a regularization-based algorithm to address this, linking the issue to challenges like adversarial examples and limitations of data augmentation.
When an online learning algorithm is used to estimate the unknown parameters of a model, the signals interacting with the parameter estimates should not decay too quickly for the optimal values to be discovered correctly. This requirement is referred to as persistency of excitation, and it arises in various contexts, such as optimization with stochastic gradient methods, exploration for multi-armed bandits, and adaptive control of dynamical systems. While training a neural network, the iterative optimization algorithm involved also creates an online learning problem, and consequently, correct estimation of the optimal parameters requires persistent excitation of the network weights. In this work, we analyze the dynamics of the gradient descent algorithm while training a two-layer neural network with two different loss functions, the squared-error loss and the cross-entropy loss; and we obtain conditions to guarantee persistent excitation of the network weights. We then show that these conditions are difficult to satisfy when a multi-layer network is trained for a classification task, for the signals in the intermediate layers of the network become low-dimensional during training and fail to remain persistently exciting. To provide a remedy, we delve into the classical regularization terms used for linear models, reinterpret them as a means to ensure persistent excitation of the model parameters, and propose an algorithm for neural networks by building an analogy. The results in this work shed some light on why adversarial examples have become a challenging problem for neural networks, why merely augmenting training data sets will not be an effective approach to address them, and why there may not exist a data-independent regularization term for neural networks, which involve only the model parameters but not the training data.