Abstract:Artificial neural networks (ANNs) are powerful tools capable of approximating any arbitrary mathematical function, but their interpretability remains limited, rendering them as black box models. To address this issue, numerous methods have been proposed to enhance the explainability and interpretability of ANNs. In this study, we introduce the application of information geometric framework to investigate phase transition-like behavior during the training of ANNs and relate these transitions to overfitting in certain models.
The evolution of ANNs during training is studied by looking at the probability distribution of its parameters. Information geometry utilizing the principles of differential geometry, offers a unique perspective on probability and statistics by considering probability density functions as points on a Riemannian manifold. We create this manifold using a metric based on Fisher information to define a distance and a velocity. By parameterizing this distance and velocity with training steps, we study how the ANN evolves as training progresses. Utilizing standard datasets like MNIST, FMNIST and CIFAR-10, we observe a transition in the motion on the manifold while training the ANN and this transition is identified with over-fitting in the ANN models considered. The information geometric transitions observed is shown to be mathematically similar to the phase transitions in physics. Preliminary results showing finite-size scaling behavior is also provided. This work contributes to the development of robust tools for improving the explainability and interpretability of ANNs, aiding in our understanding of the variability of the parameters these complex models exhibit during training.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the interpretability and explainability of parameter changes in artificial neural networks (ANNs) during the training process. Specifically, the paper studies the evolution of parameter distributions in ANNs during training by introducing an information - geometric framework, as well as the relationship between these evolutions and overfitting. The author hopes to use the concepts of distance and velocity in information geometry to reveal the phase - transition - like behaviors that may occur in ANNs during training by studying the changes in parameter probability distributions, and make an analogy with the phase - transition phenomena in physics. In addition, the paper also explores how to predict the overfitting of ANNs through the changes in information length and information velocity without directly referring to the test data set.
### Main research contents:
1. **Application of the information - geometric framework**: The paper uses the information - geometric framework to study the parameter changes in ANNs during training. Information geometry is a method of studying probability and statistics using differential geometry techniques. By regarding the probability density function as a point on a Riemannian manifold, it defines a distance measure based on Fisher information.
2. **Information length and information velocity**: The paper defines the information length \(L\) and the information velocity \(\Gamma\), which are used to quantify the total distance and the rate of change of the parameter distribution over time, respectively. These two quantities can be calculated by the following formulas:
\[
L(t):=\int_{0}^{t}\sqrt{\int dx\frac{1}{p(x,t_{1})}\left(\frac{\partial p(x,t_{1})}{\partial t_{1}}\right)^{2}}dt_{1}
\]
\[
\Gamma(t):=\sqrt{\int dx\frac{1}{p(x,t)}\left(\frac{\partial p(x,t)}{\partial t}\right)^{2}}
\]
3. **Experimental setup**: The paper uses three standard data sets, namely MNIST, Fashion - MNIST and CIFAR - 10, to study the effects of different optimizers (such as SGD, Adadelta, etc.) and hyperparameters (such as learning rate, dropout probability, noise regularization intensity, etc.) on information length and information velocity.
4. **Result analysis**: The paper finds that the changes in information length and information velocity can independently indicate the overfitting of ANNs. In particular, in shallow networks and deep networks (such as ResNet - 50), the minimum value of the second - order derivative of the information length \(\frac{d^{2}(\log L)}{d(\log t)^{2}}\) corresponds to the occurrence point of overfitting.
5. **Phase - transition analogy**: The paper further makes an analogy between the transition behaviors in information geometry and the phase - transition phenomena in physics, and finds that these transition behaviors exhibit finite - size - scaling behaviors similar to phase - transitions in finite - size systems.
### Conclusion:
By introducing the information - geometric framework, the paper provides a new method to study and explain the parameter changes in ANNs during training and their relationship with overfitting. This method not only helps to improve the interpretability and explainability of ANNs, but also can be used as an independent tool to predict overfitting, thus providing new ideas for improving the training process of ANNs.