关于neural networks

问题如下：

A team of data scientists at a global asset management firm is working on a neural network (NN) model to assess how the choice of metrics used by the firm to select equity investments relates to stock performance. In particular, the team designs the NN model to find the combination of weights for these metrics that generates the highest risk-adjusted return in the firm’s portfolios. The team optimizes the model during the training process by varying multiple parameters and observes that, during multiple runs with identical parameters, the loss function converges to different, yet still stable, values. Which of the following is the most appropriate action for the team to take to address this issue?

选项：

A.Construct a confusion matrix to train the model.

B.Apply adaptive boosting when training the model.

C.Decrease the learning rate of the model.

D.Increase the learning rate of the model.

解释：

Explanation: C is correct. The size of the step in the gradient descent algorithm is known as the learning rate. If it is too large, it may oscillate from one side of the optimization valley to the other. Decreasing the learning rate would therefore prevent finding multiple local minima and prevent overshooting the global loss function minimum.

A is incorrect. The confusion matrix is a table that is constructed for model evaluation, specifically to present the possible outcomes and results of a predictive model wherein the output variable is binary categorical.

B is incorrect. Adaptive boosting, or AdaBoost, is an ensemble technique used to improve a model’s performance training it on the errors of previous iterations. AdaBoost operates differently from a neural network and is not directly concerned with finding local minima in a gradient descent.

D is incorrect. As explained in C, increasing the learning rate could result in the gradient descent algorithm oscillating from one side of the optimization valley to the other, which is exactly the problem that needs to be addressed.

Learning Objective: Understand how neural networks are constructed and how their weights are determined.

Reference: Global Association of Risk Professionals. Quantitative Analysis. New York, NY: Pearson, 2023, Chapter 15, Machine Learning and Prediction [QA-15].