Collaboration Between Economics And Deep Learning! Calculating The Contribution Of Each Neuron In Various Metrics!
3 main points
✔️ Proposed Neuron Shapley to measure the contribution of each neuron in a given evaluation axis
✔️ Using Neuron Shapley, we found that the accuracy of deep learning models depends on a very small set of neurons (filters)
✔️ It is also possible to identify the filters that contribute to the weaknesses and biases of the deep learning model, and to modify the deep learning model by removing the identified filters.
Neuron Shapley: Discovering the Responsible Neurons
written by Amirata Ghorbani, James Zou
(Submitted on 23 Feb 2020 (v1), last revised 13 Nov 2020 (this version, v3))
Comments: Accepted to arXiv.
Subjects: Machine Learning (stat.ML); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Neural and Evolutionary Computing (cs.NE)
official
first of all
As we all know, deep learning models are starting to be used in all sorts of fields because they are much more accurate and flexible than traditional methods. On the other hand, there are several problems that need to be overcome before deep learning techniques can become commonplace in the world. One of them is that the inside of a deep learning model is a black box. This means that when a deep learning model makes a prediction that is wrong or behaves strangely, it is difficult for humans to understand what part of the model is contributing to the result.
One of the things that we take for granted in this world is the automobile. If your car is behaving strangely, you need to find the root cause of what is wrong with your car and fix it. Whether it's something wrong with the engine, the tires, or the driver himself, these can be identified. In this way, deep learning models should also be able to check what is causing the problem and remove it.
The building blocks of a deep learning model are neurons (or filters, which are sets of neurons). Prediction results are produced by the interaction of each neuron. In other words, if we can know the contribution of each neuron to the prediction result by considering the interaction of each neuron, we can interpret why the deep learning model made the prediction, and we can even modify the model according to the contribution.
The paper presented here proposes Neuron Shapley, which calculates the contribution of each neuron as described above. Neuron Shapley is an application of Shapley Value, a metric used in cooperative games, a branch of economics, to deep learning models. In the following, we introduce some assumptions in economics and then describe the proposed method.
To read more,
Please register with AI-SCHOLAR.
ORCategories related to this article
