Scholarlink

In the AIPR 2017 Workshop we presented a paper that looked at “The Anatomy of a Neural Network”. The main purpose was to show how to take a simple LeCun-based neural network, with its multiple internal operations, and extract layer-to-layer input/output pairs to form a series of bidirectional association memory matrices (AMMs) from CNN Layers. From these transformations we showed how to achieve a practical solution that met the expectations of the Universal Approximation Theorem; how we could to collapse a six-layer system into a single layer. For the AIPR 2025 Workshop we will examine how a change in the image, be it adversarial or just a naïve-rotation of the image, can be physically tracked across each of the CNN/AMMs, when each is examined in vector form, i.e., we will visually exploit the error from the first layer on to the last layer. There are many approaches to measuring layer vulnerabilities, such as autoencoders, t-SNEs, histograms, and residuals. In contrast to these methods, we do not offer a measure, we are simply demonstrating that we can take a step back and look at how the error, be it adversarial or naïve, propagates within the neural network framework to make a more informed and more immediate decision on designing a measure.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Tracking Effects Layer by Layer from Adversarial or Naive Changes to a CNN Inputer

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • James P. Larue

摘要

In the AIPR 2017 Workshop we presented a paper that looked at “The Anatomy of a Neural Network”. The main purpose was to show how to take a simple LeCun-based neural network, with its multiple internal operations, and extract layer-to-layer input/output pairs to form a series of bidirectional association memory matrices (AMMs) from CNN Layers. From these transformations we showed how to achieve a practical solution that met the expectations of the Universal Approximation Theorem; how we could to collapse a six-layer system into a single layer. For the AIPR 2025 Workshop we will examine how a change in the image, be it adversarial or just a naïve-rotation of the image, can be physically tracked across each of the CNN/AMMs, when each is examined in vector form, i.e., we will visually exploit the error from the first layer on to the last layer. There are many approaches to measuring layer vulnerabilities, such as autoencoders, t-SNEs, histograms, and residuals. In contrast to these methods, we do not offer a measure, we are simply demonstrating that we can take a step back and look at how the error, be it adversarial or naïve, propagates within the neural network framework to make a more informed and more immediate decision on designing a measure.