Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

GitHub - joy-void-joy/alexplainable

May 27, 2024 - github.com
The article discusses the development of Alexplainable, an ImageNet recognizer designed to be explainable using AI-Pull-Requests. The project aims to understand the function of each part of the network in a comprehensive manner. The process involves defining a graph of features, segmenting images, training shallow convolutional neural networks, and refactoring the graph if a node cannot be trained to have good accuracy. The article also discusses the challenges faced during the process, such as some networks not learning the proper task and the resulting feature visualization of the networks not being interpretable.

The article further outlines the remaining questions to be investigated, such as whether switching to AvgMax or MaxPool mid-training allows networks to be more interpretable, whether a claude-generated ontology is fit for recognizing images, and whether forcing the convolution layer to be plain-coded improves overall performance interpretability. The article concludes with instructions on how to use the system, including downloading sorted images, the ImageNet dataset, and Segment anything's vit_l, and running the prepare.py script.

Key takeaways:

  • Alexplainable is a project aimed at creating a semi-transparent ImageNet recognizer using AI-Pull-Requests. The goal is to understand what each part of the network does in a fully understandable way.
  • The project uses a process of defining an ontology, segmenting images, and training shallow convolutional neural networks on leaf nodes. If a node cannot be trained to have good enough accuracy in a fixed amount of time, LLMs are used to refactor the graph further.
  • Early results show that the process of defining an ontology and segmenting images works well, but there are challenges with training networks and interpreting them. Some networks have not learned the proper task when reused globally, while others show the correct behaviour.
  • There are still many questions to investigate, such as whether switching to AvgMax or MaxPool mid-training allows networks to be more interpretable, whether a claude-generated ontology is fit and/or better to recognize images, and whether the convolution layer can be forced to be plain-coded for better interpretability.
View Full Article

Comments (0)

Be the first to comment!