The Significance of Purdue’s New Tool in Understanding Neural Networks
In a world where artificial intelligence is becoming more and more present every day, understanding neural networks is becoming increasingly important. Neural networks are a type of AI inspired by our brains, but their complex architecture makes it difficult to trace errors, limiting their use. However, a new tool developed by Purdue University is changing that.
The Tool in Action
The tool, featured in a paper published in Nature Machine Intelligence, makes finding errors in neural networks as simple as spotting mountaintops from an airplane. According to David Gleich, a professor of computer science at Purdue, the tool allows us to see what the network is trying to communicate, which can help identify errors in high-stakes situations.
Using the Tool
The code for the tool is available on GitHub, and it comes with use case demonstrations. In testing, the tool caught neural networks making mistakes in databases of everything from chest X-rays and gene sequences to apparel. For example, the tool found a neural network mislabeling car images as cassette players because the pictures included tags for the cars’ stereo equipment.
Understanding Neural Networks
Neural network image recognition systems process data to identify images using a weighted firing pattern of neurons. However, the decision-making process of these networks is like a “black box” of unrecognizable numbers across multiple layers, making it difficult for humans to understand how they arrive at their conclusions.
Rather than tracing the decision-making path of a single image, Purdue’s approach visualizes the relationship that the computer sees among all the images in a database. By using a Reeb graph to map these relationships, the tool makes it possible to see spaces where the network can’t distinguish between different classifications.
Purdue’s development of a tool to better understand neural networks is a significant step forward in the understanding and use of artificial intelligence. Users can now access the tool on GitHub and learn directly from the developers. Going forward, the field of artificial intelligence can expect increased accuracy and understanding with the use of this new tool.