A month after releasing the code for one of the core algorithms behind Bing, Microsoft Corp. today made another notable contribution to the open-source community.

The company’s research division today introduced TensorWatch, an internally developed tool created to take some complexity out of artificial intelligence projects. It focuses on one aspect of the development process in particular: debugging.

Ridding code of errors is one of the most laborious and time-consuming tasks in any software project. The process is especially taxing when it comes to AI development, since the inherent complexity of machine learning models means there are far more ways they can break than a traditional program. TensorWatch aims to make bugs easier to spot by enabling developers to visualize their models in interactive graphs.

The tool generates graphs using the data that an AI produces while it’s undergoing testing. According to Microsoft, TensorWatch represents each source of information as a “stream.” A stream can include the output of a model, statistics on how much processing power it consumes and even TensorWatch graphs, among other items.

That approach has the benefit of making it easier to work with the data. A developer can reuse the same stream across several visualizations or make a graph that displays several information streams side-by-side. TensorWatch lends itself to creating everything from simple bar charts to complex, three-dimensional maps that visualize potential bugs in an interactive virtual space.

A user can zoom in on items of interest by writing queries to manipulate their graph. This feature is powered by Jupyter Notebook, a popular open-source application for experimental coding that Microsoft has baked into the TensorWatch interface.

Besides allowing developers to look for bugs in a visual fashion, the tool also promises to make troubleshooting more hardware-efficient. The credit goes to a feature Microsoft’s researchers dub “lazy logging mode.”

In a nutshell, TensorWatch can lower processing overhead by reducing the amount of data that needs to be ingested to find problem patterns. Software teams can have the tool observe only the core variables that show how well a model performs during testing.

Microsoft has shared the code for TensorWatch on GitHub. The tool joins a number of existing AI visualization and debugging tools in the open-source ecosystem. Among them is Embedding Projector, an application created by Google LLC for visualizing the complex mathematical structures that neural networks use to represent data. 

Photo: Microsoft