Artificial intelligence projects usually rely on existing code. Engineers sift through the AI models available from sources such as academic publications, pick the one deemed most suitable and then customize it to their requirements.

Canada’s DarwinAI Inc. wants to reduce the amount of work involved in the last two steps. The startup today launched from stealth with $3 million in funding and a specialized development platform that promises to automate AI projects.

DarwinAI’s software is based on research conducted by co-founder Alexander Wong at the University of Waterloo, where he serves as a professor with the systems design engineering department. The platform can customize an existing AI for the requirements of an application using its own built-in neural network.

According to DarwinAI, the software works by deriving a series of new models from the original based on user-defined parameters. Engineers can then pick the version with the particular combination of characteristics that best matches their target use case and, if necessary, make further enhancements. The startup said AI models created with its software are not only faster but also considerably more efficient.

DarwinAI claims that to have seen “extremely impressive” results so far. In one project, the startup’s software was used to generate an image classification model that proved 4.5 times more efficient than an AI created with Google LLC’s cloud-based AutoML service. DarwinAI said it also outdid Learn2Compress, an open-source tool created by the search giant to help developers build machine learning models for mobile devices.

In another project, DarwinAI said that the platform enabled engineers to create a version of Nvidia Corp.’s DetectNet object detection AI with four times better performance. The model came out ahead despite being 12 times smaller than the original, the company claimed.

DarwinAI enables developers to analyze the inner workings of the neural networks they create, which makes it possible to verify that optimizations don’t undermine accuracy. The ability to explain how an AI reaches a certain decision is particularly important when sensitive data is on the line. Organizations in areas such as healthcare, for example, need to clearly demonstrate that their software is drawing the the correct conclusions.

DarwinAI claims to be working with customers in the automotive, aerospace and consumer electronics industries. The startup is backed by Obvious Ventures, iNovia Capital and unnamed angel investors.

Image: Unsplash