Lattice is committed to promoting prudent, fair and equitable development of AI applications

One of the most exciting topics in the tech world is artificial intelligence (AI) . From smart robots in science fiction movies to the intelligent functions of real-world connected devices, AI has become one of the most powerful technologies today. However, as Spider-Man Peter Parker said: With great power comes great responsibility. The same is true for artificial intelligence. Companies that develop AI products and services have begun to be keenly aware of:

There could be problems if AI is not used in a careful, fair and equitable way.

An interesting example is the latest version of Lattice Semiconductor's sensAI solution set and its application in client devices such as PCs. Lattice has worked with major PC OEMs to combine low-power FPGAs (such as the CrossLink-NX series chips) with sensAI version 4.1 to support a range of applications, helping to improve user experience and extend laptop battery life.

These applications all use data from the PC's built-in camera to analyze the laptop user, the person behind the user, the environment around the PC, etc. as sensor input sources. The FPGA uses a trained, AI-based inference model to analyze the image data and then perform a variety of different operations. User presence detection decides to turn the screen on or off based on whether the user is detected. Attention tracking detects whether the user's attention is on the screen or shifted to somewhere outside the screen, and performs similar operations to save power. Bystander detection determines whether other people are peeking behind the user, and can turn off the screen or take other measures to protect the privacy of the data on the screen. Finally, the face framing function will ensure that the video collaboration tool can get the best possible user image and appropriately cropped image from the PC's camera.

⚠ WARNING

The key to all of these applications is that they need to accurately identify a variety of people. This seems simple, but in fact it is very challenging, especially when it comes to identifying people of color. However, many image datasets used to train AI models do not contain enough images or enough images of people with different skin colors. As a result, people with darker skin are often not accurately identified, causing applications and features to perform poorly on some populations. This unreasonable lack is not only disappointing, but also a direct manifestation of implicit bias that can permeate technology.

To avoid these kinds of problems, developers focused on AI need to be more serious about the types of datasets they use to train models and the scope of testing of model results. It’s this thoughtful, more empathetic approach to AI development that can help underrepresented groups. After all, why should the color of your skin or the way your hair is styled affect how well the technology functions? It clearly shouldn’t be, and it requires a determined, focused effort on the part of developers.

As companies continue to develop their AI software tools, they need to consider these and many other types of use cases, not only by carefully developing AI applications from a fairness and ethical perspective (as Lattice Semiconductor has pledged). And because the data used to train the models built by these tools is so important, people are increasingly looking to expand datasets, using multiple public dataset sources, and many companies are also seeking specially constructed datasets with different skin tones, headwear, and other types of elements that have been previously overlooked.

Only through these thoughtful and conscious steps can companies avoid invisible bias in today’s AI models and provide better, more accurate, and more inclusive user experiences. While many organizations may not have considered this before, there is no doubt that it will become a critical issue of widespread concern in the future.