Reducing Bias in Datasets

To have responsible AI, it's incredibly important to build high quality and representative data. You know, I actually think it's, it's critical because at its core this thing I talked about earlier is that, you know, the, the, the data is almost like the food that you feed to these machine learning systems and ultimately you are what you eat. And so the, the sort of, the data is, is sort of the one of the fundamental areas where, where these biases or these, these kinds of really critical real world issues can sort of stem from. And, and really the way we think about is like how do we, how do we build technology, how do we build tools, how do we build, how do we build things to enable either our customers or ourselves to ensure minimal bias and minimal harmful results with, with these machine learning, with the data sets that power these machine learning systems and, or how do we build systems that are going to flag that or identify when those biases may exist as proactively as possible so that then we know that we have to go fix that problem long before it ends up in the hands of a consumer or a, or some sort of critical decision making progress. So for example, we, we recently worked with like a bunch of medical researchers on, on automated medical imaging analysis. You know, this, this sort of exact problem. We worked with the MIT Media Lab and analyzed this, this exact problem. You mentioned of tons and tons of clinical images and found there were a lot more light skin images than dark skinned images. And what we did is we actually basically helped them replenish or sort of debias the data set and add more data to the unrepresented classes, either with real world data or using data augmentation or synthetic data. And we're able to significantly de bias the output or the outcomes from the machine learning algorithm. And so I think, you know, it's, it's not an easy problem in any way. You know, this is actually a sort of like, you know, a, this is one of the quote unquote hard problems of AI in the sense that there's no easy solution. It's not like, you know, I don't expect that next year we're going to get some, you know, crazy machine learning architecture that all of a sudden solves the problem of data bias or machine learning bias. It really is all about the nitty gritty details about what does your dataset look like, how do you know ahead of time when there's bias in the data set and then how do you fix that or how do you proactively resolve that and how do you keep going through that process.