I was working on some Machine Learning for a few different projects I have going on right now and found myself taking a picture of a house. The algorithm told me, “I am 68% sure it’s a purse.”
Now, that would be completely normal for anyone starting out in AI. There are many complexities in training the models properly. You can use Google’s generous models like Inception, ResNet, MobileNet, etc. and get pretty decent results. These models took plenty of time and resources to build and they can do a really good job of identifying main objects in an image.
So what happened? I live in the Middle East and the algorithm didn’t know a Middle Eastern house as well as it knew a European victorian house or a modern Silicon Valley apartment building.
Bias is nothing new. Even in AI circles it is also a lengthy and worthy ongoing discussion. The West has been throwing all of their resources towards “winning” this AI race and naturally they forget the biases held in the selective training of models.
The business intelligence and anthropological understanding we can gain from investing in AI is absolutely worth the effort. However, when we train new AI / ML models we must understand the information on which we train the model as well as our blindspots in the selection of our data/streams.
Chimamanda Ngozi Adichie expresses this concept well in her Ted Talk, The Danger of a Single Story, where she outlines the narratives she grew up with. At one point she says that when asked to draw a house, or ask, “how’s the weather,” or talking about tea time and how unnatural that was to her culture. She had adopted that speech, those appearances, and those concepts due to the large amount of Western literature she had access to. An African, indigenous to her area, would never ask, “How’s the weather?” because it never changed.
So as we are progressing with AI, how can we better average out the AI Cowboys of the West with the rest of the world’s indigenous and unique experience of life? How can AI be trained to produce a truly multicultural neurological network?