After an apparent privacy-focused shift echoed at Facebook F8, Google executives also emphasized how it has to keep responsible development of AI, privacy and security in mind. Google’s Sundar Pichai talked of bad datasets and and AI bias, “”We want to ensure that our AI models don’t reinforce bias that exists in the world.”… “Its not enough to know if a model works, we need to know how it works.” To tackle bias, Pichai said Google is working on a new model called TCAV ( (testing with concept activation vectors), it shows which elements of the training data are important to reveal biases. For instance, he explains, if classifier that is trained to identify doctors, is trained with data having mostly males wearing coats and stethoscopes, than the AI could assume that being male was an important prediction factor. Similarly, for an AI system detecting skin cancer to be effective, it will have to recognize a wide variety skin tones representative of the entire population. “There’s a lot more to do, but we are committed to building AI in a way that’s fair and works for everyone, including identifying and addressing bias in our own AI models.”

Its worth noting that avoid reinforcing biases using AI and using AI for social good were among Google’s AI principles. The company’s global advisory council for ethical issues around AI was dissolved within days of formation, after thousands of its employees protested the appointment of a right-wing thinktank leader. The eight-member Advanced Technology External Advisory Council (ATEAC) was dedicated to “the responsible development of AI.

AI in healthcare

Google’s life sciences arm Verily just received received European regulatory approval for its AI tool for detection of diabetic retinopathy, a condition affecting diabetes patients which can result in potential vision loss. Thousands of patients are being screened in India and Thailand using the AI model.

Google is now working on oncology using AI, said its head of healthcare products Lily Peng. Google new AI model can detect lung cancers with higher or same accuracy as trained radiologists, she explains. The model was trained using a neural network with de-identified lung cancer scans. Survival rates for lung cancer are 50% when detected at Stage 1, and fall to 2% when detected at Stage 4. “Very early stage cancer is miniscule and hard to see,” she said. “Many late stage patients have in their earlier scans shown subtle signs of malignancy”. Google AI model was able to detect this early signs one year before the patient was diagnosed, making the survival rates increase by 40%.

Flood forecasting using AI

Google sent early flood warnings to users’ phone in partnership with the Central Water Commission (although it isn’t clear when and where, possibly Kerala). Its expanding this detection system to cover the floodplains of the Ganga and Brahmaputra rivers for the coming monsoon season. 20% of flood fatalities in India alone occur due to the absence of consistent early warning systems, according to Jeff Dean, head of Google’s AI division. Google is helping predict flood timing, location, and accuracy, he said.

The system will better forecast where the flood will hit the hardest, and will simulate water behavior across the floodplain showing the exact areas which will be affected. To do this, it will combine thousands of satellite images to create a high res elevation maps and figure out the height of the ground. Neural networks are used to correct the image and then physics to figure out how the flooding will happen.

Predicting the extent of flood

Using satellite images to form high resolution image to detect depth of terrain