Earlier this year, Microsoft invited us to a two-day briefing on its Bangalore campus about how it is incorporating artificial intelligence (AI) into many of its business solutions, including Azure, Power BI, Teams and Office 365. As some of its business partners explain how these AI-enabled services help them, Redmond-based software giant has showcased its garage-enhanced apps such as Kaizala, Seeing AI, and Soundscape.
In a similar vein to Microsoft, Google invited us to its Ropangi Hills office in Tokyo earlier this week for a one-day briefing entitled “Solution with AI …”. The briefing was led by Google Senior Fellow and AI Lead Jeff Dean. Although Microsoft’s briefing on AI revolved around solutions that address IT business challenges, Google’s briefing addresses solutions aimed at “social good.” Product leads from Google AI explain how the company’s technology is being used in healthcare, environmental protection, agriculture and more. Google has invited some of its business partners to add input and examples during the briefing.
The briefing began with a keynote address by Dean explaining the basics of machine learning (ML), a subset of AI in which a computer is trained to recognize patterns through examples rather than programming with specific rules. He explained how neural networks could be trained to detect very large or complex patterns for humans using relatively simple mathematical functions. The ML model is developed for this purpose.
In addition to hiring their own products, Google offers ML tools with some reference implementation information for researchers and developers to create AI-enabled software. Examples of such tools include Open-Source Tensorflow Software Library, CloudML Platform, Cloud Vision API, Cloud Translate API, Cloud Speech API and Cloud Natural Language API. Google includes ML models in its offerings, including Search, Photos, Translation, Gmail, YouTube, Chrome, and more.
Dean has used the example of an air quality monitoring tool called Air Cognizer to demonstrate how tensorflow is used in everyday mobile app development. Air Cognizer is an app created in India as part of Celestini Project India 2018. It can help detect air quality levels in the surrounding area by scanning images taken with the camera of an Android device. Dean says this is just one example of developers and researchers using Google’s machine learning tools to create AI-enabled apps and services. After Dean’s introduction, other leaders from the Google AI team came on stage to talk to each other about other areas where Google’s ML efforts are making a difference.
Jeff Dean delivered the keynote address
Lily Peng, Google Health’s product manager, came on stage after Dean’s role to talk about how Google’s AI initiatives help in healthcare. “We believe that technology could have a major impact on medicine, helping to democratize access to care, divert attention from patients and help researchers make scientific discoveries,” he said during his presentation. She supported her statement by citing three specific areas where Google’s ML models have seen success: lung cancer screening, breast cancer metastasis detection, and diabetic eye disease detection.
Google’s ML model, according to the company, can analyze CT scans and predict malignant lungs in cancer screening tests. In tests conducted by Google, the company’s model detected 5 percent more cancer cases, resulting in a false positive reduction of 11 percent compared to radiologists. According to Google, early diagnosis can go a long way in treating serious diseases, but more than 80 percent of lung cancers are not detected early.
In detecting breast cancer metastases, Google says its ML model can detect 95 percent of cancer lesions in pathology images. Google claims that pathologists typically detect only 73 percent of cancer lesions. Its model can better scan medical slides, each up to 10 gigapixels in size. Google says it is also more successful at detecting false positives than doctors. Google said it found that the combination of pathologist and AI was more accurate than the one.
Google says that with the help of its sister company Verili, it is becoming increasingly successful in treating diabetic retinopathy. The company is currently using its ML model for detection of diabetic retinopathy in India and Thailand. Google believes that in many places there is a shortage of doctors and specialized equipment, which is one of the reasons why the disease is not detected early, leading to lifelong blindness in patients.
Dean outlines how Google looks at AI and ethics
Julie Catiau, product manager at Google AI, explains how the planet’s wildlife has declined by 58 percent over the past half-century. According to him, Google’s AI technology is currently helping conservationists track the sound of the humpback whale, a endangered marine species, in order to prevent them from becoming extinct. In a bioacoustic project, Google has apparently partnered with NOAA (National Oceanic and Atmospheric Administration), which has so far collected underwater audio data worth over 19 years.
Google says it was able to train its neural network (or “whale classifier”) to detect a humpback whale’s call within a 19-year-long audio data set. During his presentation, Katiyau said it was a big challenge for researchers in part because the sound of a humpback whale could easily be mistaken for another type of whale or a passing ship. Google believes that its AI technology was successful and helpful in the project because listening to a whale’s call in a data set is a huge task that would take a person too much time to complete.
Toffer White, CEO of Rainforest Connections, was one of the many partners invited by Google to attend the briefing. Using a proprietary technology, Rainforest Connection prevents illegal deforestation by logging trucks in rainforests across ten countries, hearing chainsaws and alerting local authorities. Its technology involves re-using solar-charged Android smartphones that use Google TensorFlow to analyze hearing data in real-time from within the rainforest. According to White, deforestation is a bigger cause of climate change than vehicle pollution.
Fabriadi Pratama, co-founder of the Gringo Indonesia Foundation, was one of the many partners invited by Google to the briefing. The Foundation, a recipient of the Google AI Impact Challenge, is currently using Google’s ML model to identify waste types using image recognition in the city of Denpasar, Indonesia. Pratima said in her speech that the project is effectively helping the foundation make plastics in a city where there is no formal system for waste management.
Raghu Dharmaju, Vice President, Products and Programs, Wadhwani Institute for Artificial Intelligence, was one of the partners invited by Google to attend the briefing. The institute uses a proprietary Android app with pheromone traps to scan crop samples for signs of pests, which can ruin a farmer’s crop on a large farm in India. The app uses the ML model created by Google. In his presentation, Dharmaraju said that the solution developed by the institute was significantly effective in detecting pink ballworm in cotton crop in India.
Sella Nevo, a software engineering manager at Google AI, took to the stage to talk about the company’s flood forecasting initiatives. According to him, dated, low-resolution elevation maps make it difficult to predict floods in any area. SRTM, a provider of altitude maps, highlighted data from nearly two decades ago, he said during his presentation. In a pilot project launched in Patna last year, Google was able to create high-definition altitude maps using its ML models using data from satellites and other sources for flood forecasting. It was then able to warn users about the floods at Gandhi Ghat. The flood alert was sent as a notification on the smartphone.
“The number one problem is data access, and we’ve tried to tackle it. With different types of data, we find different solutions. So, for altitude maps, the data simply doesn’t exist. We worked on different algorithms. For different satellite data, we bought most of it and put it together, “Nevo told us in an interview. According to him, Google is trying to create an elevation map that can be updated every year, unlike the ones offered by SRTM.
L to R: Jeff Dean, Sagar Savla, Julie Catiau, Sela Nevo, Lily Peng
Sagar Savla, a product manager at Google AI, took the stage to talk about Google’s Live Transcribe app. Currently available in 70 languages, the app helps deaf and hard of hearing people communicate with others by transcribing real-world speech to on-screen text. The app has been developed using Google’s ML model to ensure accuracy in transcription. For example, the app can tell if the user wants to say “new jersey” or “a new jersey” depending on the context of the sentence. Talking about the app and its development, Savla said that she used it with her grandmother, who, despite being hard of hearing, was able to join the conversation using the Live Transcribe app in Gujarati.
Julie Katiau returns to the stage to talk about Project Euphonia, a Google initiative dedicated to creating speech models that are trained to understand people with speech impediments. The initiative could combine discourse with computer vision in the future, he said during his presentation. For example, people with a speech impediment due to a nervous condition may use gestures such as blinking to communicate with others. Cattiau said the company’s ML models are currently being trained to recognize more gestures.
Tarin Klanuwat, a project researcher at the ROIS-DS Center for Open Data in the Humanities, went on stage about an ancient cursed Japanese script called Kuzushiji. Although Kujushiji has millions of books and more than a billion historical documents on record, less than 0.01 percent of the population today can read it fluently, he said during his presentation. He fears that this cultural heritage is in danger of becoming inaccessible in the future due to its misuse in modern texts.
Google says Turin and his co-researchers trained an ML model to recognize Kuzushiji characters and transcribe them into modern Japanese. According to Google, the model takes about two seconds to copy an entire page and about an hour to copy an entire book. According to test data, the model is currently capable of detecting approximately 2,300 character types with an average accuracy of 85 percent. Turin and his team are working to improve the model for the preservation of captive cultural heritage in Kujushiji texts.
Google has made sure that it is moving in the right direction when applying machine learning in the right way for social reasons. In the future, we can expect Google to take on more projects like this, where neural networks are trained to understand data sets that contain keys and clues to insoluble problems in areas that have never been tried before. At the same time, more developers and researchers should be able to incorporate Google’s open-source tensorflow library into their projects as long as Google continues to provide support and reference material for it.