If you’re looking for AI, you’ve probably heard about the amazing developments made by Google’s artificial intelligence (AI) research lab, DeepMind. The AI program developed by DeepMind is considered one of the world’s most advanced, and is credited with the first game of Go to win over a human in just one try. But what is DeepMind not doing? And how is it improving our lives?
The company first approached British public hospitals to develop software using patient data from the Royal Free London NHS Foundation Trust. Although the company has no experience in providing healthcare, they developed the software using sensitive medical data from the trust. On 18 November 2015, data containing sensitive medical records of millions of patients started flowing into third-party servers. The research was approved by the Royal Free and Google, but DeepMind did not publish its algorithms until June 2016.
Google has said repeatedly that AI is crucial for the future of Google, and it has pressured DeepMind to commercialize its research. The company has used its research to increase battery life in Android devices, as well as to cut the cost of running data centers. Meanwhile, DeepMind continues to incur losses despite the fact that its revenue keeps rising. In 2019, the company lost PS477 million, or $660 million.
The company has made several breakthroughs in AI, but these advancements have not translated into commercial success. AlphaStar, a reinforcement learning system, has mastered the real-time strategy game StarCraft 2. While AlphaStar has been a breakthrough in artificial voice quality, the company has struggled to translate these discoveries into viable business models. Its research and development has been unprofitable without a commercial model.
After the success of AlphaGo, DeepMind hired a handful of biologists to tackle the problem of protein folding. The predicted structure was very similar to the one later determined by cryo-EM. The team that helped develop AlphaFold2 was led by Stephen Brohawn, a molecular neurobiologist at the University of California, Berkeley. The team’s predictions were accurate within a half-atomic margin of error, according to the published study.
Another example of DeepMind’s recent work is improving cancer diagnosis. Its research lab at King’s Cross, London, is currently working with Google’s AOI health research team, as well as the Cancer Research U.K. Centre at Imperial College London. One goal is to improve breast cancer detection. Thousands of cancers are missed due to mammogram scans. This leads to false alarms, but DeepMind hopes to improve the accuracy of mammograms.
But the AI team at DeepMind has a different approach. Instead of relying on “deep learning neural networks,” they should hire more researchers with different viewpoints. The AI team has used evolutionary algorithms, and Hinton’s “capsule networks” approach aims to make algorithms as complex as possible. The researchers at DeepMind say their new approach is encouraging. But, the researchers also note that this is only a small sample.
While Streams describes itself as a clinical app, it doesn’t mention how many medical records DeepMind has stored. The company is currently testing its software and preparing to deliver the data to clinicians. In the meantime, they’re discussing technical infrastructure and the ethical implications of this process. There’s no doubt that DeepMind has a lot to offer, but it’s unclear what its exact role will be.
The AI has the potential to predict protein structures. Having a way to predict protein structure would be an incredible boon to medicine and the life sciences. This method would help doctors discover drugs faster. AlphaFold, a DeepMind team, came out top of the table at the latest CASP, and their AI model outperformed all the other teams. The AI could herald a new revolution in biology by unlocking the structures of proteins.