Google Brain AMA: DeepMind Collaboration, Healthcare Progress, Differential Privacy, And More

The Google Brain team, which has worked on projects such as RankBrain for Google Search, SmartReply for GMail, Google Photos, and Google Speech Recognition, did a Reddit "Ask Me Anything" (AMA) in which it unveiled its relationship with the separate DeepMind team, what it thinks about quantum computers, and much more.

Backgrounds

Perhaps surprisingly, not every single member of the Google Brain team has a PhD in machine learning. Many of them started with backgrounds such as graphic design and art history, journalism, economics, and English literature. One of the members even said he lacks a university degree, although he did also mention he taught himself programming.

What all have in common is that they channeled the skills they’ve acquired from different backgrounds into something that can be used effectively on Google Brain projects. Of course, the team also has plenty computer scientists and neuroscience PhDs on board as well, though.  

Google Brain And DeepMind

Many were curious to learn what’s the difference between Google Brain and DeepMind, and why does Google have two machine learning teams. Google Brain is Google’s in-house team of machine learning experts, whereas DeepMind was a UK company that Google acquired a few years back due to its innovative approach to artificial intelligence.

The Google Brain members said that they’ve collaborated with the DeepMind team in the past. In fact, one of the team’s interns even helped shape a core component of DeepMind’s AlphaGo AI, which allowed it to learn by playing against itself.

Considering the Google Brain team is located in Mountain View, California, and the DeepMind team is based in the UK, there can’t be as deep of a collaboration between the two as they might prefer. However, Google Brain seems to visit DeepMind’s HQ relatively often, such as when Google switched from the Torch deep learning framework to Tensorflow, and the Brain team needed help with the transition. The two also have regular meetings about using machine learning for healthcare.

Quantum A.I. Lab

Google also has a third team focusing on next-generation machine learning technologies, but this one is focused only on quantum computing-related technologies. The Quantum A.I. Lab team has had its own recent breakthroughs, such as being one of the first to create a small universal quantum computer, and then accurately simulating a hydrogen H2 molecule on it.

However, the Brain members said that their teams don’t collaborate much because their work is so different at this point in time. The quantum technology team is in the very early days of building a universal quantum computer, which may very well revolutionize everything from material science to medicine to even artificial intelligence itself. However, there is quite a way to go until that happens. In the meantime, Google Brain and DeepMind work on artificial intelligence projects that can have a real impact today, while running on conventional computers.

Healthcare

Without a doubt, the idea of using advanced machine learning technologies in the healthcare sector is one of the most exciting because it holds so much potential to help humans cure diseases.

Both the Google Brain and DeepMind teams have been working together to apply deep learning techniques to diagnosing Diabetic Retinopathy, a leading cause of preventable blindness. More such health projects should come later, but Google will probably want to tackle some of these diseases one by one at first.

Differential Privacy

Apple received a significant amount of attention this year when it started implementing differential privacy mechanisms for its data collection. Differential privacy techniques allow companies to gather meaningful data from groups rather than individuals, thus preserving a higher level of privacy for each individual user.

It’s not clear yet whether this is a top Google priority, but one of Google Brain’s members said that he’d be working on merging deep learning and differential privacy next. Considering that Google has already gotten in some hot water for accessing too much patient data in the UK, it may be a good idea for the company to pursue strong privacy techniques that can be applied to its data collection. Google could more easily obtain access to patient data if it can cryptographically guarantee that each individual’s data is truly private, even when the data is mined for information.

Google has three core teams working on artificial intelligence in various ways, which could result in all sorts of potential breakthroughs over the next decade or two. With the company now building its own custom chips for machine learning and having one of the most popular deep learning frameworks at the moment, those breakthroughs may even come sooner than we can expect.