The Future of AI: A Q&A with Brandon Allgood
Posted 3rd September 2018 by Kieran Chambers
Brandon Allgood is the Chief Technology Officer at Numerate, as well as a co-founder. In addition to being responsible for the development of Numerate’s cloud-based machine learning platforms, Brandon runs the software & data science teams, and is the technical lead on all of the companies internal drug programs, as well as external collaborations.
Here, Brandon talks about what he thinks the future holds for AI and machine learning:
Why is AI so popular now?
Well, there are a couple of reasons. It’s largely driven by the fact that the cost of compute and storage of data have effectively gone to zero. Neural networks have been around for a very long time. The algorithms behind training neural networks, back-propagation, these are not new. What has really changed is the fact that the amount of available data to train these neural networks has increased exponentially, largely due to the fact that we don’t throw a single piece of data away. Now, companies like Google and Facebook are amassing massive data sets. They’re not throwing anything away. They finally have these huge data sets of images and different types of media.
At the same time the cost of silicon chips has plummeted. Now you have these massive amounts of compute. And it’s now not just accessible to the big guys, but it’s accessible by everyone. Companies like Google and Amazon are selling compute at a fraction of a penny per compute hour. Then you have the introduction of new types of compute architectures like GPUs and specialty machine learning chips, which map directly to the problem of neural networks and other machine learning algorithms.
You have cheap chips, cheap storage, and that’s really allowed people to start thinking about new algorithms, modifications to neural networks, and that caused a big surge in the advancement on the ImageNet, a machine learning contest that’s been going on for a very long time in image recognition. That was the spark.
What does this mean for the future of AI?
As far as AI is concerned, we’re only getting started. Generally, I think in the next five years outside of our specialized field, it’s just going to integrate in daily life more and more. We’re going to notice it less and less because the entry of machine learning and AI is going to continually improve to the point where it’s no longer awkward to interact with and becomes more natural. We as engineers, will learn to how to better integrate it into human behaviour.
On the medical side or on the chemistry and biology side, we’re in a different place due to the type of problems that we are trying to solve with machine learning. If you look at where machine learning is excelling in the world today, it’s excelling at kindergarten level skills, such as recognizing letters, learning to read, organisational skills, recognizing images: these are basic human skills.
Now if you think what we’re trying to do in our field, we’re trying to leap over all that and take machine learning and AI through grad school. We’re trying to build machine learning algorithms that can help to design compounds against diseases and that can recognize causal networks in biology.
What’s holding AI back?
One of the problems in our field is that data costs so much. The data that Google and the data on which you see most machine learning that’s very successful, is relatively cheap. We’re storing all of our information on the Internet, on Google, Facebook, and other places. We’re effectively doing it for free. We’re collecting it for them to analyse and run their algorithms. I mean there’s a cost in storing it, but not that much.
Even if you look at data collected by autonomous vehicles, that’s actually very expensive data as well, but not nearly as expensive as the data that we’re collecting where for example in drug design, a single data point can cost $1,000, as high as $5,000. If you think about the fact that you need to synthetize the compound and test it in biological assay. That’s very expensive. So the amount of data in our field will never come anywhere close to the amount of data that you’ll find in these other spaces. Genomic data might get into that realm and genomic testing – should people be willing to donate their genomes. But on the chemistry side, we have a long way to go before it gets cheaper. We’re are going to have to work with small amounts of data in the meantime.
So what can we expect to see in the near future?
The hype is going to die down. I hope it won’t crash and burn because that would set us back. There’s going to be an education on what machine learning can and can’t do. As long as the hype doesn’t get too out of control, AI will start becoming part of every medicinal chemist’s, part of every biologist’s toolbox. It’s going to in many ways replace other tools that scientists have today. We, as algorithm developers and those people on the frontline developing machine learning, still have a lot of work to do in being able to build machine learning algorithms than can combine small amounts of data on a specific topic with a broader set of generalised knowledge. Think of this as giving machine learning “intuition” or getting it to make “educated guesses.”
To be more concise, in the next five years outside of our field, machine learning and AI are going to pervade everything we do. It already does and people don’t understand it. This is just going to get even more so.
In our field in the next 5 years, you’re going to see a broader acceptance of its contributions. We’re definitely going to get some successes. I often get the question, “Show me the drugs in the market that were designed by AI.” This question itself, is non-sensical given the state of the field because the machine learning models we’re building today are so narrow. If they are trained on a dataset about activity, they can’t say anything about synthesisability or reactivity, etc. We at Numerate have machine learning models for most aspects of the discovery process, but humans remain in the loop. So there is no drug that was designed entirely by machine learning models, because machine learning models are still to narrow. There are, however, compounds currently in the clinic that either would not be in the clinic today or would not look anything like they do without the use of machine learning. I can say that definitely.
So, we are going to see some high profile wins for machine learning and there will be a larger adoption of machine learning and AI across the field of medicine going all the way from target discovery, drug discovery, drug development, pharmacovigilance, compliance and beyond.
With time running out and places selling quickly, book now to avoid disappointment.
Leave a Reply