Building neural networks that smell like a brain
building neural networks that smell like a brain

Building neural networks that smell like a brain

Portrait of Guangyu Robert Yang

Credit: Guangyu Robert Yang

Advances in artificial intelligence (AI) have given machines the ability to see and speak. Training an AI system to smell, however, is a nascent endeavour. Last year, researchers at Columbia University in New York City developed a neural network that evolved an olfactory system markedly similar to that of the fruit fly Drosophila melanogaster (P. Y. Wang et al. Neuron 109, 3879–3892; 2021). The study’s leader, Guangyu Robert Yang, is now principal investigator of the MetaConscious Group at the McGovern Institute for Brain Research at the Massachusetts Institute of Technology in Cambridge. He spoke to Nature about the use of machine learning to process odour information, and the wider implications for the study of the brain.

How do you make a machine that can smell?

For the most part, it is a matter of building good sensors. There are many types in the olfactory system; the fruit fly, for example, has around 50. People have even more. They give us a rich sense of smell, and enable us to tell the difference between different types of chocolate, and check whether food is fresh. These abilities rely on having good, accurate sensors, and incorporating those into machines is hard.

Building vision into machines has the opposite issue: researchers have the sensors, but the circuitry — the network — has been really difficult to construct. We don’t fully understand the circuitry involved in olfaction, but we think it is a simpler problem than for vision.

Where do you start?

Generally speaking, we have a good understanding of the architecture of the fruit fly’s olfactory system: input neurons project to a second layer of neurons, then a third. And we know the number of neurons in each layer. This happens to resemble an artificial neural network, which means we were able to build a network that is structured in a similar way to a fruit fly’s olfactory system using standard machine-learning tools.

After we had that basic network in place, we needed to train it to do something — and that was trickier. Engineers working on machine vision have a large library of images that they use to train neural networks. We didn’t have anything like that for olfaction, so we had to develop an artificial data set. These data sets were simple compared with natural odours, and were designed around one essential property: that mixing multiple odours from a single category — apple, for instance — results in a combined smell that is still recognizably apple.

Once we trained the network using this data set, we found that the system evolved the neural connectivity that we had already observed in the fruit fly — it was almost the same system of signal processing at work in the animal and the machine. We were surprised by that, given how we trained the network. Such a simplified artificial data set would probably not work for vision.

What does the similarity between these networks tell us about olfaction?

One point of criticism that has been directed at neural networks is that the way in which they are trained — by being fed lots of data — does not reflect the biology. And therefore, even though they produce reasonable results, people say that the system can’t be relied on for understanding biological systems. But our neural network evolved an olfactory system in a different way to biological evolution, and yet we end up with the same result. Our model prioritizes efficiency, just as natural selection does, and reaches the same conclusion on how best to perform the task.

What’s the next step for modelling the brain?

Computational models have been developed for many different systems in the brain over the past ten years. The holy grail now is putting them together; these kinds of multi-system model are the main focus of our laboratory. It is a huge challenge — we don’t really know how these systems work together in the human brain, and finding out involves a lot of collaboration with neuroscientists in what are currently considered separate fields. But there has been a clear path of improvement in developing neural networks that behave like the brain, and progress is only going to accelerate. I think that in 20 years we will have made tremendous progress towards understanding how the brain functions as a whole, and these tools will help us to do that.

This interview has been edited for length and clarity.

This article is part of Nature Outlook: Smell, an editorially independent supplement produced with the financial support of third parties. About this content.

Go to the source link

Check Also

COVID-19 Has Left Millions Of Students Behind. Now What?

COVID-19 Has Left Millions Of Students Behind. Now What?

If a kid isn’t keeping up with peers academically, summer school seems like a no-brainer. …

Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.