Over the last 10 years, deep learning has transformed artificial intelligence, shaping areas as diverse as investing, online security, shopping and drug development. Now neuroscientists are using artificial neural networks to better understand the biology that originally inspired them. Deep learning provides new methods for analyzing increasingly large datasets and offers new models for exploring brain function.
Current Opinion in Neurobiology explores this trend in its April 2019 special issue, titled “Machine Learning, Big Data, and Neuroscience.” The issue was edited by Maneesh Sahani and Jonathan Pillow, both investigators with the Simons Collaboration on the Global Brain.
Pillow talked with the SCGB about how machine learning is shaping neuroscience, enabling new techniques for better analyzing neural activity, and inspiring novel approaches for understanding vision and other systems. An edited version of the conversation follows.
What sparked the theme for this issue?
Machine learning is having a major impact on all aspects of science, and in neuroscience in particular, so it seemed like a good time to highlight how machine-learning methods and models are impacting how we study the brain.
How is deep learning influencing neuroscience?
In the last five to 10 years, we have been able to use deep learning to train models that have human-level performance on recognizing faces, understanding speech and translating language. These are tasks we used to think humans were uniquely good at. Initially, people thought that by studying the human brain we could understand how the brain recognizes faces, then use that to design new computer applications for computer vision or self-driving cars.
But the opposite happened. People have realized it’s possible to train machines at better-than-human performance. Researchers are using deep neural networks as a model for intelligence, for thinking about how the brain itself works. We can train a network to recognize faces, then go look at the intermediate stages of the network to see to what degree they resemble biological networks. We can train recurrent neural networks or deep neural networks to perform some classification task, then examine the components of that network and ask if they shed light on the computations the brain carries out. What are common solutions to the problems the brain has to solve?
Advances in computing power and methods have also had a major impact on how people analyze data and how they seek to understand it. Neuroscience has undergone a revolution in the amount of data collected; the size of datasets and the number of neurons we can record from is rising rapidly. We increasingly need machine-learning methods to wrangle this data and try to gain insight into it. One common example is sorting neurons from noisy calcium imaging movies. I think of advances in machine learning as a form of statistical signal processing that has given us many new methods for processing the raw data that comes out of a neuron.
Deep learning gives us methods for relating high-dimensional neural data to high-dimensional behavior. Regression models can capture what features of neural activity relate to what features of behaviors. Or what aspects of neural activity carry information about behavior or sensory stimuli.
How are neuroscientists using machine learning to manage large datasets?
Machine learning has been essential for developing new tools for processing calcium imaging, which is noisy and has weak signals. It used to be that every lab designed its own ad hoc methods for finding neurons and extracting activity. Machine-learning tools have made a major impact by providing standardized tools that everyone in the field can rely on. Now you can extract the numbers of neurons and the activity with high accuracy. [See “Analysis Pipelines for Calcium Imaging Data” and “Computational Processing of Neural Recordings From Calcium Imaging Data.”] This also holds for spike sorting with high-density silicon probes. [See “Continuing Progress of Spike Sorting in the Era of Big Data”]
What are some innovative ways researchers are putting machine-learning techniques to use?
Danielle Bassett’s group at the University of Pennsylvania is using machine learning to harness information from large-scale neuroimaging datasets to design new treatments for psychiatric illness, bridging machine learning and network science. [See “Harnessing Networks and Machine Learning in Neuropsychiatric Care.”] You can aggregate many datasets on connectivity or synthesize results of many different kinds of imaging studies to make recommendations about the diagnosis or type of illness and predict treatment responses for different kinds of illnesses.
Raj Rao [at the University of Washington] described advances in brain-machine interfaces. We tend to think of these devices as being able to read out brain activity or move an artificial limb or a cursor on a screen. But Rao talks about how they could be used in rehabilitation after injury to induce plasticity in neural circuits. Instead of just reading out signals, we can also induce patterns in the brain that might enhance plasticity, to restore memory or enhance movement in limbs that lost function. [See “Towards Neural Co-processors for the Brain: Combining Decoding and Encoding in Brain-Computer Interfaces.”]
What about more basic applications?
David Barrett [of DeepMind], Ari Morcos [DeepMind], and Jakob Macke [Technical University of Munich] described combining knowledge from biological and artificial networks to allow the networks to shed light on each other. The cutting edge of AI research is to understand why deep networks are so good at learning certain kinds of tasks. They did a nice job connecting this to why brains are so good at certain kinds of tasks, such as speech processing and sensory tasks. In both fields, we want to understand why networks can learn efficiently from data or learn complex functions. We can use techniques from neuroscience to study deep neural networks or apply deep-learning techniques to the brain.
Shreya Saxena and John Cunningham [Columbia University] wrote a riff on the neuron doctrine, the proposal from the 1800s that the nervous system is made up of discrete cells. Saxena and Cunningham propose instead a neural population doctrine, in which we should think about neural populations as the fundamental computational unit. This is a radical idea. There hasn’t been a lot of work on how groups of neurons working together subserve some computation that is useful for biological functions. If you’re looking at one neuron at a time, you would be blind to this level of computation.
What’s an example of a key neuroscience insight that emerged from machine learning?
I would point to work from Josh Tenenbaum [Massachusetts Institute of Technology]. His work is shedding light on why certain parts of the visual system compute what they do in order to recognize objects. Tenenbaum and collaborators have a way of combining prior knowledge of the structure of the world we have in our brains with the fast inference of neural networks to develop a new way of trying to understand how the brain recognizes objects. That’s something we haven’t made a lot of progress on using traditional methods.
What would you like to see going forward?
I would like to see more opportunities for experimentalists and theorists to work closely together. Still too often, experimentalists collect datasets that they then share with theorists, who do their analysis on it. I want to see more active collaboration between the two, where theorists help design experiments and experimentalists help analyze data. I think that’s increasingly happening. A lot of a papers in the issue were co-authored by experimentalists and theorists.
How can the field enhance that kind of collaboration?
The BRAIN Initiative is a major venue for encouraging experimentalists to talk to theorists, as well as for promoting data sharing. For many years, it was hard to get access to good data. The SCGB is another good example. The first round of funding went to individual PIs or pairs of PIs. In the next funding round, we were encouraged to form larger groups. That’s another exciting development. In this round, I’m part of the International Brain Lab, a group of 21 scientists working together to understand decision-making and foraging behavior using recordings from the entire mouse brain. The collaboration has a nearly even split between experimentalists and theorists, and the two groups have been working closely together on everything from the design of the task, to models of behavior, to new methods for the analysis of electrophysiological data. I think the opportunities for combining different levels of analysis with large-scale behavioral and neural datasets are going to provide a lot of exciting opportunities for new science.