[dropcap]I[/dropcap]n the mid twentieth century, artificial intelligence researchers invented a new type of computational system that could detect patterns in images – a daunting task for previous technology. Because this new system comprised highly interconnected information-processing nodes, resembling the organization and function of the brain, it became known as an artificial neural network.
At that time, neuroscience was still in its infancy, and our understanding of the brain was limited. Scientists knew that neurons could pass signals to other neurons. They had some idea that the connections between neurons were flexible, and that connection strengths could change. And by peering at cells through a microscope it was easy to extrapolate that the total number of neuronal connections in the brain was astronomical. But basic information about the brain’s operation was still mysterious. Nobody had a clue how the human brain’s 89 billion neurons were subdivided into functional groups, how electrochemical fluctuations encoded information, or how neural circuits processed electrical signals. Thus, the similarity between artificial neural networks and biological neural networks didn’t extend very far.
At least, it didn’t initially.
“… the researchers designing systems are also desperately trying to understand how they work.”
Today, neural networks resemble biological brains more vividly. These artificial systems can perform complicated tasks with surprising intelligence: Researchers are currently developing systems that can learn how to drive a car just by observing a human driver, or that can cooperate seamlessly with humans to solve problems jointly. And the secret to the performance of these advanced neural nets is a complex and inscrutable system of connections buried in so-called hidden layers. The more hidden layers a deep learning neural network has, the more remarkable its problem-solving ability – and the less anyone can understand how it’s working.
Hence, we have reached a peculiar stage in the history of technology wherein the researchers designing systems are also desperately trying to understand how they work.
To investigate the intricate computation occurring deep inside neural nets that classify images, one strategy involves systematically feeding the network different images and singling out one hidden node at a time to find out what image properties cause that node to activate. In a neural net that can identify cupcakes in photos, there might be a hidden node that responds to blue stripes angled at 45 degrees. Or, there might be a node that responds to pink frosting in the center of the frame. By discovering the image properties uniquely recognized by each of many hidden nodes, researchers can start to piece together the function of the hidden layers, and how the composition of these layers can decode information about the image – from pixel to cupcake.
“Neural networks are more aptly named than their inventors ever realized.”
This same strategy is a staple of neuroscientific research. Foundational studies of the brain’s visual system homed in on the precise properties of light and the visual field that activated specific neurons in different regions of the brain. With this method, neuroscientists learned that there are numerous brain areas in the visual system that each respond to different aspects of visual images – some neurons encode the region of space that a visual stimulus inhabits, some neurons encode colors, and other neurons encode more complex properties like object identity. And now that these neurons’ functional properties are clear, neuroscientists are able to form theories about how different visual areas connect, work together to decipher visual information, and distribute this information throughout the rest of the brain.
It seems then that neural networks are more aptly named than their inventors ever realized. Neural network researchers are using a strategy to study their creations identical to one neuroscientists use to study the brain, which leads to some thought-provoking speculation: What other neuroscientific research methods could be useful for studying neural networks?
It’s possible to imagine how fMRI, tractography, optogenetics, or event-related potential techniques could be tailored to the study of artificial neural networks. In neuroscience, these popular and powerful methods each capture a different type of data, and so can be used to test different types of hypotheses. The brain is too complex to ever yield complete knowledge of every neuron’s activity at every moment in time, so research questions focus on specific aspects of neural operation: the location of activity in the brain, whether a type of cell is necessary for some behavior, or the time course of a specific neural process. Then, findings from different research programs can be compared and woven together to form a theoretical understanding of how the brain works. This same broad strategy could be applied to the study of artificial neural networks, the ever-increasing complexity of which also thwarts detailed mechanistic understanding.
If advanced neural networks can be directed to analyze their own functioning, would that change how we view ourselves?
If we extrapolate further, to the bleeding edge of neuroscience, we tread into the realm of science fiction. Neuroimaging technologies have been steadily advancing, but the most methodological progress is being made in data analysis. Using the same fMRI data that has been available for decades, neuroscientists are now devising sophisticated statistical tools to answer new questions that were once thought to be unapproachable. Many of these advanced analytical tools, such as multi-voxel pattern analysis, support vector machines, and representational similarity analysis are machine learning applications – they are powered by the same technology that drives artificial neural networks. So, if researchers studying artificial neural networks find success in the adaptation of neuroscience methods to their own work, their efforts might eventually include these recent machine learning applications, at which point neural networks would be deployed in the analysis of themselves.
Introspection, the capacity to gaze inward and reflect on the very mental processes that underlie our inquisitiveness, is often considered to be a defining trait of humanity that sets us apart from other animals. But if advanced neural networks can be directed to analyze their own functioning, would that change how we view ourselves? Would artificially intelligent systems need to be recognized on equal standing with us? Or would we simply need to strike one possible essentially human trait off of the ledger of human nature?
Before we start worrying about losing our unique place in the universe, we can take some small comfort in one likely scenario. Namely, it’s possible that self-reflective neural networks would be more successful in deciphering their functioning than we are as humans. As the great American psychologist William James described, our introspection is “like seizing a spinning top to catch its motion, or trying to turn up the gas quickly enough to see the darkness.” In other words, we have the capacity for introspection, but true introspective understanding is elusive. So our uniqueness would then be preserved: In the club of ineffectual self-reflection, we could still be the sole members.
• The Dark Secret at the Heart of AI. Will Knight, MIT Technology Review. 4/11/2017 https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/
• Hubel D.H., Wiesel T.N. (1962) Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. The Journal of Physiology 160(1):106-154.
• Vogt, N. (2018) Machine learning in neuroscience. Nature Methods 15, 33 (2018)
• James, W. (1890) The principles of psychology. NY, US: Henry Holt and Company.