How Machine Learning is Used for Mapping the Brain
This article explains the world of connectomics and why artificial intelligence is essential to its success. In this article I go over 5 main points:
- What is connectomics?
- What are the implications of a full connectome?
- Why is AI necessary for mapping the brain?
- Key Takeaways
Basically, I start with an explanation of connectomics, then dive into AI in relation to this process. So, let’s get started!
- What is Connectomics?
Connectomics is a subfield of neuroscience that uses electron microscopy to image the brain in 3D nano-resolution. It aims to understand the brain and its functions by looking at its complex wiring. Accomplishing this requires cutting brains 1000 fifths as thin as a human hair, and then imaging everything in that section with an electron microscope.
These pictures can be thought of as frames of a movie, so if you put them one after another you are moving through the space of the brain, allowing you to see every neuron, and every connection from neuron to the next, or synapses. Synapses are the vesicles that contain neurotransmitters, which is how chemical and electrical messages move from one neuron to the next, through exocytosis.
Many neuroscientists believe that memories are stored primarily in the synapses between neurons, and hypothesize that new memories are formed when these synapses weaken or strengthen, as well as when new synapses form between neurons.
However, the whole human brain has a mass of around 1500 g, and the cerebral cortex has about 16 billion neurons (accounting for 20% overall) and 150 trillion synapses. This amounts to 160K km of axons in white matter. As you can imaging, connectomics, or mapping all of this, takes a lot of effort, as well as digital storage. But before we get into all of that, let’s go over the meaning of mapping a full connectome, and how this could alter how we understand multiple fields of science, including psychology and neuroscience.
2. What are the Implications of a Full Connectome?
A connectome is the complete map of the neural connections in a brain. The human connectome is currently incomplete, and there is only one connectome known: a worm with 300 neurons, and 7000 connections. A human has 100 billion connections. It took 12 years of work to get the C. elegans connectome, meaning that if we want to get anywhere near completing the human connectome, we need more automated and sophisticated technology to speed up the process (AI spoilers!!).
But ultimately, scientists are studying connectomics because they believe that human mental behavior, such as depression, schizophrenia, and intelligence correlate to specific features in the brain. So, the goal of the Human Connectome Project (HCP), is to look at individual features in the brain that relate to behavior, high level cognitive, emotional and motivational behaviors, and find commonalities. This could lead to increased discoveries of why humans do, feel, and think as they do, building on what we already know to become more precise and accurate.
If you still aren’t excited about connectomics, it’s literally in movies. Check out this clip from Iron Man 3 to see a cool connectome reference: https://www.youtube.com/watch?v=rggpAI__HDo&list=PL-oIXvOkx-RCK4U2lck3GbCRsXfBLEG-R&index=163
This isn’t a fantasy, this is a possibility!
However, as previously mentioned, the mapping the full connectome is an extensive process, and takes massive amounts of data. There is no room for mistakes.
Why is AI Necessary for Mapping the Brain?
If you haven’t gotten the feeling yet, I’ll tell you again that mapping the connectome is a painstakingly slow process.
This slow process is an image segmentation problem. Before, the process was separated into two parts: using an edge detecting algorithm to find the boundaries between neurons (classification), and another to group together parts that were not separated which would provide the shape fo the full neuron. Recently, however, scientists have started using Recurrent Neural Networks as well as Recurrent Convolutional Neural Networks to find a specific pixel location and then fill in the region with extreme accuracy, even predicting which pixels are a part of the same object (neuron) as that original location in an iterative process. This speed up the connectomic process by years.
A Recurrent Neural Network (RNN) works for connectomics because as input data flows into the model’s single layer, it is processed like a traditional neural network, but with current context. After processing, data is output with new context. Then, the network recurs its new state to itself . This goes on until all data points are processed. Since the context changes at every step, RNNs are good for sequential data, which is exactly how the images from the electron microscopy appear. Remember when I said it was like playing through a film? The RNN, combined with the Convolutional Neural Network’s image processing, is kind of like making sure every frame of the movie matches up to the last.
These algorithms learn from both examples and experience to produce its output, or prediction. They have already been applied successfully in some other fields you may know, such as facial recognition. Why not use it for neural networks?
At SyConn, neuroscientists mapping the brain of a songbird used Convolutional Neural Networks. By training the algorithms to recognize synapses and neuron forms, the Syconn network can map the brain to a point of such accuracy that humans don’t even need to check for errors, the algorithm does that too.
With AI, we get one step closer to unlocking the secrets of the mind.
Key Takeaways:
So, let’s review the big points of connectomics and AI.
- Connectomics is the mapping of every neuron and synapse in the brain in order to learn more about how it functions, such as our personalities and emotions, as well as our mental disorders.
- This is a tedious process, because of the amount of neurons and connections the human nervous system has.
- Because of the previous point, AI is required to complete the job faster and with greater accuracy.
- Neuroscientists work with Recurrent Neural Networks and Convolutional Neural Networks to both predict the connectome and check for errors.
