MICrONS, short for Machine Intelligence from Cortical Networks, is a cross-institutional project that has generated a publicly available dataset encompassing the morphology of over 200,000 cells and over 500 million synapses from a 1 cubic millimeter piece of mouse visual cortex (along with a smaller, proof-of-principle dataset). The data was collected over five years using electron microscopy to resolve the fine details of cells and their connections. The goal of the project was to collect data regarding synaptic connectivity and brain wiring for the purposes of improving an area of machine learning called computer vision. This data is also informative for neuroscientists and clinicians hoping to study how wiring can influence cognition, behavior, and neurological disorders.
First researchers at the Center for Neuroscience and Artificial Intelligence at Baylor College of Medicine recorded activity from neurons in the visual cortex while the area was activated, using a movie as natural stimuli. Then, scientists at the Allen Institute for Brain Science sectioned the brain into 27,000 slices and 150 million images were taken using an electron microscope. From there, the images were sent to Sebastian Seung’s lab at PNI, so that the cells and circuitry could be reconstructed into three dimensional images. The machine learning that took place at PNI initially required researchers to train artificial neural networks to recognize the features of cell bodies, dendrites, axons and synapses. To accomplish this, researchers in Sebastian’s lab would manually color example images by hand for weeks until they felt that the neural network could accurately label them independently, thereby saving the researchers from having to manually label millions of images by hand, a clearly impossible task. Ultimately, this pipeline and collaboration between the Baylor College of Medicine, the Allen Brain Institute and the Princeton Neuroscience Institute will allow researchers to associate neural activity and computations within visual cortex with the ultra-detailed morphology of these cortical neurons and their connections.
Visual cortex was used for this experiment because of its ability to extract and detect specific features from a raw set of stimuli. Due to this feature, the wiring diagram of this region was influential to the design of the first artificial neural networks and continues to play an important role in a subset of machine learning called computer vision; one of the goals for this project was to enhance the ability of artificial neural networks to identify novel features from a raw set of data or stimuli. In addition, MICrONS researchers hope that other scientists will utilize this publicly available data to perform their own analyses, thereby expanding the contributions that MICrONS will have throughout the neuroscience and computer science communities.
The data is hosted online by the Brain Observatory Storage Service and Database (BossDB). Both Amazon Web Services (AWS) and Google have also contributed to storage of data, tools and making the data freely accessible. The project was funded by the Intelligence Advanced Research Projects Activity (IARPA), National Institutes of Health (NIH), National Science Fund (NSF), Allen Institute for Brain Science and Princeton University.
by Chris Suriano