The most detailed 3D map of the human brain is published in Science! GPT-4 parameters are only equivalent to 0.2% of humans

The human brain tissue is the size of a sesame seed, and its synapse size is equivalent to one GPT-4!

Google and Harvard teamed up to conduct nanoscale modeling of a partial human brain, and the paper has been published in Science.

This is the largest and most detailed replica of the human brain to date, showing for the first time the network of synaptic connections in the brain.

With ultra-high resolution, the reconstruction, called H01, has revealed some previously unseen details about the human brain.

Corresponding author of the project, Professor Lichtman of Harvard University, said that no one has really seen such a complex synaptic network before.

This modeling result will help to gain an in-depth understanding of the workings of the brain and inspire further research on brain functions and diseases.

It is also worth mentioning that this study involved 1 cubic millimeter of human brain tissue, but the amount of data generated was as high as 1.4PB.

According to a study based on the volume of the human brain, modeling the entire human brain would generate 1.76 ZB of data, and the current most advanced supercomputer has a storage capacity of only 7/10,000 ZB. Less than 0.4% of a single state of a single human brain.

Even if we take all the servers on the entire Internet, we can only store 9 human brains.

At the same time, 1 cubic millimeter of brain tissue contains 57,000 cells and 150 million synapses, and the number of synapses in the entire brain is as high as quadrillion .

In contrast, the number of parameters of GPT-4 is only 2 trillion, which is only 0.2% of the number of synapses in the human brain. According to this calculation, it is the size of a sesame seed when placed in the brain.

Some people lamented that AGI may be far away again...

Nanoscale modeling brings new discoveries

Specifically, the researchers obtained a sample of temporal lobe cortex tissue from a 45-year-old female epilepsy patient, which was approximately 1 cubic millimeter in size.

After the samples were quickly fixed, stained and embedded in resin, the researchers used an ultramicrotome with an automatic collection device to cut 5019 continuous sections with a thickness of approximately 33.9 nanometers.

The researchers then used a multi-beam scanning electron microscope to image each slice at a resolution of 4×4 nanometers/pixel, obtaining original two-dimensional image data with a total size of approximately 1.4PB.

Next, the researchers used computational tools to splice and align these two-dimensional images, and reconstruct three-dimensional voxel data.

After that, they used a machine learning algorithm called flood-filling networks (FFN) to segment the neuron morphology of the entire voxel, and manually corrected the segmentation errors. , and finally reconstructed the three-dimensional shape of all cells, synapses, blood vessels and other structures in this 1 cubic millimeter of brain tissue.

FFN was proposed by Google Brains in 2018. The basic idea is to start from a seed point and recursively expand around it, marking all voxels connected to it until it encounters the background or the boundaries of other objects.

At the same time, they also used machine learning models to automatically identify synaptic locations and distinguish excitatory and inhibitory synapses.

In the end, the team successfully modeled 1 cubic millimeter of brain tissue at the nanometer level, containing more than 50,000 cell nuclei and 150 million synapses, as well as 230 mm of super Small veins.

On this basis, by analyzing the reconstructed cell morphology, the researchers identified the main cell type composition of the brain area.

Of the total 57,180 cells, 49,080 are neurons and glial cells, and 8,100 are related to blood vessels. Among neurons and glial cells, the number of the latter is about twice that of the former.

Among neurons, 65.5% are pyramidal neurons with spikes, and 29.1% are interneurons with smooth processes; among glial cells, oligodendrocytes are the most common.

Researchers developed a machine learning model to automatically identify synapse locations and their types (excitatory/inhibitory).

This brain area contains a total of about 150 million synapses, of which 111 million are excitatory synapses and the other 39 million are inhibitory synapses. The distribution of excitatory and inhibitory synapses at different cortical levels There are also some differences in density.

By analyzing the synaptic input received by each neuron, the researchers found that the vast majority (96.49%) of axons only form one synapse with their target cell, but a few axons can form multiple synapses. (up to more than 50) and establish a particularly strong connection with the target cells.

Further analysis found that such polysynaptic "strong connections" are prevalent in both excitatory and inhibitory axons, and their number is significantly higher than that formed randomly expected levels at synapses.

The researchers speculate that among a large number of random weak connections, a specific few axons may regulate the activity of neurons through deliberately formed strong connections.

In addition, the researchers also analyzed a special type of pyramidal neurons in detail.

There are two mirror-symmetrical orientations of the basal dendrites of these "triangular" and "compass" cells, suggesting that they may have different functions.

However, the author also stated that the relevant samples were from epilepsy patients. Although no obvious pathological changes were found under a light microscope, it cannot be ruled out that long-term epilepsy or drug treatment may have an impact on The connections, or structure, of cortical tissue have some more subtle effects.

In other words, the universality of this model may need to be further verified, but at least it has unveiled another layer of the synaptic network.

In order for people to use the modeling results to discover more mysteries, the research team has made all original data, modeling results and related tools open source.

All data tools are open source

The author has established an online interactive data visualization platform Neuroglancer, which other researchers can use to explore the H01 data set at different scales.

It includes all original electron microscope slice images, as well as segmentation results of neuron morphology, synapse location and excitability/inhibition, as well as labels of different types of cells. Users can flexibly observe the data set micro and macro structures.

In addition to the data, the author also open sourced CREST, a tool for exploring synaptic connections between neurons, and CAVE, an online collaborative correction platform deeply integrated with Neuroglancer, to help other researchers Explore and analyze this unprecedented large-scale human brain data set from every angle.

The author said that making this result open source will provide the academic community with a physical basis for studying the structure and function of the human brain, and provide a reference for disease research.

Although H01 has brought unprecedented detailed information, compared with the entire human brain, these data are just the tip of the iceberg of this huge organ. In the future, similar studies will be needed on more regions and levels of the human brain. For nanoscale imaging and three-dimensional reconstruction, the author also calls on the academic community to work together.

One More Thing

The release of the H01 series of data coincides with the 10th anniversary of the establishment of Google Research’s Connectomics team.

Previously, the team also released a Drosophila brain map containing 25,000 neurons and millions of connections between them.

Last year, the team also announced that it would cooperate with a number of universities and spend US$33 million to map the hippocampus in the mouse brain. This project is also the focus of the team's next step.

The H01 map released this time was first released as a data set and preprint paper in June 2021. After optimization and a deeper analysis of synaptic characteristics, the official version of the paper was unveiled today .

Paper address: https://www.science.org/doi/10.1126/science.adk4858

参考链接：
[1]https://research.google/blog/a-browsable-petascale-reconstruction-of-the-human-cortex/。
[2]https://www.sciencealert.com/amazingly-detailed-images-reveal-a-single-cubic-millimeter-of-human-brain-in-3d。
[3]https://news.ycombinator.com/item?id=40313193。

The above is the detailed content of The most detailed 3D map of the human brain is published in Science! GPT-4 parameters are only equivalent to 0.2% of humans. For more information, please follow other related articles on the PHP Chinese website!