If you can get past the coining of yet another new word to talk about neural network designs, think about this: modeling the actual synapse can be a way to improve digital systems. In other words, a lot of people are thinking about the cognitive process itself – but some others, oddly enough to some, are studying the physical tissue that is doing all of that thinking in the biological universe. Not only that – they’re peering in, drawing, analyzing, using microscopes and pencils – and coming up with some pretty exciting observations.

All of this is ultimately really interesting and important to modern research.

Here, MIT Professor Nir Shavit talks about “connectomics” and work pioneered by Santiago Ramon y Cahal (Cahal, along with Camillo Golgi, received the Nobel Prize in Physiology/Medicine in 1906) and how that impacts modern efforts to study neural tissue under a microscope.

Some of what the early scientists did, he explains, involved intuition, because they couldn’t actually see the synaptic connections.

“(More recently) we studied the conductivity in neural tissue, and tried to get … maps of these neural networks,” he says. “And … the idea here is that in order to see the synapses, we actually need to use electron microscopy.”

Describing a pipeline approach, Shavit reveals a pretty radical method:

“We take an animal, we take its brain, we embed the brain in a resin, then we take that resin, and we slice it 10s of 1000s of times … put it on a tape, take the tape… put it on disks, and stick these discs into an electron microscope. And with the electron microscope, now we can get the resolution that we need in order to actually see the connections. … if you have this image…now what we can do is actually take these images, (here’s a block of images from an electron microscope,) and … in each slice, we can identify where the neuron is (by) segmenting it, and from this, we can get the actual connectivity between these neurons.”

In displaying the cubic rendering and building a 3D model, he adds, the quality of the work depends on the quality of the image, and here’s where it really gets interesting. Shavit notes that it will take years to do the research the conventional way, but then suggests that you can just isolate the more important parts of that cube with a smart data read. He suggests spending “beam time” on just the synaptic connections.

Ruminating on 3D modeling

If you can follow this, you’re seeing that what you’re actually doing is using artificial intelligence to help fine-tune the study of biological neural networks.

So one type of neural network is allowing us to study another.

Video: Some insights into connectivity and more can be gleaned by looking more closely at neural net activity

It has a little bit of a similarity to the idea that all sorts of general-purpose AI are going to help us with specialized AIs, and vice versa.

Sometimes it gets sort of hard to keep all of the threads straight! Not to mention that in the mad rush toward AI, the idea of AIs working on other AIs is a major part of what gives a lot of people, frankly, the creeps…

Anyway, going back to this video, Shavit suggests these things can be in every lab, eventually.

He specifies that you want to focus on the membranes, and to some extent ignore all of the other interstitial tissue and data.

He presents this focus on ‘regions of interest’ that he says leads to hybrid accuracy and speed models with what he calls a “smart electron microscope” process. (The video shows more detail in predicting where difficulties will be, and how this method works.)

“Right now, we’re just in the beginning of building this technology, you can get about (700%) improvement, you know, and you can do that with almost no error added to it,” Shavit says. “And the future, of course, is to use this everywhere. Hopefully, you can make labs all over the world use this kind of technology. And we can use it not only for biology, but also for material science, for detecting errors in VLSI circuits, those kinds of things.”

In closing, Shavit gives props to the teams involved in this sort of ground-breaking work.

“This is joint work between Harvard, MIT, and the people at Thermo Fisher, which really bring their expertise in building microscopes to this problem,” he says.