SAN DIEGO Harnessing the photoconductive properties of novel silicon "neurochips," researchers at the University of California, San Diego have been able to directly observe the physical changes that neurons undergo during learning.
By non-invasively firing specific neurons grown on a silicon substrate, the researchers were able to identify the exact physical changes, which were observed with a fluorescent tracer. Laboratory observations during learning regimes were able to verify the specific physical changes resulting in short- and long-term memories.
"I wouldn't call it the holy grail of all of neuroscience, but in terms of memory and learning in the brain, its one of the things that we have been searching for, for a very long time," said Michael Colicos, post-doctoral fellow to professor Yukiko Goda at the University of California, San Diego (UCSD). "For many years, most people have assumed that when we form new memories, new connections are made between our brain cells, and there has been some evidence of this. But no one has been able to actually film it happening dynamically and live, stimulating them in a culture environment like we've been able to do."
Also on the team were biochemistry professor Michael J. Sailor and post-doctoral fellow Boyce Collins.
The Goda-led research team designed a custom device using a smooth silicon surface identical to normal chips, but on which living brain cells were grown. Unlike all previous cell culture setups, wherein cells were treated roughly, the smooth silicon surface enabled the researchers to create an ideal environment for the growing brain cells. Previously, electrode puncture wounds limited cell lifetimes to five to six hours. But Goda's approach should be able to keep cells alive for as long as a year, although the current experiment was limited to six weeks.
Instead of using electrodes, the researchers took advantage of the "photoconductive" properties of silicon. The effect can be used to lower the resistance below any specific neuron merely by shining a narrow-beam light onto it. Thus, a high-frequency pulse delivered to the whole substrate will only fire the neuron on which the laser is focused. In effect,the light opens a gate between the neuron and the substrate.
"We refer to it as photoconductive stimulation, because the brain cells are grown on a thin layer of silicon. Then, by using a light shining on the surface of the chip, we actually change the conductivity of the silicon in the part that is right underneath the neuron.
"So you observe the chip under a microscope and observe the neuronal culture growing there," said Colicos. "And if you want to fire one specific neuron, you can target in on that neuron, shine a light on it and then give it a short pulse of electricity, and that will cause the neuron to fire."
When the neuron fires, according to Colicos, it fires in exactly the same manner as in the brain, even though it has been pushed over its threshold. Normally, neurons gather input pulses together according to a resistance-capacitance curve, where they integrate charge until they pass a preset threshold level. In the experiment, the applied electronic pulse floods in just enough charge to take the neuron over its threshold, although the neuron immediately resets and resumes normal operation.
Brain changes marked
To take the first step toward future neural circuitry, the research team tested out the theory that learning results from a physical change that strengthens the connections between selected neurons. Using a fluorescent marker, the researchers were able to show how short- and long-term memories result from different physical effects in the brain. Short-term memories, it turns out, result from the instant assembly of more filaments to strengthen the skin of the cell temporarily, whereas long-term memories result from the growing of a new synapses to strengthen the connection permanently.
"Since we have this fluorescent tag, we can actually watch what is happening inside the cell, both while we are firing it, to make the two neurons communicate with each other and in the long term, after we have fired it repeatedly in a certain pattern. We were finally able to see what the long-term changes were in terms of the structure and connectivity between neurons," said Colicos.
The neurons used were from the hippocampus, the area of the human brain thought to be crucial in forming memories. First the selected neurons were pulsed randomly, as if from the normal daily sensory input of uneventful experiences short-term memory with the result that more filaments instantly appeared inside the axon "output wire" to adjacent neurons. The team filmed the filaments inside the temporarily activated cell and noted that they moved toward the neurons to which the activated cell was connected. Filaments in neighboring neurons moved away from the activated cell. Both of those cell changes were temporary, lasting only about three to five minutes before disappearing permanently within 10 minutes.
Long-term memories, from repeated pulse patterns, resulted in new synapses being grown at the end of axons in less than an hour. The synapses actually split at their ends and form new synapses. At the end of trials as long as six weeks, the researchers observed no changes in these newly formed synaptic connections, indicating that they were permanent.
"If we repeated the same pattern of high-frequency pulses for as little as four times in an hour, we found that at the end of that the axon had grown a new synaptic connection," said Colicos.
For the future, according to Colicos, experiments will attempt to verify the flip side of the "strengthening" (long-term potentiation) theory that the current research verified. That is, the team will investigate the "decay" (long-term depression) part of the synaptic connection.
Synapses 'lose' old info
If learning strengthens connections, then the slow decay of memories should theoretically weaken old connections. To prove how connections are strengthened, the researchers used a fast series of high-frequency pulses, so their research into how the brain forgets will use a slow series of low-frequency pulses.
"We think that slow, low-frequency pulses over a long period of time might produce synaptic loss, or forgetting. You just have this slow repetitious background pattern that slowly negates old things out of the system," said Colicos.
Once the team knows how to add and subtract connections wherever they want to put them, Colicos envisions designing computing circuitry that performs analog computations. For instance, the vision and associative memory capabilities of real neural networks routinely performs massively parallel operations that are impossible even for the biggest supercomputers. Future neural designers could wire sensor subsystems with integral object segmentation and recognition using living neurons to outperform the fastest DSPs.
"Neural networks can perform very complex integrations and associations of information. Since we are non-invasively firing the neurons and they can survive for a very long time, you could potentially feed information into them and then, by observing how they processed it, read data back out of it," said Colicos.
Future cultured-brain-cell experiments may design and troubleshoot neural circuitry in the same manner that a gate array might be designed today. Like a gate array, such a neural chip would hold an array of neurons that could be induced to grow into a particular kind of data-processing configuration by stimulating its neurons in a pattern that results in its growing the particular connections that realize that data-processing task.
An audio recording of reporter R. Colin Johnson's full interview with Michael Colicos can be found online at AmpCast.com/RColinJohnson.