August 8, 2013 | 5
Stacked layers have fascinated us humans for as long as we’ve sought to organize the universe around us. Authors and artists have developed primitive conceptions of heaven, earth and hell into elaborate hierarchies of celestial and infernal spheres. Medieval alchemists placed metals and planets in a symbolic hierarchy, with gold and the sun at the very top. Even today, we read the ancient history of our planet in its descending layers of geological strata; and our schools, governments and militaries – along with many other professions and organizations – are structured according to layered hierarchical models.
Appearances, however, can be deceiving. Since the early twentieth century, many neuroscience professors have taught that the brain is organized according to a strict hierarchical model, in which incoming signals are processed one layer at a time. But a new discovery may be poised to overturn a century of this thinking, and blaze a trail toward a very different model of brain function.
The outer surface of your cerebrum - that cauliflowery mass that makes up most of your brain – is covered with a thin coating of dense gray matter. This coating is known as the cerebral cortex, after the Latin word for “rind” or “bark.” Unlike tree bark, though, the cerebral cortex’s layers aren’t stacked neatly, one on top of the next – they’re interlaced in a multilayered quilt all across the outer surface of your brain. Still, a rough hierarchy does seem to persist all over the cortex: Layer 4, in the middle of the cortex, responds to simple stimuli, then passes its output on to Layers 2 and 3, which perform some further processing, and finally shuttle their output inward to Layers 5 and 6.
This roughly hierarchical pattern, it turns out, is repeated all throughout your cortex, from brain areas that recognize shapes and sounds all the way up to areas that recognize specific songs, feelings and even abstract ideas: Simple stimuli trigger reactions in Layer 4, which passes its responses on through the hierarchy for further integration and processing.
What’s more, many cortical neurons prefer to connect with other neurons in the same local area, creating what’s known as a “cortical column” – a densely packed tower of interconnected neurons dedicated to recognizing one specific type of stimulus. Many scientists have come to see these cortical columns – of which there are about two million in a human brain – as the basic building blocks of cortical computation. If we could simulate the precise behavior of a whole brain-wide network of cortical columns, some researchers say, we might be very close to engineering a genuine artificial intelligence.
But as today’s cutting-edge neuroscience labs take closer looks at the cortex’s circuitry, this column-centric hierarchical model isn’t entirely holding up. While it’s clear that some kind of hierarchical processing takes place in the cortex, it may not be column-centric – and the cortical layers may serve very different purposes than anyone expected.
Deep below your cortex – below your entire cerebrum, in fact – sits your thalamus, an ancient almond-shaped brain area that relays signals around the cortex. Brain-wiring experts have known since the early twentieth century that the thalamus sends signals mainly to Layer 4 of the cortex – the layer that appears to integrate simple stimuli into more complex signals for processing by the Layers 2 and 3, then by Layers 5 and 6.
For decades, this model of thalamocortical wiring seemed so simple as to be almost beyond doubt. In this view, Layer 4 receives input from the thalamus, and helps assemble those signals into patterns for Layers 2 and 3, then Layers 5 and 6, to recognize and interpret. “We’ve always had this idea that the thalamus’s role is to activate Layer 4 of the cortex,” says Randy Bruno, a neuroscientist at Columbia University. while Layers 2 and 3 represent more complex features, then in turn send their own signals on to more the complex Layers 5 and 6.
But this June, Bruno and his team published a research paper that calls the old hierarchy into serious question. “We measured the degree of connectivity between the thalamus and Layers 5 and 6 of the cortex,” Bruno says, “and that connectivity turns out to be much stronger than expected.”
So Bruno began to wonder: What if, instead of just sending information to Layer 4, to be passed up a layered cortical hierarchy, the thalamus feeds the same information to multiple cortical layers simultaneously? Maybe the cortical “hierarchy” isn’t so hierarchical after all. Maybe the outermost layers work more like a democracy, where several parallel chains of command all converge on the same decisions, then pass the results on to their higher-ups.
Intrigued by this lead, Bruno and his team designed a round of experiments to test just how strong a role the thalamus played in higher-layer processing. Using local anesthesia to “turn off” individual cortical layers in living animals, the investigators checked to see how long it takes each cortical layer to be activated by incoming sensory stimuli. What they found surprised them: Layer 5 responds at the same speed as Layer 4. This, as it turned out, was the first hint hint of something big.
Next, the team counted synaptic connections between thousands of individual neurons, and kept track of how many thalamic neurons were synaptically connected to neurons in Layer 5. That connectivity turned out to be much stronger than anyone had predicted. And finally, the researchers “turned off” Layer 4 altogether while recording the electrochemical activity of neurons in Layer 5. At last, they’d be able to confirm or deny what everyone had assumed for so long: That Layer 4 sat at the bottom of the cortical hierarchy, received most of its input straight from the thalamus, and sent signals on to the waiting Layers 2 and 3. If they were right, Layer 5 also received some of its input straight from the thalamus, and the hierarchy might be structured very differently from existing models.
“I expected the proportion of thalamic connectivity would be fifty-fifty, at best, between Layer 4 and Layer 5,” Bruno says; “but we were shocked to find that almost 100 percent of the synaptic input to neurons in Layers 5 and 6 comes directly from the thalamus.” This means that most of the input to Layers 5 and 6 isn’t actually passed up through Layer 4 and on through Layers 2 and 3, as most everyone had assumed for the past century. Instead, the thalamus seems to send two parallel copies of the same information to two different places: Layer 4 and Layer 5. “If we have that kind of arrangement,” Bruno says, “I can reach no other conclusion than that our cortex has two similar but distinct sensory processing systems living right on top of one another. It’s almost as if you have two brains built into one cortex.”
“These findings may pave the way for a new computational theory of how the cortex processes information,” says Rajesh Rao, a neuroscientist at the University of Washington, Seattle, and an expert on cortical wiring. “The notion of a strict cortical hierarchy, especially a feedforward hierarchy of processing layers, is losing ground to a more nuanced view of the cortex as an interconnected and distributed network.”
One area where Bruno’s discoveries are likely to have a major impact is in the study of receptive fields: The range of inputs to which a given neuron responds. Traditionally, most researchers have thought that neurons in each layer help assemble and shape the receptive fields of neurons in the next layer up – but the fact that the thalamus sends signals directly to Layers 4 and 5 seems to contradict that assumption. “The point of our paper,” Bruno says, “is that the connections between the layers are probably used for something other than constructing receptive fields. I think a lot of those connections are there for something much more interesting.”
So if lower layers aren’t constructing receptive fields, what are they up to? “One possibility,” Bruno says, “is that Layers 2, 3 and 4 are interested in context, while the lower-ranking layers – Layers 5 and 6 – create simple stimulus-response loops.” Another possibility is that Layers 2 and 3 are crucial for learning. Intriguing as this idea sounds, it’s still at the theoretical stage, not yet verified by hard data – but Bruno points out that Layers 2 and 3 are sparsely active, responding only to very select input configurations. Experiments with computerized neural networks have demonstrated that this kind of sparse coding is ideal for learning new information, which suggests that Layers 2 and 3 may be adapted for learning and predicting patterns in incoming signals from Layers 4, 5 and 6. “We’re investigating this idea in our lab right now,” Bruno says.
In other words, researchers may soon verify a more democratic model of cortical behavior, in which Layers 4, 5 and 6 all cross-check their own interpretations of incoming signals with each other in order to agree on a result. “Dr. Bruno’s results suggest these neurons throughout the brain have direct access to incoming sensory information, and can combine this information with predictions or contextual information from higher cortical areas,” Rao says.
Not only could these realizations rewrite the book on cortical layers – they may also alter our idea of what a cortical column is, and what it does. “Instead of viewing a module as a cortical column with a strict flow of information from the middle to upper layers and then to deeper layer neurons,” Rao says, “it may be more useful to regard it as a computational unit that integrates inputs from a diverse set of sources.” Lower-layer neurons may integrate their interpretations of sensory and contextual information, then convey this information to areas of the brain that deal with action selection and other behavioral responses – while higher-layer neurons may compare sensory information with predictions and other contextual information from higher-order areas, and convey the results back to the higher-order areas. Both these processes may be taking place simultaneously, column by column, all across the cortex.
Even if we need to make some revisions to our textbooks, though, Bruno’s findings don’t negate the fact that Layer 4 responds to simpler stimuli; higher layers to more complex inputs. Nor do they undermine the fact that the thalamus sends a large proportion of its signals to Layer 4. But Bruno’s results still put our traditional explanation for cortical function on trial. “Maybe we’ve fooled ourselves into thinking that because these two information-processing systems sit on top of one another and are interconnected, they’re all parts of one structure,” Bruno says. “That, to me, is the most intriguing question this research raises.”
Images: by Randy Bruno
Secrets of the Universe: Past, Present, FutureX