Abstractions of the Mind Before Data Were So Abundant, Computer Models of the Brain Were Simple

OUTLOOK COGNITIVE HEALTH

NEURAL MODELLING Abstractions of the mind Before data were so abundant, computer models of the brain were simple. Information is now much more plentiful — but some argue that models should remain uncomplicated.

BY KELLY RAE CHI testing hypotheses on how the brain deals with amused Eliasmith, it did not surprise him. uncertainty in functions such as attention and Markram is well known for taking a different he first major results of the Blue Brain decision-making. approach to modelling, as he did in the Blue

Project, a detailed simulation of a bit of There is a widespread preference for Brain Project. His strategy is to build in every MARIO WAGNER rat neocortex about the size of a grain hypothesis-driven approaches in the brain- possible detail to derive a perfect imitation of Tof coarse sand, were published last year1. The modelling community. Some models might be the biological processes in the brain with the model represents 31,000 brain cells and 37 mil- very small and detailed, for example, focusing hope that higher functions will emerge — a lion synapses. It runs on a supercomputer and on a single synapse. Others might explore the ‘bottom-up’ approach. Researchers such as is based on data collected over 20 years. Fur- electrical spiking of whole neurons, the com- Eliasmith and Pouget take a ‘top-down’ strat- thermore, it behaves just like a speck of brain munication patterns between brain areas, or egy, creating simpler models based on our tissue. But therein, say critics, lies the problem. even attempt to recapitulate the whole brain. knowledge of behaviour. These skate over “It’s the best biophysical model we have of any But ultimately a model needs to answer ques- certain details, instead focusing on testing brain, but that’s not enough,” says Christof tions about brain function if we are to advance hypotheses about brain function. Koch, a neuroscientist at the Allen Institute for our understanding of cognition. Rather than dismiss the criticism, Eliasmith Brain Science in Seattle, Washington, which took Markram’s comment on board and added has embarked on its own large-scale brain- FROM TOP TO BOTTOM bottom-up detail to Spaun. He selected a modelling effort. The trouble with the model is Blue Brain is not the only sophisticated model handful of frontal cortex neurons, which were that it holds no surprises: no higher functions to have hit the headlines in recent years. In relatively simple to begin with, and swapped or unexpected features have emerged from it. late 2012, theoretical neuroscientist Chris them for much more complicated neurons — Some neuroscientists, including Koch, say Eliasmith at the University of Waterloo in ones that account for multiple ion channels that this is because the model was not built Canada unveiled Spaun, a whole-brain model and changes in electrical activity over time. with a particular hypothesis about cognitive that contains 2.5 million neurons (a fraction Although these complicated neurons were processes in mind. Its success will depend on of the human brain’s estimated 86 billion). more biologically realistic, Eliasmith found whether specific questions can be asked of Spaun has a digital eye and a robotic arm, and that they brought no improvement to Spaun’s it. The irony, says neuroscientist Alexandre can reason through eight complex tasks such performance on the original eight tasks. “A Pouget, is that deriving answers will require as memorizing and reciting lists, all of which good model doesn’t introduce complexity for drastic simplification of the model, “unless we involve multiple areas of the brain2. Neverthe- complexity’s sake,” he says. figure out how to adjust the billions of param- less, Henry Markram, a neurobiologist at the eters of the simulations, which would seem Swiss Federal Institute of Technology in Laus- SIMPLIFY, SIMPLIFY, SIMPLIFY to be a challenging problem to say the least”. anne who is leading the Blue Brain Project, For many years, computational models of the By contrast, Pouget’s group at the University noted3 at the time: “It is not a brain model.” brain were what theorists call unconstrained: of Geneva, Switzerland, is generating and Although Markram’s dismissal of Spaun there were not enough experimental data to

results. He found, for instance, that neurons In physics, the marriage between experiment oscillated when first presented with a pattern and theory led to the development of unifying that was moving at constant velocity — a result principles. And although neuroscientists might that he took to Werner Reichardt, who was hope for a similar revelation in their field, the also taken aback. “He didn’t expect his model brain (and biology in general) is inherently to show that,” says Borst. They confirmed the more noisy than a physical system, says com- results in real neurons, and continued to refine putational neuroscientist Gustavo Deco of the and expand Reichardt’s model to gain insight Pompeu Fabra University in Barcelona, Spain, into how the visual system detects motion. who is an investigator on the Human Brain Pro- In the realm of bottom-up models, the ject. Deco points out that equations describing greatest success has come from a set of equa- the behaviour of neurons and synapses are non- tions developed in 1952 to explain how flow linear, and neurons are connected in a variety of ions in and out of a nerve cell produces an of ways, interacting in both a feedforward and a axon potential. These Hodgkin–Huxley equa- feedback manner. That said, there are examples tions are “beautiful and inspirational”, says of theory allowing neuroscientists to extract neurobiologist Anthony Zador of Cold Spring general principles, such as how the brain bal- Harbor Laboratory in New York, adding that ances excitation and inhibition, and how neu- they have allowed many scientists to make rons fire in synchrony, Wang says. predictions about how neuronal excitability Complex neuroscience often requires works. The equations, or their variants, form huge computational resources. But it is not some of the basic building blocks of many of a want of supercomputers that limits good, today’s larger brain models of cognition. theory-driven models. “It is a lack of knowledge about experimental facts. We need more map onto the models or to fully test them. GAMBLE IN DETAILS facts and maybe more ideas,” Borst says. For instance, scientists could record electrical Although many theoretical neuroscientists Those who crave vast amounts of computer activity, but from only one neuron at a time, do not see value in pure bottom-up power misunderstand the real challenge which limited their ability to represent neural approaches such as that taken by the Blue facing scientists who are trying to unravel networks. Back then, brain models were simple Brain Project, they do not dismiss bottom-up the mysteries of the brain, Borst contends. out of necessity. models entirely. These types of data-driven “I still don’t see the need for simulating one In the past decade, an array of technologies brain simulations have the benefit of remind- million neurons simultaneously in order to has provided more information. Imaging tech- ing model-builders what they do not know, understand what the brain is doing,” he says, nology has revealed previously hidden parts of which can inspire new experiments. And referring to the large-scale simulation linked the brain. Researchers can control genes to iso- top-down approaches can often benefit from with the Human Brain Project. “I’m sure we late particular functions. And emerging statisti- the addition of more detail, says theoretical can reduce that to a handful of neurons and cal methods have helped to describe complex neuroscientist Peter Dayan of the Gatsby get some ideas.” phenomena in simpler terms. These techniques Computational Neuroscience Unit at Uni- Computational neuroscientist Andreas are feeding newer generations of models. versity College London. “The best kind of Herz, of the Ludwig-Maximilians University in Nevertheless, most theorists think that a modelling is going top-down and bottom-up Munich, Germany, agrees. “We make best pro- good model includes only the details needed simultaneously,” he says. gress if we focus on specific elements of neural to help answer a specific Borst, for example, is now approaching the computation,” he says. For example, a single question. Indeed, one “We make Reichardt detector from the bottom up to cortical neuron receives input from thousands of the most challeng- best progress explore questions such as how neurotrans- of other cells, but it is unclear how it processes ing aspects of model if we focus mitter receptors on motion-sensitive neurons this information. “Without this knowledge, building is working out on specific interact. And Eliasmith’s more complex Spaun attempts to simulate the whole brain in a seem- which details are impor- elements has allowed him to do other types of experi- ingly biologically realistic manner are doomed tant to include and of neural ment that he couldn’t before — in particular, he to fail,” he adds. which are acceptable to computation.” can now mimic the effect of sodium-channel At the same time, supercomputers do allow ignore. “The simpler the blockers on the brain. researchers to build details into their models model is, the easier it is to analyse and under- Also taking a multiscale approach is and see how they compare to the originals, as stand, manipulate and test,” says cognitive and neuroscientist Xiao-Jing Wang of New with Spaun. Eliasmith has used Spaun and its computational neuroscientist Anil Seth of the York University Shanghai in China, whose variations to see what happens when he kills University of Sussex in Chichester, UK. group described a large-scale model of the neurons or tweaks other features to investigate An oft-cited success in theoretical interaction of circuits across different regions ageing, motor control or stroke damage in the neuroscience is the Reichardt detector — a of the macaque brain4. The model is built, in brain. For him, adding complexity to a model simple, top-down model for how the brain part, from his previous, smaller models of local has to serve a purpose. “We need to build big- senses motion — proposed by German physi- neuronal circuits that show how neurons in a ger and bigger models in every direction, more cist Werner Reichardt in the 1950s. “The big group fire in time. To scale up to the entire neurons and more detail,” he says. “So that we advantage of the Reichardt model for motion brain, Wang had to include the strength of can break them.” ■ detection was that it was an algorithm to begin the feedback between areas. Only now has he with,” says neurobiologist Alexander Borst of got the right data — thanks to the burgeoning Kelly Rae Chi is a freelance science writer the Max Planck Institute of Neurobiology in field of connectomics (the study of connection based in Cary, North Carolina. Martinsried, Germany. “It doesn’t speak about maps within an organism’s nervous system) — 1. Markram, H. et al. Cell 163, 456–492 (2015). neurons at all.” to build in this important detail, he says. Wang 2. Eliasmith, C. et al. Science 338, 1202–1205 (2012). 3. Sanders, L. Science News http://go.nature.com/ When Borst joined the Max Planck Society is using his model to study decision-making, j1t8lu (2012). in the mid-1980s, he ran computational simula- the integration of sensory information and 4. Chaudhuri, R., Knoblauch, K., Gariel, M.-A., Kennedy, tions of the Reichardt model, and got surprising other cognitive processes. H. & Wang, X.-J. Neuron 88, 419–431 (2015).