Category: Uncategorized

Mark Harnett named Vallee Foundation Scholar

Anne Trafton |

The Bert L and N Kuggie Vallee Foundation has named McGovern Institute investigator Mark Harnett a 2018 Vallee Scholar. The Vallee Scholars Program recognizes original, innovative, and pioneering work by early career scientists at a critical juncture in their careers and provides $300,000 in discretionary funds to be spent over four years for basic biomedical research. Harnett is among five researchers named to this year’s Vallee Scholars Program.

Harnett, who is also the Fred and Carole Middleton Career Development Assistant Professor in the Department of Brain and Cognitive Sciences, is being recognized for his work exploring how the biophysical features of neurons give rise to the computational power of the brain. By exploiting new technologies and approaches at the interface of biophysics and systems neuroscience, research in the Harnett lab aims to provide a new understanding of the biology underlying how mammalian brains learn. This may open new areas of research into brain disorders characterized by atypical learning and memory (such as dementia and schizophrenia) and may also have important implications for designing new, brain-inspired artificial neural networks.

The Vallee Foundation was established in 1996 by Bert and Kuggie Vallee to foster originality, creativity, and leadership within biomedical scientific research and medical education. The foundation’s goal to fund originality, innovation, and pioneering work “recognizes the future promise of these scientists who are dedicated to understanding fundamental biological processes.” Harnett joins a list of 24 Vallee Scholars, including McGovern investigator Feng Zhang, who have been appointed to the program since its inception in 2013.

Feng Zhang named winner of the 2018 Keio Medical Science Prize

Anne Trafton |

Feng Zhang and Masashi Yanagisawa have been named the 2018 winners of the prestigious Keio Medical Science Prize. Zhang is being recognized for the groundbreaking development of CRISPR-Cas9-mediated genome engineering in cells and its application for medical science. Zhang is an HHMI Investigator and the James and Patricia Poitras Professor of Neuroscience at MIT, an associate professor in MIT’s Departments of Brain and Cognitive Sciences and Biological Engineering, an investigator at the McGovern Institute for Brain Research, and a core member of the Broad Institute of MIT and Harvard. Masashi Yanagisawa, Director of the International Institute for Integrative Sleep Medicine at the University of Tsukuba, is being recognized for his seminal work on sleep control mechanisms.

“We are delighted that Feng is now a Keio Prize laureate,” says McGovern Institute Director Robert Desimone. “This truly recognizes the remarkable achievements that he has made at such a young age.”

The Keio Medical Prize is awarded to a maximum of two scientists each year, and is now in its 23rd year. The prize is offered by Keio University, and the selection committee specifically looks for laureates that have made an outstanding contribution to medicine or the life sciences. The prize was initially endowed by Dr. Mitsunada Sakaguchi in 1994, with the express condition that it be used to commend outstanding science, promote medical advances in medicine and the life sciences, expand researcher networks, and contribute to the well-being of humankind. The winners receive a certificate of merit, medal, and a monetary award of 10 million yen.

Feng Zhang is a molecular biologist who has contributed to the development of multiple molecular tools to accelerate our understanding of human disease and create new therapeutic modalities. During his graduate work Zhang contributed to the development of optogenetics, a system for activating neurons using light, which has advanced our understanding of brain connectivity. Zhang went on to pioneer the deployment of the microbial CRISPR-Cas9 system for genome engineering in eukaryotic cells. The ease and specificity of the system has led to its widespread use across the life sciences and it has groundbreaking implications for disease therapeutics, biotechnology, and agriculture. Zhang has continued to mine bacterial CRISPR systems for additional enzymes with useful properties, leading to the discovery of Cas13, which targets RNA, rather than DNA, and may potentially be a way to treat genetic diseases without altering the genome. He has also developed a molecular detection system called SHERLOCK based on the Cas13 family, which can sense trace amounts of genetic material, including viruses and alterations in genes that might be linked to cancer.

“I am tremendously honored to have our work recognized by the Keio Medical Prize,” says Zhang. “It is an inspiration to us to continue our work to improve human health.”

The prize ceremony will be held on December 18th 2018 at Keio University in Tokyo, Japan.

New sensors track dopamine in the brain for more than a year

Anne Trafton |

Dopamine, a signaling molecule used throughout the brain, plays a major role in regulating our mood, as well as controlling movement. Many disorders, including Parkinson’s disease, depression, and schizophrenia, are linked to dopamine deficiencies.

MIT neuroscientists have now devised a way to measure dopamine in the brain for more than a year, which they believe will help them to learn much more about its role in both healthy and diseased brains.

“Despite all that is known about dopamine as a crucial signaling molecule in the brain, implicated in neurologic and neuropsychiatric conditions as well as our ability to learn, it has been impossible to monitor changes in the online release of dopamine over time periods long enough to relate these to clinical conditions,” says Ann Graybiel, an MIT Institute Professor, a member of MIT’s McGovern Institute for Brain Research, and one of the senior authors of the study.

Michael Cima, the David H. Koch Professor of Engineering in the Department of Materials Science and Engineering and a member of MIT’s Koch Institute for Integrative Cancer Research, and Rober Langer, the David H. Koch Institute Professor and a member of the Koch Institute, are also senior authors of the study. MIT postdoc Helen Schwerdt is the lead author of the paper, which appears in the Sept. 12 issue of Communications Biology.

Long-term sensing

Dopamine is one of many neurotransmitters that neurons in the brain use to communicate with each other. Traditional systems for measuring dopamine — carbon electrodes with a shaft diameter of about 100 microns — can only be used reliably for about a day because they produce scar tissue that interferes with the electrodes’ ability to interact with dopamine.

In 2015, the MIT team demonstrated that tiny microfabricated sensors could be used to measure dopamine levels in a part of the brain called the striatum, which contains dopamine-producing cells that are critical for habit formation and reward-reinforced learning.

Because these probes are so small (about 10 microns in diameter), the researchers could implant up to 16 of them to measure dopamine levels in different parts of the striatum. In the new study, the researchers wanted to test whether they could use these sensors for long-term dopamine tracking.

“Our fundamental goal from the very beginning was to make the sensors work over a long period of time and produce accurate readings from day to day,” Schwerdt says. “This is necessary if you want to understand how these signals mediate specific diseases or conditions.”

To develop a sensor that can be accurate over long periods of time, the researchers had to make sure that it would not provoke an immune reaction, to avoid the scar tissue that interferes with the accuracy of the readings.

The MIT team found that their tiny sensors were nearly invisible to the immune system, even over extended periods of time. After the sensors were implanted, populations of microglia (immune cells that respond to short-term damage), and astrocytes, which respond over longer periods, were the same as those in brain tissue that did not have the probes inserted.

In this study, the researchers implanted three to five sensors per animal, about 5 millimeters deep, in the striatum. They took readings every few weeks, after stimulating dopamine release from the brainstem, which travels to the striatum. They found that the measurements remained consistent for up to 393 days.

“This is the first time that anyone’s shown that these sensors work for more than a few months. That gives us a lot of confidence that these kinds of sensors might be feasible for human use someday,” Schwerdt says.

Paul Glimcher, a professor of physiology and neuroscience at New York University, says the new sensors should enable more researchers to perform long-term studies of dopamine, which is essential for studying phenomena such as learning, which occurs over long time periods.

“This is a really solid engineering accomplishment that moves the field forward,” says Glimcher, who was not involved in the research. “This dramatically improves the technology in a way that makes it accessible to a lot of labs.”

Monitoring Parkinson’s

If developed for use in humans, these sensors could be useful for monitoring Parkinson’s patients who receive deep brain stimulation, the researchers say. This treatment involves implanting an electrode that delivers electrical impulses to a structure deep within the brain. Using a sensor to monitor dopamine levels could help doctors deliver the stimulation more selectively, only when it is needed.

The researchers are now looking into adapting the sensors to measure other neurotransmitters in the brain, and to measure electrical signals, which can also be disrupted in Parkinson’s and other diseases.

“Understanding those relationships between chemical and electrical activity will be really important to understanding all of the issues that you see in Parkinson’s,” Schwerdt says.

The research was funded by the National Institute of Biomedical Imaging and Bioengineering, the National Institute of Neurological Disorders and Stroke, the Army Research Office, the Saks Kavanaugh Foundation, the Nancy Lurie Marks Family Foundation, and Dr. Tenley Albright.

Can the brain recover after paralysis?

by The Brain |

Why is it that motor skills can be gained after paralysis but vision cannot recover in similar ways? – Ajay, Puppala

Thank you so much for this very important question, Ajay. To answer, I asked two local experts in the field, Pawan Sinha who runs the vision research lab at MIT, and Xavier Guell, a postdoc in John Gabrieli’s lab at the McGovern Institute who also works in the ataxia unit at Massachusetts General Hospital.

“Simply stated, the prospects of improvement, whether in movement or in vision, depend on the cause of the impairment,” explains Sinha. “Often, the cause of paralysis is stroke, a reduction in blood supply to a localized part of the brain, resulting in tissue damage. Fortunately, the brain has some ability to rewire itself, allowing regions near the damaged one to take on some of the lost functionality. This rewiring manifests itself as improvements in movement abilities after an initial period of paralysis. However, if the paralysis is due to spinal-cord transection (as was the case following Christopher Reeve’s tragic injury in 1995), then prospects for improvement are diminished.”

“Turning to the domain of sight,” continues Sinha, “stroke can indeed cause vision loss. As with movement control, these losses can dissipate over time as the cortex reorganizes via rewiring. However, if the blindness is due to optic nerve transection, then the condition is likely to be permanent. It is also worth noting that many cases of blindness are due to problems in the eye itself. These include corneal opacities, cataracts and retinal damage. Some of these conditions (corneal opacities and cataracts) are eminently treatable while others (typically those associated with the retina and optic nerve) still pose challenges to medical science.”

You might be wondering what makes lesions in the eye and spinal cord hard to overcome. Some systems (the blood, skin, and intestine are good examples) contain a continuously active stem cell population in adults. These cells can divide and replenish lost cells in damaged regions. While “adult-born” neurons can arise, elements of a degenerating or damaged retina, optic nerve, or spinal cord cannot be replaced as easily lost skin cells can. There is currently a very active effort in the stem cell community to understand how we might be able to replace neurons in all cases of neuronal degeneration and injury using stem cell technologies. To further explore lesions that specifically affect the brain, and how these might lead to a different outcome in the two systems, I turned to Xavier Guell.

“It might be true that visual deficits in the population are less likely to recover when compared to motor deficits in the population. However, the scientific literature seems to indicate that our body has a similar capacity to recover from both motor and visual injuries,” explains Guell. “The reason for this apparent contradiction is that visual lesions are usually not in the cerebral cortex (but instead in other places such as the retina or the lens), while motor lesions in the cerebral cortex are more common. In fact, a large proportion of people who suffer a stroke will have damage in the motor aspects of the cerebral cortex, but no damage in the visual aspects of the cerebral cortex. Crucially, recovery of neurological functions is usually seen when lesions are in the cerebral cortex or in other parts of the cerebrum or cerebellum. In this way, while our body has a similar capacity to recover from both motor and visual injuries, motor injuries are more frequently located in the parts of our body that have a better capacity to regain function (specifically, the cerebral cortex).”

In short, some cells cannot be replaced in either system, but stem cell research provides hope there. That said, there is remarkable plasticity in the brain, so when the lesion is located there, we can see recovery with training.

Do you have a question for The Brain? Ask it here.

Why do I talk with my hands?

by The Brain |

This is a very interesting question sent to us by Gabriel Castellanos (thank you!) Many of us gesture with our hands when we speak (and even when we do not) as a form of non-verbal communication. How hand gestures are coordinated with speech remains unclear. In part, it is difficult to monitor natural hand gestures in fMRI-based brain imaging studies as you have to stay still.

“Performing hand movements when stuck in the bore of a scanner is really tough beyond simple signing and keypresses,” explains McGovern Principal Research Scientist Satrajit Ghosh. “Thus ecological experiments of co-speech with motor gestures have not been carried out in the context of a magnetic resonance scanner, and therefore little is known about language and motor integration within this context.”

There have been studies that use proxies such as co-verbal pushing of buttons, and also studies using other imaging techniques, such as electroencephalography (EEG) and magnetoencephalography (MEG), to monitor brain activity during gesturing, but it would be difficult to precisely spatially localize the regions involved in natural co-speech hand gesticulation using such approaches. Another possible avenue for addressing this question would be to look at patients with conditions that might implicate particular brain regions in coordinating hand gestures, but such approaches have not really pinpointed a pathway for coordinating speech and hand movements.

That said, co-speech hand gesturing plays an important role in communication. “More generally co-speech hand gestures are seen as a mechanism for emphasis and disambiguation of the semantics of a sentence, in addition to prosody and facial queues,” says Ghosh. “In fact, one may consider the act of speaking as one large orchestral score involving vocal tract movement, respiration, voicing, facial expression, hand gestures, and even whole body postures acting as different instruments coordinated dynamically by the brain. Based on our current understanding of language production, co-speech or gestural events would likely be planned at a higher level than articulation and therefore would likely activate inferior frontal gyrus, SMA, and others.”

How this orchestra is coordinated and conducted thus remains to be unraveled, but certainly the question is one that gets to the heart of human social interactions.

Do you have a question for The Brain? Ask it here.

Constructing the striatum

by Sabbi Lall |

The striatum, the largest nucleus of the basal ganglia in the vertebrate brain, was historically thought to be a homogeneous group of cells. This view was overturned in a classic series of papers from MIT Institute Professor, Ann Graybiel. In previous work, Graybiel, who is also an investigator at MIT’s McGovern Institute, found that the striatum is highly organized, both structurally and functionally and in terms of connectivity. Graybiel has now collaborated with Z. Josh Huang’s lab at Cold Spring Harbor Laboratory to map the developmental lineage of cells that give rise to this complex architecture. The authors found that different functions of the striatum, such as execution of actions as opposed to evaluation of outcomes, are defined early on as part of the blueprint that constructs this brain region, rather than sculpted through a later mechanism.

Graybiel and colleagues tracked what is happening early in development by driving cell-specific fluorescent markers that allowed them to follow the progenitors that give rise to cells in the striatum. The striatum is known, thanks to Graybiel’s early work, to be organized into compartments called striosomes and the matrix. These have distinct connections to other brain regions. Broadly speaking, while striosomes are linked to value-based decision-making and reinforcement-based behaviors, the matrix has been linked to action execution. These regions are further subdivided into direct and indirect pathways. The direct pathway neurons are involved in releasing inhibition in other regions of the basal ganglia and thus actively promote action. Neurons projecting into the indirect pathway, instead inhibit “unwanted” actions that are not part of the current “cortical plan.” Based on their tracking, Graybiel and colleagues were indeed able to build a “fate map” that told them when the cells that build these different regions of the striatum commit to a functional path during development.

“It was already well known that individual neurons have lineages that can be traced back to early development, and many such lineages are now being traced,” says Graybiel. “What is so striking in what we have found with the Huang lab is that the earliest specification of lineages we find—at least with the markers that we have used—corresponds to what later become the two major neurochemically distinct compartments of the striatum, rather than many other divisions that might have been specified first. If this is so, then the fundamental developmental ground plan of the striatum is expressed later by these two distinct compartments of the striatum.”

Building the striatum turns out to be a symphony of organization embedded in lateral ganglion eminence cells, the source of cells during development that will end up in the striatum. Progenitors made early in development are somewhat committed: they can only generate spiny projection neurons (SPNs) that are striosomal. Following this in time, cells that will give rise to matrix SPNs appear. There is then a second mechanism laid over this initial ground plan that is switched on in both striosomal and matrisomal neurons and independently gives rise to neurons that will connect into direct as opposed to indirect pathways. This latter specification of direct-indirect pathway neurons is less rigid, but there is an overarching tendency for neurons expressing a certain neurotransmitter, dopamine, to appear earlier in developmental time. In short, progenitors move through an orchestrated process where they generate spiny projection neurons that can first sit in any area of the striatum, then where the ultimate fate of cells is more restricted at the level of striosome or matrix, and finally choices are made in both regions regarding indirect-direct pathway circuitry. Remarkably, these results suggest that even at the very earliest development of the striatum, its ultimate organization is already laid down in a way that distinguishes value-related circuit from movement-related circuits.

“What is thrilling,” says Graybiel, “is that there are lineage progressions— the step by step laying out of the brain’s organization— the turn out to match the striosome-matrix architecture of the striatum the were not even known to exist 40 years ago!”

The striatum is a hub regulating movement, emotion, motivation, evaluation, and learning, and linked to disorders such as Parkinson’s Disease and persistent negative valuations. This means that understanding its construction has important implications, perhaps even, one day, for rebuilding a striatum affected by neurodegeneration. That said, the findings have broader implications. Consider the worm, specifically, C. elegans. The complete lineage of cells that make up this organism is known, including where each neuron comes from, what it connects to, and its function and phenotype. There’s a clear relationship between lineage and function in this relatively simple organism with its highly stereotyped nervous system. Graybiel’s work suggests that in the big picture, early development in the forebrain is also providing a game plan. In this case, however, this groundwork underpins for circuits that underlie extremely complex behaviors, those that come to support the volitional and habitual behaviors that make up part of who we are as individuals.

 

A social side to face recognition by infants

Anne Trafton |

When interacting with an infant you have likely noticed that the human face holds a special draw from a very young age. But how does this relate to face recognition by adults, which is known to map to specific cortical regions? Rebecca Saxe, Associate Investigator at MIT’s McGovern Institute and John W. Jarve (1978) Professor in Brain and Cognitive Sciences, and her team have now considered two emerging theories regarding early face recognition, and come up with a third proposition, arguing that when a baby looks at a face, the response is also social, and that the resulting contingent interactions are key to subsequent development of organized face recognition areas in the brain.

By a certain age you are highly skilled at recognizing and responding to faces, and this correlates with activation of a number of face-selective regions of the cortex. This is incredibly important to reading the identities and intentions of other people, and selective categorical representation of faces in cortical areas is a feature shared by our primate cousins. While brain imaging tells us where face-responsive regions are in the adult cortex, how and when they emerge remains unclear.

In 2017, functional magnetic resonance imaging (fMRI) studies of human and macaque infants provided the first glimpse of how the youngest brains respond to faces. The scans showed that in 4-6 month human infants and equivalently aged macaques, regions known to be face-responsive in the adult brain are activated when shown movies of faces, but not in a selective fashion. Essentially fMRI argues that these specific, cortical regions are activated by faces, but a chair will do just as well. Upon further experience of faces over time, the specific cortical regions in macaques became face-selective, no longer responding to other objects.

There are two prevailing ideas in the field of how face preference, and eventually selectivity, arise through experience. These ideas are now considered in turn by Saxe and her team in an opinion piece in the September issue of Trends in Cognitive Sciences, and then a third, new theory proposed. The first idea centers on the way we dote over babies, centering our own faces right in their field of vision. The idea is that such frequent exposures to low level face features (curvilinear shape etc.) will eventually lead to co-activation of neurons that are responsive to all of the different aspects of facial features. If these neurons stimulated by different features are co-activated, and there’s a brain region where these neurons are also found together, this area with be stimulated eventually reinforcing emergence of a face category-specific area.

A second idea is that babies already have an innate “face template,” just as a duckling or chick already knows to follow its mother after hatching. So far there is little evidence for the second proposition, and the first fails to explain why babies seek out a face, rather than passively look upon and eventually “learn” the overlapping features that represent “face.”

Saxe, along with postdoc Lindsey Powell and graduate student Heather Kosakowski, instead now argue that the role a face plays in positive social interactions comes to drive organization of face-selective cortical regions. Taking the next step, the researchers propose that a prime suspect for linking social interactions to the development of face-selective areas is the medial prefrontal cortex (mPFC), a region linked to social cognition and behavior.

“I was asked to give a talk at a conference, and I wanted to talk about both the development of cortical face areas and the social role of the medial prefrontal cortex in young infants,” says Saxe. “I was puzzling over whether these two ideas were related, when I suddenly saw that they could be very fundamentally related.”

The authors argue that this relationship is supported by existing data that has shown that babies prefer dynamic faces and are more interested in faces that engage in a back and forth interaction. Regions of the mPFC are also known to activated during social interactions and known to be activated during exposure to dynamic faces in infants.

Powell is now using functional near infrared spectroscopy (fNIRS), a brain imaging technique that measures changes in blood flow to the brain, to test this hypothesis in infants. “This will allow us to see whether mPFC responses to social cues are linked to the development of face-responsive areas.”

In Daniel Deronda, the novel by George Eliot, the protagonist says “I think my life began with waking up and loving my mother’s face: it was so near to me, and her arms were round me, and she sang to me.” Perhaps this type of positively valenced social interaction, reinforced by the mPFC, is exactly what leads to the particular importance of faces and their selective categorical representation in the human brain. Further testing of the hypothesis proposed by Powell, Kosakowski, and Saxe will tell.

Neuroscientists get at the roots of pessimism

Anne Trafton |

Many patients with neuropsychiatric disorders such as anxiety or depression experience negative moods that lead them to focus on the possible downside of a given situation more than the potential benefit.

MIT neuroscientists have now pinpointed a brain region that can generate this type of pessimistic mood. In tests in animals, they showed that stimulating this region, known as the caudate nucleus, induced animals to make more negative decisions: They gave far more weight to the anticipated drawback of a situation than its benefit, compared to when the region was not stimulated. This pessimistic decision-making could continue through the day after the original stimulation.

The findings could help scientists better understand how some of the crippling effects of depression and anxiety arise, and guide them in developing new treatments.

“We feel we were seeing a proxy for anxiety, or depression, or some mix of the two,” says Ann Graybiel, an MIT Institute Professor, a member of MIT’s McGovern Institute for Brain Research, and the senior author of the study, which appears in the Aug. 9 issue of Neuron. “These psychiatric problems are still so very difficult to treat for many individuals suffering from them.”

The paper’s lead authors are McGovern Institute research affiliates Ken-ichi Amemori and Satoko Amemori, who perfected the tasks and have been studying emotion and how it is controlled by the brain. McGovern Institute researcher Daniel Gibson, an expert in data analysis, is also an author of the paper.

Emotional decisions

Graybiel’s laboratory has previously identified a neural circuit that underlies a specific kind of decision-making known as approach-avoidance conflict. These types of decisions, which require weighing options with both positive and negative elements, tend to provoke a great deal of anxiety. Her lab has also shown that chronic stress dramatically affects this kind of decision-making: More stress usually leads animals to choose high-risk, high-payoff options.

In the new study, the researchers wanted to see if they could reproduce an effect that is often seen in people with depression, anxiety, or obsessive-compulsive disorder. These patients tend to engage in ritualistic behaviors designed to combat negative thoughts, and to place more weight on the potential negative outcome of a given situation. This kind of negative thinking, the researchers suspected, could influence approach-avoidance decision-making.

To test this hypothesis, the researchers stimulated the caudate nucleus, a brain region linked to emotional decision-making, with a small electrical current as animals were offered a reward (juice) paired with an unpleasant stimulus (a puff of air to the face). In each trial, the ratio of reward to aversive stimuli was different, and the animals could choose whether to accept or not.

This kind of decision-making requires cost-benefit analysis. If the reward is high enough to balance out the puff of air, the animals will choose to accept it, but when that ratio is too low, they reject it. When the researchers stimulated the caudate nucleus, the cost-benefit calculation became skewed, and the animals began to avoid combinations that they previously would have accepted. This continued even after the stimulation ended, and could also be seen the following day, after which point it gradually disappeared.

This result suggests that the animals began to devalue the reward that they previously wanted, and focused more on the cost of the aversive stimulus. “This state we’ve mimicked has an overestimation of cost relative to benefit,” Graybiel says.

The study provides valuable insight into the role of the basal ganglia (a region that includes the caudate nucleus) in this type of decision-making, says Scott Grafton, a professor of neuroscience at the University of California at Santa Barbara, who was not involved in the research.

“We know that the frontal cortex and the basal ganglia are involved, but the relative contributions of the basal ganglia have not been well understood,” Grafton says. “This is a nice paper because it puts some of the decision-making process in the basal ganglia as well.”

A delicate balance

The researchers also found that brainwave activity in the caudate nucleus was altered when decision-making patterns changed. This change, discovered by Amemori, is in the beta frequency and might serve as a biomarker to monitor whether animals or patients respond to drug treatment, Graybiel says.

Graybiel is now working with psychiatrists at McLean Hospital to study patients who suffer from depression and anxiety, to see if their brains show abnormal activity in the neocortex and caudate nucleus during approach-avoidance decision-making. Magnetic resonance imaging (MRI) studies have shown abnormal activity in two regions of the medial prefrontal cortex that connect with the caudate nucleus.

The caudate nucleus has within it regions that are connected with the limbic system, which regulates mood, and it sends input to motor areas of the brain as well as dopamine-producing regions. Graybiel and Amemori believe that the abnormal activity seen in the caudate nucleus in this study could be somehow disrupting dopamine activity.

“There must be many circuits involved,” she says. “But apparently we are so delicately balanced that just throwing the system off a little bit can rapidly change behavior.”

The research was funded by the National Institutes of Health, the CHDI Foundation, the U.S. Office of Naval Research, the U.S. Army Research Office, MEXT KAKENHI, the Simons Center for the Social Brain, the Naito Foundation, the Uehara Memorial Foundation, Robert Buxton, Amy Sommer, and Judy Goldberg.

Testing the limits of artificial visual recognition systems

Anne Trafton |

While it can sometimes seem hard to see the forest from the trees, pat yourself on the back: as a human you are actually pretty good at object recognition. A major goal for artificial visual recognition systems is to be able to distinguish objects in the way that humans do. If you see a tree or a bush from almost any angle, in any degree of shading (or even rendered in pastels and pixels in a Monet), you would recognize it as a tree or a bush. However, such recognition has traditionally been a challenge for artificial visual recognition systems. Researchers at MIT’s McGovern Institute for Brain Research and Department of Brain and Cognitive Sciences (BCS) have now directly examined and shown that artificial object recognition is quickly becoming more primate-like, but still lags behind when scrutinized at higher resolution.

In recent years, dramatic advances in “deep learning” have produced artificial neural network models that appear remarkably similar to aspects of primate brains. James DiCarlo, Peter de Florez Professor and Department Head of BCS, set out to determine and carefully quantify how well the current leading artificial visual recognition systems match humans and other higher primates when it comes to image categorization. In recent years, dramatic advances in “deep learning” have produced artificial neural network models that appear remarkably similar to aspects of primate brains, so DiCarlo and his team put these latest models through their paces.

Rishi Rajalingham, a graduate student in DiCarlo’s lab conducted the study as part of his thesis work at the McGovern Institute. As Rajalingham puts it “one might imagine that artificial vision systems should behave like humans in order to seamlessly be integrated into human society, so this tests to what extent that is true.”

The team focused on testing so-called “deep, convolutional neural networks” (DCNNs), and specifically those that had trained on ImageNet, a collection of large-scale category-labeled image sets that have recently been used as a library to train neural networks (called DCNNIC models). These specific models have thus essentially been trained in an intense image recognition bootcamp. The models were then pitted against monkeys and humans and asked to differentiate objects in synthetically constructed images. These synthetic images put the object being categorized in unusual backgrounds and orientations. The resulting images (such as the floating camel shown above) evened the playing field for the machine models (humans would ordinarily have a leg up on image categorization based on assessing context, so this was specifically removed as a confounder to allow a pure comparison of specific object categorization).

DiCarlo and his team found that humans, monkeys and DCNNIC models all appeared to perform similarly, when examined at a relatively coarse level. Essentially, each group was shown 100 images of 24 different objects. When you averaged how they did across 100 photos of a given object, they could distinguish, for example, camels pretty well overall. The researchers then zoomed in and examined the behavioral data at a much finer resolution (i.e. for each single photo of a camel), thus deriving more detailed “behavioral fingerprints” of primates and machines. These detailed analyses of how they did for each individual image revealed strong differences: monkeys still behaved very consistently like their human primate cousins, but the artificial neural networks could no longer keep up.

“I thought it was quite surprising that monkeys and humans are remarkably similar in their recognition behaviors, especially given that these objects (e.g. trucks, tanks, camels, etc.) don’t “mean” anything to monkeys” says Rajalingham. “It’s indicative of how closely related these two species are, at least in terms of these visual abilities.”

DiCarlo’s team gave the neural networks remedial homework to see if they could catch up upon extra-curricular training by now training the models on images that more closely resembled the synthetic images used in their study. Even with this extra training (which the humans and monkeys did not receive), they could not match a primate’s ability to discern what was in each individual image.

DiCarlo conveys that this is a glass half-empty and half-full story. Says DiCarlo, “The half full part is that, today’s deep artificial neural networks that have been developed based on just some aspects of brain function are far better and far more human-like in their object recognition behavior than artificial systems just a few years ago,” explains DiCarlo. “However, careful and systematic behavioral testing reveals that even for visual object recognition, the brain’s neural network still has some tricks up its sleeve that these artificial neural networks do not yet have.”

Dicarlo’s study begins to define more precisely when it is that the leading artificial neural networks start to “trip up”, and highlights a fundamental aspect of their architecture that struggles with categorization of single images. This flaw seems to be unaddressable through further brute force training. The work also provides an unprecedented and rich dataset of human (1476 anonymous humans to be exact) and primate behavior that will help act as a quantitative benchmark for improvement of artificial neural networks.

 

Image: Example of synthetic image used in the study. For category ‘camel’, 100 distinct, synthetic camel images were shown to DCNNIC models, humans and rhesus monkeys. 24 different categories were tested altogether.

Charting the cerebellum

Anne Trafton |

Small and tucked away under the cerebral hemispheres toward the back of the brain, the human cerebellum is still immediately obvious due to its distinct structure. From Galen’s second century anatomical description to Cajal’s systematic analysis of its projections, the cerebellum has long drawn the eyes of researchers studying the brain.  Two parallel studies from MIT’s McGovern institute have recently converged to support an unexpectedly complex level of non-motor cerebellar organization, that would not have been predicted from known motor representation regions.

Historically the cerebellum has primarily been considered to impact motor control and coordination. Think of this view as the cerebellum being the chain on a bicycle, registering what is happening up front in the cortex, and relaying the information so that the back wheel moves at a coordinated pace. This simple view has been questioned as cerebellar circuits have been traced to the basal ganglia and to neocortical regions via the thalamus. This new view suggests the cerebellum is a hub in a complex network, with potentially higher and non-motor functions including cognition and reward-based learning.

A collaboration between the labs of John Gabrieli, Investigator at the McGovern Institute for Brain Research and Jeremy Schmahmann, of the Ataxia Unit at Massachusetts General Hospital and Harvard Medical School, has now used functional brain imaging to give new insight into the cerebellar organization of non-motor roles, including working memory, language, and, social and emotional processing. In a complementary paper, a collaboration between Sheeba Anteraper of MIT’s Martinos Imaging Center and Gagan Joshi of the Alan and Lorraine Bressler Clinical and Research Program at Massachusetts General Hospital, has found changes in connectivity that occur in the cerebellum in autism spectrum disorder (ASD).

A more complex map of the cerebellum

Published in NeuroImage, and featured on the cover, the first study was led by author Xavier Guell, a postdoc in the Gabrieli and Schmahmann labs. The authors used fMRI data from the Human Connectome Project to examine activity in different regions of the cerebellum during specific tasks and at rest. The tasks used extended beyond motor activity to functions recently linked to the cerebellum, including working memory, language, and social and emotional processing. As expected, the authors saw that two regions assigned by other methods to motor activity were clearly modulated during motor tasks.

“Neuroscientists in the 1940s and 1950s described a double representation of motor function in the cerebellum, meaning that two regions in each hemisphere of the cerebellum are engaged in motor control,” explains Guell. “That there are two areas of motor representation in the cerebellum remains one of the most well-established facts of cerebellar macroscale physiology.”

When it came to assigning non-motor tasks, to their surprise, the authors identified three representations that localized to different regions of the cerebellum, pointing to an unexpectedly complex level of organization.

Guell explains the implications further. “Our study supports the intriguing idea that while two parts of the cerebellum are simultaneously engaged in motor tasks, three other parts of the cerebellum are simultaneously engaged in non-motor tasks. Our predecessors coined the term “double motor representation,” and we may now have to add “triple non-motor representation” to the dictionary of cerebellar neuroscience.”

A serendipitous discussion

What happened next, over a discussion of data between Xavier Guell and Sheeba Arnold Anteraper of the McGovern Institute for Brain Research that culminated in a paper led by Anteraper, illustrates how independent strands can meet and reinforce to give a fuller scientific picture.

The findings by Guell and colleagues made the cover of NeuroImage.
The findings by Guell and colleagues made the cover of NeuroImage.

Anteraper and colleagues examined brain images from high-functioning ASD patients, and looked for statistically-significant patterns, letting the data speak rather than focusing on specific ‘candidate’ regions of the brain. To her surprise, networks related to language were highlighted, as well as the cerebellum, regions that had not been linked to ASD, and that seemed at first sight not to be relevant. Scientists interested in language processing, immediately pointed her to Guell.

“When I went to meet him,” says Anteraper, “I saw immediately that he had the same research paper that I’d been reading on his desk. As soon as I showed him my results, the data fell into place and made sense.”

After talking with Guell, they realized that the same non-motor cerebellar representations he had seen, were independently being highlighted by the ASD study.

“When we study brain function in neurological or psychiatric diseases we sometimes have a very clear notion of what parts of the brain we should study” explained Guell, ”We instead asked which parts of the brain have the most abnormal patterns of functional connectivity to other brain areas? This analysis gave us a simple, powerful result. Only the cerebellum survived our strict statistical thresholds.”

The authors found decreased connectivity within the cerebellum in the ASD group, but also decreased strength in connectivity between the cerebellum and the social, emotional and language processing regions in the cerebral cortex.

“Our analysis showed that regions of disrupted functional connectivity mapped to each of the three areas of non-motor representation in the cerebellum. It thus seems that the notion of two motor and three non-motor areas of representation in the cerebellum is not only important for understanding how the cerebellum works, but also important for understanding how the cerebellum becomes dysfunctional in neurology and psychiatry.”

Guell says that many questions remain to be answered. Are these abnormalities in the cerebellum reproducible in other datasets of patients diagnosed with ASD? Why is cerebellar function (and dysfunction) organized in a pattern of multiple representations? What is different between each of these representations, and what is their distinct contribution to diseases such as ASD? Future work is now aimed at unraveling these questions.