the cog-sci collective

we're thinking about machines thinking about our brains

[neurosciencestuff]

Neurodegenerative Diseases Identified Using Artificial Intelligence

Researchers have developed an artificial intelligence platform to detect a range of neurodegenerative disease in human brain tissue samples, including Alzheimer’s disease and chronic traumatic encephalopathy, according to a study conducted at the Icahn School of Medicine at Mount Sinai and published in the Nature medical journal Laboratory Investigation. Their discovery will help scientists develop targeted biomarkers and therapeutics, resulting in a more accurate diagnosis of complex brain diseases that improve patient outcomes.

The buildup of abnormal tau proteins in the brain in neurofibrillary tangles is a feature of Alzheimer’s disease, but it also accumulates in other neurodegenerative diseases, such as chronic traumatic encephalopathy and additional age-related conditions. Accurate diagnosis of neurodegenerative diseases is challenging and requires a highly-trained specialist.  

Researchers at the Center for Computational and Systems Pathology at Mount Sinai developed and used the Precise Informatics Platform to apply powerful machine learning approaches to digitized microscopic slides prepared using tissue samples from patients with a spectrum of neurodegenerative diseases. Applying deep learning, these images were used to create a convolutional neural network capable of identifying neurofibrillary tangles with a high degree of accuracy directly from digitized images.

“Utilizing artificial intelligence has great potential to improve our ability to detect and quantify neurodegenerative diseases, representing a major advance over existing labor-intensive and poorly reproducible approaches,” said lead investigator John Crary, MD, PhD, Professor of Pathology and Neuroscience at the Icahn School of Medicine at Mount Sinai. “Ultimately, this project will lead to more efficient and accurate diagnosis of neurodegenerative diseases.”

This is the first framework available for evaluating deep learning algorithms using large-scale image data in neuropathology. The Precise Informatics Platform allows for data managements, visual exploration, object outlining, multi-user review, and evaluation of deep learning algorithm results.

Researchers at the Center for Computational and Systems Pathology at Mount Sinai have used use advanced computer science and mathematical techniques coupled with cutting-edge microscope technology, computer vision, and artificial intelligence to more accurately classify a broad array of diseases.

“Mount Sinai is the largest academic pathology department in the country and processes more than 80 million tests a year, which offers researchers access to a broad set of data that can be used to improve testing and diagnostics, ultimately leading to better diagnosis and patient outcomes,” said author Carlos Cordon-Cardo, MD, PhD, Chair of the Department of Pathology at the Mount Sinai Health System and Professor of Pathology, Genetics and Genomic Sciences, and Oncological Sciences at the Icahn School of Medicine.

[the-cog-sci-collective]

Awesome crossover between neuroscience and AI. Interested to see where these developments lead!

[the-cog-sci-collective]

fidget spinners: annoying or therapeutic?

As 2019 winds to a close, we’re taking a look at one of the decade’s most infamous trends: the little rotating gadgets known as fidget spinners. What, you might ask, does a fidget spinner have to do with cognitive science? Well, it’s not the toy itself, but rather the process of fidgeting that may have a positive effect on the brain’s ability to concentrate.

The digital age has seen a rise in computer-based jobs that involve long hours sitting at a desk, engaging in somewhat repetitive work. That’s where fidgeting comes in – it breaks the monotony without detracting from your focus. To quote Wired magazine:

“Society increasingly demands mental work while enforcing unhealthy, sedentary physical habits. Fidgeting is a way to cope.”

Outside of the workplace, fidget spinners may have benefits in the classroom, according to researchers at UC Davis. In the study, children with ADHD (attention deficit hyperactivity disorder) were asked to focus on a specific moving arrow while ignoring the other arrows in the animation. The results showed that when the children moved around, their performance on the task improved. These findings indicate that small movements, while they might appear distracting, can actually increase kids’ focus.

But are fidget spinners really the answer? The small bits of plastic are fairly inexpensive, yet they reek of America’s consumer culture, which demands that if you want a solution to your problems – be it the mind-numbing quality of your desk job or your difficulty concentrating in class – you’ll have to pay extra for it. And while that may be true in some cases, perhaps as we head into the new decade, it’s time to return to the forms of fidgeting that require only a pencil, your hands and feet, and whatever else is on hand. 

Citations:

https://www.wired.com/2017/05/fidgeting-makes-you-smarter/

https://www.universityofcalifornia.edu/news/fidgeting-may-help-children-with-adhd-perform-better-school

[the-cog-sci-collective]

can dreaming improve our memory?

“I was in the building for the study of dreams, downstairs in the cafeteria, and I was explaining to a rhinoceros that I was preparing my dreams as my hand-bags, with many objects which could be useful in case” - participant in study

has anyone ever told you that going to sleep after studying will help you remember the information for your test the following day? moreover, what if you dreamed about what you had studied? would that further help remember the information?

a study published last Wednesday in the journal Scientific Reports explores the connection between dream content and memory consolidation. thirty-two “high dream recallers” explored new visuo-olfactory interactive episodes (when they selected areas on a computer image, specific scents were released) for 3 consecutive days. during the nights following each episode, participants woke up at 5am and their usual waking time to record their dreams. a total of 120 dreams were reported and elements related to the learning phase, in which the subjects were exposed to the images and smells, were identified in 37 of the dreams, either learning-related or experiment-related. 

the researchers found that participants with learning-related or experiment related dreams performed similarly to other participants on odor-recognition tasks, but significantly outperformed the other participants on visuo-spatial recognition tasks. this supports their hypothesis that the learning phase is loosely incorporated into dreams and could be related to memory consolidation.

source: https://www.nature.com/articles/s41598-019-51497-y#Sec18

image: https://www.google.com/url?sa=i&source=images&cd=&ved=2ahUKEwjdooK9j8zlAhVYop4KHZ16BLYQjRx6BAgBEAQ&url=https%3A%2F%2Fhow2drawanimals.com%2F8-animals%2F23-draw-rhinoceros.html&psig=AOvVaw1l2ZRK-GGiODZ_9iszyFMb&ust=1572804493009482

[neurosciencestuff]

Babies who hear two languages at home develop advantages in attention

The advantages of growing up in a bilingual home can start as early as six months of age, according to new research led by York University’s Faculty of Health. In the study, infants who are exposed to more than one language show better attentional control than infants who are exposed to only one language. This means that exposure to bilingual environments should be considered a significant factor in the early development of attention in infancy, the researchers say, and could set the stage for lifelong cognitive benefits.

The research was conducted by Ellen Bialystok, Distinguished Research Professor of Psychology and Walter Gordon Research Chair of Lifespan Cognitive Development at York University and Scott Adler, associate professor in York’s Department of Psychology and the Centre for Vision Research, along with lead author Kyle J. Comishen, a former Master’s student in their lab. It was published today in Developmental Science.

The researchers conducted two separate studies in which infants’ eye movements were measured to assess attention and learning. Half of the infants who were studied were being raised in monolingual environments while others were being raised in environments in which they heard two languages spoken approximately half of the time each. The infants were shown images as they lay in a crib equipped with a camera and screen, and their eye movements were tracked and recorded as they watched pictures appear above them, in different areas of the screen. The tracking was conducted 60 times for each infant.

“By studying infants – a population that does not yet speak any language – we discovered that the real difference between monolingual and bilingual individuals later in life is not in the language itself, but rather, in the attention system used to focus on language,” says Bialystok, co-senior author of the study. “This study tells us that from the very earliest stage of development, the networks that are the basis for developing attention are forming differently in infants who are being raised in a bilingual environment. Why is that important? It’s because attention is the basis for all cognition.”

In the first study, the infants saw one of two images in the centre of the screen followed by another image appearing on either the left or right side of the screen. The babies learned to expect that if, for example, a pink and white image appears in the centre of the screen, it would be followed by an attractive target image on the left; If a blue and yellow image appeared in the centre, then the target would appear on the right. All the infants could learn these rules.

In the second study, which began in the same way, researchers switched the rule halfway through the experiment. When they tracked the babies’ eye movements, they found that infants who were exposed to a bilingual environment were better at learning the new rule and at anticipating where the target image would appear. This is difficult because they needed to learn a new association and replace a successful response with a new contrasting one.

“Infants only know which way to look if they can discriminate between the two pictures that appear in the centre,” said Adler, co-senior author of the study. “They will eventually anticipate the picture appearing on the right, for example, by making an eye movement even before that picture appears on the right. What we found was that the infants who were raised in bilingual environments were able to do this better after the rule is switched than those raised in a monolingual environment.”

Anything that comes through the brain’s processing system interacts with this attentional mechanism, says Adler. Therefore, language, as well as visual information, can influence the development of the attentional system.

Researchers say the experience of attending to a complex environment in which infants simultaneously process and contrast two languages may account for why infants raised in bilingual environments have greater attentional control than those raised in monolingual environments.

In previous research, bilingual children and adults outperformed monolinguals on some cognitive tasks that require them to switch responses or deal with conflict. The reason for those differences were thought to follow from the ongoing need for bilinguals to select which language to speak. This new study pushes back the explanation to a time before individuals are actively using languages and switching between them.

[aiweirdness]

Halloween costumes by the neural net GPT-2

In my opinion, one of the best applications of neural networks is for generating Halloween costumes. Thanks to a dataset of over 7,100 costumes crowdsourced from readers of this blog, I’ve been able to generate Halloween costumes with progressively more powerful neural networks. In 2017, I used char-rnn, which learned to generate costumes starting from no knowledge of English (Statue of Pizza, the Fail Witch, Spartan Gandalf, and Professor Panda were some of its inventions). In 2018, I used textgen-rnn, also training from scratch, and teamed up with the New York Times to illustrate the costumes (some of my favorites were Sexy Wizard and Ruth Bader Hat Guy).

Now, as of 2019, there are much more powerful text-generating neural nets around. One of these is GPT-2, trained by OpenAI on a huge dataset of text from the internet. Using the connections it’s gleaned from this huge general dataset, GPT-2 can generate recognizable (if often weird) lists, mushrooms, British snacks, crochet patterns, and even a to-do list for a horrible goose.

So, I trained the 355-M size of GPT-2 (the largest I can currently finetune for free via Max Woolf’s collab notebook)

GPT-2 is good at costumes. Many of its inventions could easily have come from the training data. In fact, the neural net did tend to memorize the training data and repeat it back to me - technically this is what I asked for when I asked it to predict the training data. (The neural net is trying to give me exactly what I ask for, which isn’t necessarily exactly what I want.) I was using a handy script to filter out duplicates (thanks to John Tebbutt), and even so I had to check several of these to make sure they weren’t near copies of the training data. My previous Halloween costume generators would not have been smart enough to come up with things like “jackalope” or “Carl Sagan”, but GPT-2 has seen these words used online in similar contexts to things that ARE in the training data, and it makes the connection.

Gothy Terminator jackalope vampire cat Eye of Sauron incognito llama space cow Vampire Rock Scooby Gadget a raised eyebrow Battle worm Mastodon Swamp girl Carl Sagan A space squirrel walking carpet Frizzle the witch Cleopatra on vacation gothy giraffe Sexy Lego Batman skeleton

Oh yes, the sexy characters. The neural net definitely picked that up from the training data, and innovated admirably, bringing in words that it knew from the internet (barnacle, groundhog, and bunsen burner were not in the list of Halloween costumes), and adding a sexy twist. This is impressive (if somewhat horrifying) work. None of these were in its training data, but I wouldn’t be surprised if some of them exist.

Burlesque Horse Sexy Bulldog Sexy Egg Sexy Parsley Sexy Barnacle Sexy Walrus sexy locust Sexy Titmouse Sexy Hummingbird Sweet Potato Burlesque Sexy Groundhog Sexy bitcoin Sexy DNA Sexy Rubber Duck Sexy Bunsen burner Butt-Monks Sexy gingerbread man Sexy Flying Dutchman Sexy Chimneywatcher Sexy Flames Of War Sexy English Tea Party

And the neural net was pretty good at designing identifiable characters, even if they are a bit on the weird side.

A spangled Auroch manatee M. Bison the Clown Prince of Darkness Gingerbread Man guinea pig Skin Fairy sentient stone fast food bald eagle Fairy root vegetable Ghost in a packet of potato chips cozy coconut Kelpie the mage Crochet monster Walrus rider Star skunk Slytherin AI priest A skunk in a moose suit Semi-molten Kool Aid Man Time Lord Power Ranger The Power Dinosaur Space Oystermonger Deadly Snow Monkey An evil cupcake basic plumber’s equine Spooky mother hen The Bozo the Destroyer Eight Ball of Wrath Ursula, Queen of the Fart Science A poker player in possession of an onion

There are hints, though, that this is the work of an AI rather than the work of someone who understands what costumes are and how they work. These, for example, take somewhat ordinary costume concepts and then make them unnecessarily difficult.

Batman on egg Vampire in hot tub A Hidden Jesus Statue Zombie ice cream cone penguin as a Newt A wizard encased in a icicle Zombie fisherman on a quest Computer generated horse(?) telephone that accepts up to 4 numbers Third Eye Blind Photographed By Dorothy Zombie fisherman w/ lady diegrove tied around foot

And the following costumes are clearly the product of a glitchy AI:

Meat Belt Eyeballed Balloon Men Green beans in bun 10,000 Hands Favorite Caterpillar The Oatmeal Tree 102 SNOWBALLS in a basket Pie and Jell-O List of leg parts world´s nicest fart Pineapple wrapped sasquatch Is it a Snake, a Watermelon, or a Bush? Putting Turtles on Decor Fish tank ‘n chair ROBO-ACCIDENT pajamas made of wood and spiders Ssssssssssexy SSSssssssstinky Ssssssssssssexy ssssssssssssssssexy setup 9 × 11 party trick Smagma Monster Commentary couldn’t be heard over the squawking of clocks Poltergeist might be entertaining, but he’s harder to read in Hungarian Cereal Implanting Device blueberry sipping fizzy pop with eyes of ice blueberry sipping fizzy pop with fake blood on it A sarcastic, racist noble using progressively tinier body parts as a human shield

Bonus content! The above costumes are all from temperature 1.2; I also tried a higher temperature setting, but the generated costumes were at an expert level of chaos (I would like to see someone attempt to go as “hypnopotamus embroidered death”) Enter your email here to get them!

You can order my book You Look Like a Thing and I Love You! It’s out November 5 2019.

Amazon - Barnes & Noble - Indiebound - Tattered Cover - Powell’s

[the-cog-sci-collective]

Happy Halloween from the neural nets of @lewisandquark!

[the-cog-sci-collective]

a pokémon-sized window into the human brain

(Image credit: Kalanit Grill-Spector)

Everything we learn to recognize changes our brain in some way. With functional neuroimaging, we can now measure how the human brain has learned to organize all the things we see. For example, some regions of the visual brain respond most to faces while others respond most to scenes, but not all object categories can claim their own dedicated region (for example, there are no shoe or car regions). Interestingly, these regions are spatially organized in pretty much the same way from brain to brain. These facts raise a number of deep questions: why do some things have localized brain regions while others don’t, and why are these regions where they are? 

A recent study by Gomez, Barnett and Grill-Spector in Nature Human Behavior  used an exceptionally clever approach to gain insight into these questions by measuring the brains of Pokémon experts. The Pokémon experts who participated in the study had all played the handheld video game heavily throughout their childhood, beginning between ages 5 and 8. Gomez et al. show that this extensive early visual experience led to large-scale changes in their adult brain responses. A region in the occipitotemporal sulcus responded more to pictures of Pokémon than to other depicted objects. In contrast, no Pokémon region was found in another group of nonexperts. 

These results alone are worth dwelling on for a moment. Certainly, playing Pokémon for years of one’s life must change the brain at some level of structure. However, what is remarkable is that these changes happened at such a large scale: centimetres of cortex encompassing millions of neurons now have different functional properties as a result of this intensive Pokémon playing. We knew that this kind of experience-based regional formation was possible, as children learning to read develop a nearby region that responds most to words. In fact, young monkeys taught to distinguish letters also develop letter-selective regions in their brain, regions that do not develop in adult monkeys who are given the same training. Along with these previous studies, the current finding of a Pokémon-preferring brain region really drives home just how amazing the plasticity of our developing visual system is.

source: http://konklab.fas.harvard.edu/Papers/Janini_2019_NatHumBeh.pdf

[the-cog-sci-collective]

(mostly) right...for the wrong reasons

you’re in the midst of writing an email when suddenly, in grey text, google suggests the next sentence. mind reading? actually, it’s an application of natural language inference (nli) , the process by which computers determine if one sentence “entails” another. while google might seem to predict your email structure with uncanny accuracy, it can only handle the simple stuff; “let me know...”, “...does that work for you?”, and other phrases found in millions of emails. so at what point do these programs fail, and how can we fix it? that’s what computational linguists r. thomas mccoy, ellie pavlick, and tal linzen are trying to find out. 

this summer, they developed hans, a dataset used to test nli programs on sentences specially crafted to reveal gaps in the program’s understanding. if the programs consistently make incorrect predictions for the hans sentences, the technique being used to make those predictions is flawed. the result? “models trained on MNLI, including the state-of-the-art model BERT, perform very poorly on HANS.” 

why? it boils down to this: all these machine-learning models can do, really, is pick up patterns in their training data. if that data is insufficient, the programs end up making sweeping generalizations about superficial trends rather than learning the underlying causes for those patterns. mccoy, pavlick, and linzen hope that by providing the programs with more complex examples, they will naturally correct their bad assumptions. 

perhaps one day google will be writing our emails for us. 

[the-cog-sci-collective]

mediterranean-style diets linked to better brain function in older adults

The Mediterranean diet is rich in fruits, vegetables, whole grains, beans, potatoes, nuts, olive oil and fish. Processed foods, fried and fast foods, snack foods, red meat, poultry and whole-fat dairy foods are infrequently eaten on the Mediterranean diet.

The MIND diet is a version of the Mediterranean diet that includes 15 types of foods. Ten are considered "brain-healthy:" green leafy vegetables, other vegetables, nuts, berries, beans, whole grains, seafood, poultry, olive oil, and wine. Five are considered unhealthy: red meat, butter and stick margarine, cheese, pastries, sweets and fried/fast foods.

Researchers examined information from 5,907 older adults who participated in the Health and Retirement Study. The participants filled out questionnaires about their eating habits. Researchers then measured the participants' cognitive abilities -- mostly on their memory and attention skills.

The researchers compared the diets of participants to their performance on the cognitive tests. They found that older people who ate Mediterranean and MIND-style diets scored significantly better on the cognitive function tests than those who ate less healthy diets. In fact, older people who ate a Mediterranean-style diet had 35% lower risk of scoring poorly on cognitive tests. Even those who ate a moderate Mediterranean-style diet had 15% lower risk of doing poorly on cognitive tests. The researchers noted similar results for people who ate MIND-style diets.

This study suggests that eating Mediterranean and MIND-style diets is linked to better overall cognitive function in older adults, said the researchers. What's more, older adults who followed these healthy diets had lower risks for having cognitive impairment in later life, noted the researchers.

source: https://www.sciencedaily.com/releases/2017/07/170725154208.htm

image: https://www.google.com/url?sa=i&source=images&cd=&ved=2ahUKEwj6ye2r_8vlAhUgGTQIHbd3CmgQjRx6BAgBEAQ&url=https%3A%2F%2Fwww.themediterraneandish.com%2Fmediterranean-diet-breakfast%2F&psig=AOvVaw2SgvKxPUiibMVAGIgJ-05N&ust=1572800161264116

[the-cog-sci-collective]

evolving science of the spotless mind

what place does machine learning occupy in the field of neuroscience? according to Glaser et al., a prominent one. his team presents four applications of supervised ML to the study of the human brain – solving engineering problems, identifying predictive variables, evaluating simple models of the brain, and modeling brain behavior itself. 

Glaser’s review begins with the much-discussed “predictive power” of supervised ML as a method of solving engineering problems. essentially, the algorithm is “trained” on a large quantity of pre-labeled data, then expected to correctly label data in another dataset. though this method may seem too specific to satisfy a broad range of ML tasks, it’s surprisingly versatile. in neuroscience alone, it can detect regions of white matter in the brain, determine susceptibility to neurodegenerative disease, even convert sensory information into neural signals. the use of ML for such tasks has been well-studied since the mid 2000s; much of Glaser’s discussion simply summarizes these developments. he touches on potential caveats, such as overfitting and the need to hand-select a model that suits your datasets. 

next, Glaser explores the use of ML to identify predictive variables. this is done by weighting the algorithms with certain variables, then evaluating the success of the model. for instance, a model could be fed data about patients’ DNA and asked to predict the likelihood of developing Alzheimer’s disease. based on the results, the researchers could determine whether one section of the gene is most strongly correlated with the risk of disease. the main advantage: ML notices patterns in multi-dimensional vector space that aren’t obvious to the human mind, and therefore can better find variables responsible for trends. the main caveat: correlation doesn’t imply causation. but this issue is not specific to ML – in fact, it crops up in nearly every statistical study. Glaser concludes that the ML approach, given its aptitude for unusual regressions, is a promising method for isolating potentially important neuroscience-related variables.

from here, the review takes a speculative turn. the two final sections look at ML as either a benchmark for models of neural processes, or as a model for the human brain itself. the former possibility is far more cautious – after all, ML is a veritable black box. we can verify that its predictions are accurate – making it a good performance benchmark for other models – but as for internal structure, much is unknown. as Glaser points out, it’s not a great idea to explain an unknown (the brain) with another unknown (ML). however, certain parallels have been found between ML neural network structure and actual neuron connections in the brain. for instance, certain photoreceptors in the retina seem to process data in a similar feed-forward method to ML algorithms. there are discrepancies as well – hardly any human behaviors are learned in the same way as machine learning, making ML a good model for performance, not acquisition. there are no structures in algorithms to mimic the neurotransmitters, glial cells, and other support structures within the brain. neurons are also binary (either signalling or not signalling), while most ML nodes are continuous in signal. Glaser cautions the use of ML to model the brain, but acknowledges that further developments in neural net design as well as unsupervised and reinforcement learning, could yield structures more similar to the human brain. 

[neurosciencestuff]

Navigating Our Thoughts: Fundamental Principles of Thinking

It is one of the most fundamental questions in neuroscience: How do humans think? Until recently, we seemed far from a conclusive answer. However, scientists from the Max-Planck-Institute for Human Cognitive and Brain Sciences (MPI CBS) in Leipzig, Germany, have combined the available evidence, in collaboration with Nobel Prize laureate Edvard Moser from the Kavli Institute for Systems Neuroscience in Trondheim, Norway and Peter Gärdenfors from Lund University in Sweden, to paint a new, comprehensive picture in the journal Science: Humans think using their brain’s navigation system.

When we navigate our environment, two important cell types are active in our brain. Place cells in the hippocampus and grid cells in the neighboring entorhinal cortex form a circuit that allows orientation and navigation. The team of scientists suggests that our inner navigation system does much more. They propose that this system is also key to ‘thinking’, explaining why our knowledge seems to be organized in a spatial fashion.

“We believe that the brain stores information about our surroundings in so-called cognitive spaces. This concerns not only geographical data, but also relationships between objects and experience,” explains Christian Doeller, senior author of the paper and the new director at the MPI CBS.

The term ‘cognitive spaces’ refers to mental maps in which we arrange our experience. Everything that we encounter has physical properties, whether a person or an object, and can therefore be arranged along different dimensions. “If I think about cars, I can order them based on their engine power and weight for example. We would have racing cars with strong engines and low weights as well as caravans with weak engines and high weight, as well as all combinations in between,” says Doeller. “We can think about our family and friends in a similar way; for example, on the basis of their height, humor, or income, coding them as tall or short, humorous or humorless, or more or less wealthy.” Depending on the dimensions of interest individuals might be stored mentally closer together or further away.

A Theory of Human Thinking

In their proposal, Doeller and his team combine individual threads of evidence to form a theory of human thinking. The theory begins with the Nobel Prize-winning discoveries of place and grid cells in rodents’ brains, which were subsequently shown to exist in humans. Both cell types show patterns of activity representing the animal’s position in space, for example, while it forages for food. Each position in space is represented by a unique pattern of activity. Together, the activity of place and grid cells allows the formation of a mental map of the surroundings, which is stored and reactivated during later visits.

The very regular activation pattern of grid cells can also be observed in humans—but importantly, not only during navigation through geographical spaces. Grids cells are also active when learning new concepts, as shown by a study from 2016. In that study, volunteers learned to associate pictures of birds, which only varied in the length of their necks and legs, with different symbols, such as a tree or a bell. A bird with a long neck and short legs was associated with the tree whereas a bird with a short neck and long legs belonged to the bell. Thus, a specific combination of bodily features came to be represented by a symbol.

In a subsequent memory test, performed in a brain scanner, volunteers indicated whether various birds were associated with one of the symbols. Interestingly, the entorhinal cortex was activated, in much the same way as it is during navigation, providing a coordinate system for our thoughts.

“By connecting all these previous discoveries, we came to the assumption that the brain stores a mental map, regardless of whether we are thinking about a real space or the space between dimensions of our thoughts. Our train of thought can be considered a path though the spaces of our thoughts, along different mental dimensions,” Jacob Bellmund, the first author of the publication, explains.

Mapping New Experience

“These processes are especially useful for making inferences about new objects or situations, even if we have never experienced them,” the neuroscientist continues. Using existing maps of cognitive spaces humans can anticipate how similar something new is to something they already know by putting it in relation to existing dimensions. If they’ve already experienced tigers, lions, or panthers, but have never seen a leopard, we would place the leopard in a similar position as the other big cats in our cognitive space. Based on our knowledge about the concept ‘big cat’, already stored in a mental map, we can adequately react to the encounter with the leopard. “We can generalize to novel situations, which we constantly face, and infer how we should behave”, says Bellmund.

[the-cog-sci-collective]

the thought train is an interesting conception, but it fails to address one fundamental aspect of thought: the process of actually having a thought. instead, this article focuses on information storage and recall, a slightly different task. most of us think in sentences, or sentence fragments. how do we piece together the words in our head? is it one word appended to another sequentially, or are we capable of creating thoughts more holistically? these are interesting questions.

[the-cog-sci-collective]

not-so-deep (language) learning

a recent mit article has confirmed what has been long suspected – machines need grammar to learn language. although an interesting insight, this revelation is somewhat disappointing, for one reason: people don’t need grammar to learn language.

chomskyan theory describes human language ability as a dormant seed: far from a full-formed plant, but containing all the structures necessary to flourish under the right conditions. it’s uncertain exactly what these structures are, but given the limited language input we receive as children, there’s no way we can learn all the necessary grammar structures to wield the master tool that is language by age four.

however, computers can (supposedly) process much more information than a toddler can. if language acquisition is based on input of grammatical info, computers should pick up language much better than humans, right? wrong. in fact, the best parsers on the market are still unable to comprehend sentences that would take english speakers only a few seconds to grasp, such as:

"the complex houses married and single soldiers and their families."

granted, this isn’t the easiest sentence to understand; the sentence structure initially tricks us into thinking that “complex” is the adjective,“married” the verb, and “houses” the plural noun subject. in fact, “complex” is the noun, “houses” the verb, and “married” the adjective.

but most of us can still figure it out, while the google and stanford parsers routinely fail. so humans still have a better grasp of the rules of language than computers, despite not being fed millions of grammatical rules.

what does this mean? it means that if we really want computers to rival humans in their grasp of our language, we’d better start with understanding how we grasp language. instead of banking on computers’ info-processing strength to build language models, we might try more delicate approach, modeled after the intricacies of the human mind.

because at this point, the human mind is still the best model we’ve got.

[neurosciencestuff]

The neurobiology of social aggression

Duke-NUS researchers have discovered that a growth factor protein, called brain-derived neurotrophic factor (BDNF), and its receptor, tropomyosin receptor kinase B (TrkB) affects social dominance in mice. The research has implications for understanding the neurobiology of aggression and bullying.

“Humans and rodents are social animals. Our every interaction follows rules according to a social hierarchy. Failure to navigate this hierarchy can be detrimental.” explained senior author A/Prof. Hyunsoo Shawn Je, from the Neuroscience and Behavioural Disorders Signature Research Programme at Duke-NUS Medical School. “Our paper may be the first to demonstrate that specific molecular signalling pathways in specialised nerve cells, in a particular location in the brain, are important for the balanced navigation of social hierarchies.”

Difficulties in navigating these hierarchies can lead to problems like aggression and bullying. “Given the heavy societal cost of bullying and aggression, understanding the biological causes is a step towards their effective prevention and treatment,” A/Prof. Je added.

Activity within the brain is mediated by circuits made up of excitatory neurons, which ramp up activity, and GABA-ergic interneurons, which inhibit and quiet the excitatory activity. Previous studies have shown that BDNF-TrkB signalling is important for the maturation of GABA-ergic interneurons and the development of nerve circuits in the brain. But researchers have not been able to pinpoint the behavioural consequences of disrupted BDNF-TrkB signalling.

A/Prof. Je’s team generated transgenic mice in which the TrkB receptor was removed specifically in the GABAergic interneurons in the area of the brain regulating emotional and social behaviour, known as the corticolimbic system. The transgenic mice exhibited unusual aggressive behaviour when housed together with normal mice. To understand the origin of this behaviour, the team conducted behavioural tests. They found that the mice were not being aggressive to protect their territory. They were also not being aggressive because they were stronger; the transgenic mice were injured more than other mice during acts of aggression. Instead, their aggressive behaviour was a result of increased fighting for status and dominance over other mice in the group.

The researchers found that due to the loss of BDNF-TrkB, GABA-ergic interneurons in these transgenic mice supplied weaker inhibition to surrounding excitatory cells, which became overactive. They proceeded to shut down excitatory neurons in a specific area of the transgenic mice brains, which re-established the “excitatory/inhibitory” balance and which “instantaneously reversed the abnormal social dominance”, says Duke-NUS post-doctoral research fellow Dr. Shawn Pang Hao Tan, who was the first author of the paper.

A significant amount of research has focused on the roles of family and peer networks on aggressive behaviour. This study, together with other recently published findings, demonstrates that genetic and biological factors can play an unexpected role in social behaviours, said Je.

[the-cog-sci-collective]

this post made me think of all the ways in which our brains are susceptible to influence... for now, we can increase aggression in mice, but what of humans? will we see a future where soldiers and WWE fighters undergo gene therapy or neurosurgery to make them more eager to fight? 

[the-cog-sci-collective]

can a computer walk like an egyptian?

a recent study has developed a computational model that, when prioritizing certain spatial variables, can consistently replicate the balanced lateral foot placement of a human. unfortunately, just because a computer can move like a human doesn’t mean it can think like one. it’s a fallacy. 

*fallacy demo*

“when we program the algorithm to make calculations based on these variables, it can successfully place the simulated feet in such a way that the simulated person is balanced”

“therefore, since humans can also place their feet in such a way that they are balanced, their brains must be taking into account the same variables”

basically, an observation of the same phenomenon can’t prove cause. just because i dunk a basketball and my friend dunks another one doesn’t mean he did it the same way i did. 

the researchers in this particular study follow the minimum intervention principle, that is, that the brain only reacts to occurrences that can disrupt its goal. this seems reasonable, and it’s well backed by motor-control studies. the problem here is the researchers’ presumption to know what characteristics a brain might consider disruptive. sure, from a purely physical standpoint, it’s enough to show that certain obvious factors are correlated to human motion, but how can you truly know that the brain is prioritizing such factors over others? returning to the basketball analogy, what if i sidestep a small rock on my way to dunk the ball, and my friend doesn’t notice it, steps on it, but still makes the shot? couldn’t it be that there are multiple ways of achieving the same goal? 

on the other hand, perhaps there is some relationship between our movement regulation and that seen in the computer simulation. but it’s a “perhaps”. 

and if we’re really serious about understanding how the brain makes its calculations, we’re going to have to do better than “perhaps”.

*thanks to @compneuropapers for the original article*

[the-cog-sci-collective]

how must robots trust?

 here’s a problem we might not think of when considering human-robot interactions: how does a robot know whether or not to trust a human? after all, much of artificial learning is integration of sensory information; given the great variation in human activity, it will be important for robots to be able to distinguish between correct and incorrect behavior rather than absorbing information like naive sponges, unable to sift the useful from the confounding.

the royal society recently approached this dilemma, using traditional developmental robotic algorithms and a simple, boolean-based bayesian decision making system. the researchers defined robotic trust as a combination of theory of mind (assigning emotions and beliefs to others), probabilistic unified trust (making determinations of trustworthiness based off of previous experience), and episodic memory (storing interactions as reference points).

an interesting aspect of the study was the development of “immature” and “mature” algorithms. apparently, children under the age of four can’t understand that other people have different perceptions of the world than their own. this understanding of motives is essential for trust, so ideally the “immature” algorithm would be, like a four-year-old, incapable of determining whether someone is trustworthy. and indeed it was.

with the imposition of human attributes (such as human trust) onto artificially intelligent robots comes the question: is the human mind the best basis for evaluating trustworthiness? to what extent can human-based algorithms successfully guide robotic behavior? in this study, the actions being suggested to the robot were repetitive and near-identical, but the researchers hope to extend the findings to “elderly care, robotic-assisted surgery, joint-action hri (human-robot interaction) in the performing arts and autonomous driving”. directions for surgery would certainly vary between practitioners, roads are notorious for unexpected conditions, and elderly people may waver in and out of awareness, making their determinations of ‘right’ and ‘wrong’ quite subjective.

on top of this, if we’re following asimov’s robotic laws, robots may never disobey a command from a human – but if a human is deemed untrustworthy, what is the robot to do? the tension between security and morality is lurking behind these investigations into robotic trust, and the anthropomorphism of robots in general.

*a big thanks to @compneuropapers​, who posted about the study by the royal society*

links: 

https://royalsocietypublishing.org/doi/10.1098/rstb.2018.0032

https://www.verywellmind.com/theory-of-mind-4176826

photo creds:

https://www.fifthdomain.com/opinion/2018/10/08/rethinking-the-concept-of-trust/

[the-cog-sci-collective]

if you gaze long enough into an abyss, the abyss will gaze back into you

nietzsche

welcome to this blog, whose goal is to gaze deeper into the abyss of the human mind (hence cog-sci) but thinks it’s a rather lofty goal to undertake alone (hence collective)

perhaps what we examine may show us something of ourselves.