The risk of jobs being replaced by automation varies by country (credit: World Bank Development Report, 2016)
A new report from the Oxford Martin School and Citi considers the risks of job automation to developing countries, estimated to range from 55% in Uzbekistan to 85% in Ethiopia — a substantial share in major emerging economies, including China and India (77% and 69% respectively).
The Future is Not What is Used to Be provides in-depth analysis of the vulnerabilities of countries and cities to job automation, explores what automation will mean for traditional models of economic growth, and considers how governments can prepare for the potentially disruptive impacts of job automation on society.
47% of US jobs are at risk from automation, but not all cities have the same job risk (credit: Berger, Frey and Osborne/Citi Research report, 2015)
Key areas of analysis in the report include:
While manufacturing productivity has traditionally enabled developing countries to close the gap with richer countries, automation is likely to impact negatively on their ability to do this, and new growth models will be required.
The impact of automation may be more disruptive for developing countries, due to lower levels of consumer demand and limited social safety nets. With automation and developments in 3D printing likely to drive companies to move manufacturing closer to home, developing countries risk :premature de-industrialisation.”
Even within countries, the impact of automation will lead to the divergence of the fortunes of different cities. While a number of cities may have been affected by, for example, offshoring of manufacturing in the past, the expanding scope of automation now means that even low-end service jobs are at risk, making a different set of cities vulnerable.
Risk of total U.S. employment at risk continues since the 2013 study at about 47 percent. Cities in the U.S. most at risk included Fresno and Las Vegas; least at risk include Boston, Washington DC, and New York. In relatively skilled cities, such as Boston, only 38% of jobs are susceptible to automation. In Fresno, by contrast, the equivalent figure is 54%. The computing revolution has been closely linked to the fortunes of U.S. cities, with cities that became centers of information technology gaining a comparative advantage in new job creation that has persisted since. The tendency of skilled jobs to cluster in initially skilled cities has, since the computer revolution of the 1980s, contributed to increased income disparities between cities.
Digital industries have not created many new jobs. Since 2000, just 0.5% of the US workforce has shifted into new technology industries, most of which are directly associated with digital technologies.
The largest number of job openings in the coming decades is projected to be in the health sector, which is expected to add more than 4 million new jobs in the U.S. from 2012 to 2022.
25 least computerizable jobs in U.S. (credit: Carl Benedikt Frey and Michael A. Osborne/Oxford Martin School)
Most investors by surveyed by Citi feel automation poses a major challenge to societies and policymakers, but are optimistic that automation and technology will help to boost productivity over time, and believe that investment in education will be the most effective policy response to the potential negative impacts of automation.
“When it comes to cities, the risk is clear: those that specialize in skills that are easily automatable stand to lose, while the ones that manage the industrial renewal process, particularly by creating new industries, stand to gain,” said Frey, Co-Director of the Oxford Martin Programme on Technology and Employment, and Oxford Martin Citi Fellow.
Kathleen Boyle, Citi GPS Managing Editor, acknowledges that mindsets need to change, saying: “A key challenge of the 21st century will be recognizing that accelerating technological change is going to affect both employment and society.
“The magnitude of the challenge ahead needs to be recognized and an agenda set for policy to address issues such as educational needs, to minimize the negative affect of automation on workers. And it is crucial that this conversation starts now.”
Individual brain cells within a neural network are highlighted in this image obtained using a fluorescent imaging technique (credit: Sandra Kuhlman/CMU)
Carnegie Mellon University is embarking on a five-year, $12 million research effort to reverse-engineer the brain and “make computers think more like humans,” funded by the U.S. Intelligence Advanced Research Projects Activity (IARPA). The research is led by Tai Sing Lee, a professor in the Computer Science Department and the Center for the Neural Basis of Cognition (CNBC).
A “Human Genome Project” for the brain’s visual system
“MICrONS is similar in design and scope to the Human Genome Project, which first sequenced and mapped all human genes,” Lee said. “Its impact will likely be long-lasting and promises to be a game changer in neuroscience and artificial intelligence.”
The researchers will attempt to discover the principles and rules the brain’s visual system uses to process information. They believe this deeper understanding could serve as a springboard to revolutionize machine learning algorithms and computer vision.
In particular, the researchers seek to improve the performance of artificial neural networks — computational models for artificial intelligence inspired by the central nervous systems of animals. Interest in neural nets has recently undergone a resurgence thanks to growing computational power and datasets. Neural nets now are used in a wide variety of applications in which computers can learn to recognize faces, understand speech and handwriting, make decisions for self-driving cars, perform automated trading and detect financial fraud.
How neurons in one region of the visual cortex behave
“But today’s neural nets use algorithms that were essentially developed in the early 1980s,” Lee said. “Powerful as they are, they still aren’t nearly as efficient or powerful as those used by the human brain. For instance, to learn to recognize an object, a computer might need to be shown thousands of labeled examples and taught in a supervised manner, while a person would require only a handful and might not need supervision.”
To better understand the brain’s connections, Sandra Kuhlman, assistant professor of biological sciences at Carnegie Mellon and the CNBC, will use a technique called “two-photon calcium imaging microscopy” to record signaling of tens of thousands of individual neurons in mice as they process visual information, an unprecedented feat. In the past, only a single neuron, or tens of neurons, typically have been sampled in an experiment, she noted.
“By incorporating molecular sensors to monitor neural activity in combination with sophisticated optical methods, it is now possible to simultaneously track the neural dynamics of most, if not all, of the neurons within a brain region,” Kuhlman said. “As a result we will produce a massive dataset that will give us a detailed picture of how neurons in one region of the visual cortex behave.”
A multi-institution research team
Other collaborators are Alan Yuille, the Bloomberg Distinguished Professor of Cognitive Science and Computer Science at Johns Hopkins University, and another MICrONS team at the Wyss Institute for Biologically Inspired Engineering, led by George Church, professor of genetics at Harvard Medical School.
The Harvard-led team, working with investigators at Cold Spring Harbor Laboratory, MIT, and Columbia University, is developing revolutionary techniques to reconstruct the complete circuitry of the neurons recorded at CMU. The database, along with two other databases contributed by other MICrONS teams, unprecedented in scale, will be made publicly available for research groups all over the world.
In this MICrONS project, CMU researchers and their collaborators in other universities will use these massive databases to evaluate a number of computational and learning models as they improve their understanding of the brain’s computational principles and reverse-engineer the data to build better computer algorithms for learning and pattern recognition.
“The hope is that this knowledge will lead to the development of a new generation of machine learning algorithms that will allow AI machines to learn without supervision and from a few examples, which are hallmarks of human intelligence,” Lee said.
The CNBC is a collaborative center between Carnegie Mellon and the University of Pittsburgh. BrainHub is a neuroscience research initiative that brings together the university’s strengths in biology, computer science, psychology, statistics and engineering to foster research on understanding how the structure and activity of the brain give rise to complex behaviors.
The MICrONS team at CMU allso includes Abhinav Gupta, assistant professor of robotics; Gary Miller, professor of computer science; Rob Kass, professor of statistics and machine learning and interim co-director of the CNBC; Byron Yu, associate professor of electrical and computer engineering and biomedical engineering and the CNBC; Steve Chase, assistant professor of biomedical engineering and the CNBC; and Ruslan Salakhutdinov, one of the co-creators of the deep belief network, a new model of machine learning that was inspired by recurrent connections in the brain, who will join CMU as an assistant professor of machine learning in the fall.
Other members of the team include Brent Doiron, associate professor of mathematics at Pitt, and Spencer Smith, assistant professor of neuroscience and neuro-engineering at the University of North Carolina.
Not all machine-intelligence experts are on board with reverse-engineering the brain. In a Facebook post today, Yann LeCun, Director of AI Research at Facebook and a professor at New York University, asked the question in a recent lecture, “Should we copy the brain to build intelligent machines?” “My answer was ‘no, because we need to understand the underlying principles of intelligence to know what to copy. But we should draw inspiration from biology.’”
Portuguese researchers have demonstrated the first swarm of intelligent aquatic surface robots in a real-world environment.
Swarms of aquatic robots have the potential to scale to hundreds or thousands of robots and cover large areas, making them ideal for tasks such as environmental monitoring, search and rescue, and maritime surveillance. They can replace expensive manned vessels and can put the crew out of danger in many maritime missions.
“Swarm robotics is a paradigm shift: we rely on many small, simple and inexpensive robots, instead of a single or a few large, complex and expensive robots,” said Anders Christensen, Ph.D., the principal investigator in the project.
The researchers, from the Institute of Telecommunications at University Institute of Lisbon and from University of Lisbon in Portugal, used nature-inspired approaches for designing their robotic swarm. Instead of manually programming the robots to carry out a mission, evolutionary algorithms are used to synthesize the controller of each robot.
Evolutionary algorithms mimic Darwinian evolution to automatically generate the artificial neural network that allows the robots to carry out the missions autonomously, without a human operator or a central control station. The team demonstrated the capabilities of a swarm with up to ten robots in various collective tasks, including area monitoring, navigation to waypoint, aggregation, and dispersion.
Each robot costs only €300 in materials. The hull of the robots is built from CNC-machined polystyrene foam, and fitted with 3D-printed components. Each robot is equipped with GPS and compass, it can communicate with neighboring robots using Wi-Fi, and the software runs on an onboard $40 Raspberry Pi 2 computer.
The team is now developing the second generation of their aquatic robots, which will be equipped with more advanced sensors and able to carry out long-term missions at sea. The research was funded by Fundação para a Ciência e Tecnologia, Portugal.
AAAI Video Competition | A Sea of Robots
Abstract of Evolution of Collective Behaviors for a Real Swarm of Aquatic Surface Robots
Studies on swarm robotics systems have shown the potential of large-scale multirobot systems based on decentralized control. So far, studies have been mostly conducted in simulation, and the few that have been conducted on real robots have been confined to laboratory environments. In this paper, we present experiments with an autonomous swarm of up to ten aquatic surface robots, conducted in a real and uncontrolled environment. We evolve neural network-based controllers in simulation for canonical swarm robotics tasks, and then assess their performance on real hardware. Our results show that the evolved controllers generally transfer well to the real robots, displaying a similar performance to that obtained in simulation, and similar behavior patterns. We also show that the evolved control maintain key properties of swarm intelligence-based control, namely scalability, flexibility, and robustness. We conclude with a proof-of-concept experiment in which the swarm performs an environmental monitoring task by combining multiple evolved controllers.
According to the CDC, Pseudomonas aeruginosa is a common cause of healthcare-associated infections, including pneumonia, bloodstream infections, urinary tract infections, and surgical site infections. Some strains of P. aeruginosa have been found to be resistant to nearly all or all antibiotics. (illustration credit: CDC)
A new machine-learning technique can uncover previously unknown features of organisms and their genes in large datasets, according to researchers from the Perelman School of Medicine at the University of Pennsylvania and the Geisel School of Medicine at Dartmouth University.
For example, the technique learned to identify the characteristic gene-expression patterns that appear when a bacterium is exposed in different conditions, such as low oxygen and the presence of antibiotics.
The technique, called “ADAGE” (Analysis using Denoising Autoencoders of Gene Expression), uses a “denoising autoencoder” algorithm, which learns to identify recurring features or patterns in large datasets — without being told what specific features to look for (that is, “unsupervised.”)*
Last year, Casey Greene, PhD, an assistant professor of Systems Pharmacology and Translational Therapeutics at Penn, and his team published, in an open-access paper in the American Society for Microbiology’s mSystems, the first demonstration of ADAGE in a biological context: an analysis of two gene-expression datasets of breast cancers.
Tracking down gene patterns of a multi-drug-resistant bacterium
The new study, published Jan. 19 in an open-access paper in mSystems, was more ambitious. It applied ADAGE to a dataset of 950 gene-expression arrays publicly available at the time for the multi-drug-resistant bacterium Pseudomonas aeruginosa. This bacterium is a notorious pathogen in the hospital and in individuals with cystic fibrosis and other chronic lung conditions; it’s often difficult to treat due to its high resistance to standard antibiotic therapies.
The data included only the identities of the roughly 5,000 P. aeruginosa genes and their measured expression levels in each published experiment. The goal was to see if this “unsupervised” learning system could uncover important patterns in P. aeruginosa gene expression and clarify how those patterns change when the bacterium’s environment changes — for example, when in the presence of an antibiotic.
Even though the model built with ADAGE was relatively simple — roughly equivalent to a brain with only a few dozen neurons — it had no trouble learning which sets of P. aeruginosa genes tend to work together or in opposition. To the researchers’ surprise, the ADAGE system also detected differences between the main laboratory strain of P. aeruginosa and strains isolated from infected patients. “That turned out to be one of the strongest features of the data,” Greene said.
“We expect that this approach will be particularly useful to microbiologists researching bacterial species that lack a decades-long history of study in the lab,” said Greene. “Microbiologists can use these models to identify where the data agree with their own knowledge and where the data seem to be pointing in a different direction … and to find completely new things in biology that we didn’t even know to look for.”
Support for the research came from the Gordon and Betty Moore Foundation, the William H. Neukom Institute for Computational Science, the National Institutes of Health, and the Cystic Fibrosis Foundation.
* In 2012, Google-sponsored researchers applied a similar method to randomly selected YouTube images; their system learned to recognize major recurring features of those images — including cats of course.
Abstract of ADAGE-Based Integration of Publicly Available Pseudomonas aeruginosa Gene Expression Data with Denoising Autoencoders Illuminates Microbe-Host Interactions
The increasing number of genome-wide assays of gene expression available from public databases presents opportunities for computational methods that facilitate hypothesis generation and biological interpretation of these data. We present an unsupervised machine learning approach, ADAGE (analysis using denoising autoencoders of gene expression), and apply it to the publicly available gene expression data compendium for Pseudomonas aeruginosa. In this approach, the machine-learned ADAGE model contained 50 nodes which we predicted would correspond to gene expression patterns across the gene expression compendium. While no biological knowledge was used during model construction, cooperonic genes had similar weights across nodes, and genes with similar weights across nodes were significantly more likely to share KEGG pathways. By analyzing newly generated and previously published microarray and transcriptome sequencing data, the ADAGE model identified differences between strains, modeled the cellular response to low oxygen, and predicted the involvement of biological processes based on low-level gene expression differences. ADAGE compared favorably with traditional principal component analysis and independent component analysis approaches in its ability to extract validated patterns, and based on our analyses, we propose that these approaches differ in the types of patterns they preferentially identify. We provide the ADAGE model with analysis of all publicly available P. aeruginosa GeneChip experiments and open source code for use with other species and settings. Extraction of consistent patterns across large-scale collections of genomic data using methods like ADAGE provides the opportunity to identify general principles and biologically important patterns in microbial biology. This approach will be particularly useful in less-well-studied microbial species.
Abstract of Unsupervised feature construction and knowledge extraction from genome-wide assays of breast cancer with denoising autoencoders
Big data bring new opportunities for methods that efficiently summarize and automatically extract knowledge from such compendia. While both supervised learning algorithms and unsupervised clustering algorithms have been successfully applied to biological data, they are either dependent on known biology or limited to discerning the most significant signals in the data. Here we present denoising autoencoders (DAs), which employ a data-defined learning objective independent of known biology, as a method to identify and extract complex patterns from genomic data. We evaluate the performance of DAs by applying them to a large collection of breast cancer gene expression data. Results show that DAs successfully construct features that contain both clinical and molecular information. There are features that represent tumor or normal samples, estrogen receptor (ER) status, and molecular subtypes. Features constructed by the autoencoder generalize to an independent dataset collected using a distinct experimental platform. By integrating data from ENCODE for feature interpretation, we discover a feature representing ER status through association with key transcription factors in breast cancer. We also identify a feature highly predictive of patient survival and it is enriched by FOXM1 signaling pathway. The features constructed by DAs are often bimodally distributed with one peak near zero and another near one, which facilitates discretization. In summary, we demonstrate that DAs effectively extract key biological principles from gene expression data and summarize them into constructed features with convenient properties.
Go is played on a grid of black lines (usually 19×19). Game pieces, called stones, are played on the line intersections. (credit: Goban1/Wikipedia)
A deep-learning computer system called AlphaGo created by Google’s DeepMind team has defeated reigning three-time European Go champion Fan Hui 5 games to 0 — the first time a computer program has ever beaten a professional Go player, reports Google Research blog today (Jan. 27) — a feat previously thought to be at least a decade away.
“AlphaGo uses general machine-learning techniques to allow it to improve itself, just by watching and playing games,” according to David Silver and Demis Hassabis of Google DeepMind. Using a vast collection of more than 30 million Go moves from expert players, DeepMind researchers trained their system to play Go on its own.
To achieve that, AlphaGo, as described in a paper in Nature today, combines a state-of-the-art tree search with two deep neural networks, each containsing many layers with millions of neuron-like connections needed to deal with Go’s vast search space — more than a googol (10100) times larger than chess (a number greater than there are atoms in the universe).
“We first trained [one of the two networks] on 30 million moves from games played by human experts, until it could predict the human move 57% of the time …. But our goal is to beat the best human players, not just mimic them, Silver and Hassabis said. “To do this, AlphaGo learned to discover new strategies for itself, by playing thousands of games between its neural networks, and gradually improving them using a trial-and-error process known as reinforcement learning.”
This figure from the Nature article shows the Elo rating (a 230 point gap corresponds to a 79% probability of winning) and approximate rank of AlphaGo (both single machine and distributed versions), the European champion Fan Hui (a professional 2-dan), and the strongest other Go programs, evaluated over thousands of games. Pale pink bars show the performance of other programs when given a four move headstart. (credit: David Silver et al./Nature)
AlphaGo’s next challenge will be to play the top Go player in the world over the last decade, Lee Sedol. The match will take place this March in Seoul, South Korea.
“While games are the perfect platform for developing and testing AI algorithms quickly and efficiently, ultimately we want to apply these techniques to important real-world problems,” the researchers say. “Because the methods we have used are general purpose, our hope is that one day they could be extended to help us address some of society’s toughest and most pressing problems, from climate modelling to complex disease analysis.”
Google DeepMind | Ground-breaking AlphaGo masters the game of Go
Abstract of Mastering the game of Go with deep neural networks and tree search
The game of Go has long been viewed as the most challenging of classic games for artificial intelligence owing to its enormous search space and the difficulty of evaluating board positions and moves. Here we introduce a new approach to computer Go that uses ‘value networks’ to evaluate board positions and ‘policy networks’ to select moves. These deep neural networks are trained by a novel combination of supervised learning from human expert games, and reinforcement learning from games of self-play. Without any lookahead search, the neural networks play Go at the level of state-of-the-art Monte Carlo tree search programs that simulate thousands of random games of self-play. We also introduce a new search algorithm that combines Monte Carlo simulation with value and policy networks. Using this search algorithm, our program AlphaGo achieved a 99.8% winning rate against other Go programs, and defeated the human European Go champion by 5 games to 0. This is the first time that a computer program has defeated a human professional player in the full-sized game of Go, a feat previously thought to be at least a decade away.
Pressure sensors wrap around and conform to the shape of the fingers while still accurately measuring pressure distribution. (credit: 2016 Someya Laboratory)
Doctors may one day be able to physically screen for breast cancer using pressure-sensitive rubber gloves to detect tumors, thanks to a transparent, bendable, and sensitive pressure sensor newly developed by Japanese and American teams.
Conventional pressure sensors can’t measure pressure changes accurately once they are twisted or wrinkled, making them unsuitable for use on complex and moving surfaces, and they can’t be miniaturized below 100 micrometers (0.1 millimeters) thickness because of limitations in current production methods.
To address these issues, an international team of researchers led by Dr. Sungwon Lee and Professor Takao Someya of the University of Tokyo’s Graduate School of Engineeringhas developed a nanofiber-type pressure sensor made from carbon nanotubes and graphene that can measure pressure distribution of rounded surfaces such as an inflated balloon and maintain its sensing accuracy even when bent over a radius of 80 micrometers, equivalent to just twice the width of a human hair. The sensor is roughly 8 micrometers thick and can measure the pressure in 144 locations at once.
The device demonstrated in this study consists of organic transistors, electronic switches made from carbon and oxygen-based organic materials, and a pressure-sensitive nanofiber structure. Carbon nanotubes and graphene were added to an elastic polymer to create nanofibers with a diameter of 300 to 700 nanometers, which were then entangled with each other to form a transparent, thin and light porous structure.
The material may also have applications in improving the touch sensitivity in robots.
Abstract of A transparent bending-insensitive pressure sensor
Measuring small normal pressures is essential to accurately evaluate external stimuli in curvilinear and dynamic surfaces such as natural tissues. Usually, sensitive and spatially accurate pressure sensors are achieved through conformal contact with the surface; however, this also makes them sensitive to mechanical deformation (bending). Indeed, when a soft object is pressed by another soft object, the normal pressure cannot be measured independently from the mechanical stress. Here, we show a pressure sensor that measures only the normal pressure, even under extreme bending conditions. To reduce the bending sensitivity, we use composite nanofibres of carbon nanotubes and graphene. Our simulations show that these fibres change their relative alignment to accommodate bending deformation, thus reducing the strain in individual fibres. Pressure sensitivity is maintained down to a bending radius of 80 μm. To test the suitability of our sensor for soft robotics and medical applications, we fabricated an integrated sensor matrix that is only 2 μm thick. We show real-time (response time of ∼20 ms), large-area, normal pressure monitoring under different, complex bending conditions.
In a computational reconstruction of brain tissue in the hippocampus, Salk and UT-Austin scientists found the unusual occurrence of two synapses from the axon of one neuron (translucent black strip) forming onto two spines on the same dendrite of a second neuron (yellow). Separate terminals from one neuron’s axon are shown in synaptic contact with two spines (arrows) on the same dendrite of a second neuron in the hippocampus. The spine head volumes, synaptic contact areas (red), neck diameters (gray) and number of presynaptic vesicles (white spheres) of these two synapses are almost identical. (credit: Salk Institute)
Salk researchers and collaborators have achieved critical insight into the size of neural connections, putting the memory capacity of the brain far higher than common estimates. The new work also answers a longstanding question as to how the brain is so energy efficient, and could help engineers build computers that are incredibly powerful but also conserve energy.
“This is a real bombshell in the field of neuroscience,” says Terry Sejnowski, Salk professor and co-senior author of the paper, which was published in eLife. “We discovered the key to unlocking the design principle for how hippocampal neurons function with low energy but high computation power. Our new measurements of the brain’s memory capacity increase conservative estimates by a factor of 10 to at least a petabyte (1 quadrillion or 1015 bytes), in the same ballpark as the World Wide Web.”
“When we first reconstructed every dendrite, axon, glial process, and synapse* from a volume of hippocampus the size of a single red blood cell, we were somewhat bewildered by the complexity and diversity amongst the synapses,” says Kristen Harris, co-senior author of the work and professor of neuroscience at the University of Texas, Austin. “While I had hoped to learn fundamental principles about how the brain is organized from these detailed reconstructions, I have been truly amazed at the precision obtained in the analyses of this report.”
10 times more discrete sizes of synapses discovered
The Salk team, while building a 3D reconstruction of rat hippocampus tissue (the memory center of the brain), noticed something unusual. In some cases, a single axon from one neuron formed two synapses reaching out to a single dendrite of a second neuron, signifying that the first neuron seemed to be sending a duplicate message to the receiving neuron.
At first, the researchers didn’t think much of this duplicity, which occurs about 10 percent of the time in the hippocampus. But Tom Bartol, a Salk staff scientist, had an idea: if they could measure the difference between two very similar synapses such as these, they might glean insight into synaptic sizes, which so far had only been classified in the field as small, medium and large.
To do this, researchers used advanced microscopy and computational algorithms they had developed to image rat brains and reconstruct the connectivity, shapes, volumes and surface area of the brain tissue down to a nanomolecular level.
The scientists expected the synapses would be roughly similar in size, but were surprised to discover the synapses were nearly identical.
Salk scientists computationally reconstructed brain tissue in the hippocampus to study the sizes of connections (synapses). The larger the synapse, the more likely the neuron will send a signal to a neighboring neuron. The team found that there are actually 26 discrete sizes that can change over a span of a few minutes, meaning that the brain has a far great capacity at storing information than previously thought. Pictured here is a synapse between an axon (green) and dendrite (yellow). (credit: Salk Institute)
“We were amazed to find that the difference in the sizes of the pairs of synapses were very small, on average, only about eight percent different in size. No one thought it would be such a small difference. This was a curveball from nature,” says Bartol.
Because the memory capacity of neurons is dependent upon synapse size, this eight percent difference turned out to be a key number the team could then plug into their algorithmic models of the brain to measure how much information could potentially be stored in synaptic connections.
It was known before that the range in sizes between the smallest and largest synapses was a factor of 60 and that most are small.
But armed with the knowledge that synapses of all sizes could vary in increments as little as eight percent between sizes within a factor of 60, the team determined there could be about 26 categories of sizes of synapses, rather than just a few.
“Our data suggests there are 10 times more discrete sizes of synapses than previously thought,” says Bartol. In computer terms, 26 sizes of synapses correspond to about 4.7 “bits” of information. Previously, it was thought that the brain was capable of just one to two bits for short and long memory storage in the hippocampus.
“This is roughly an order of magnitude of precision more than anyone has ever imagined,” says Sejnowski.
What makes this precision puzzling is that hippocampal synapses are notoriously unreliable. When a signal travels from one neuron to another, it typically activates that second neuron only 10 to 20 percent of the time.
“We had often wondered how the remarkable precision of the brain can come out of such unreliable synapses,” says Bartol. One answer, it seems, is in the constant adjustment of synapses, averaging out their success and failure rates over time. The team used their new data and a statistical model to find out how many signals it would take a pair of synapses to get to that eight percent difference.
The researchers calculated that for the smallest synapses, about 1,500 events cause a change in their size/ability (20 minutes) and for the largest synapses, only a couple hundred signaling events (1 to 2 minutes) cause a change.
“This means that every 2 or 20 minutes, your synapses are going up or down to the next size. The synapses are adjusting themselves according to the signals they receive,” says Bartol.
“Our prior work had hinted at the possibility that spines and axons that synapse together would be similar in size, but the reality of the precision is truly remarkable and lays the foundation for whole new ways to think about brains and computers,” says Harris. “The work resulting from this collaboration has opened a new chapter in the search for learning and memory mechanisms.” Harris adds that the findings suggest more questions to explore, for example, if similar rules apply for synapses in other regions of the brain and how those rules differ during development and as synapses change during the initial stages of learning.
“The implications of what we found are far-reaching,” adds Sejnowski. “Hidden under the apparent chaos and messiness of the brain is an underlying precision to the size and shapes of synapses that was hidden from us.”
A model for energy-efficient computers
The findings also offer a valuable explanation for the brain’s surprising efficiency. The waking adult brain generates only about 20 watts of continuous power—as much as a very dim light bulb. The Salk discovery could help computer scientists build powerful and ultraprecise, but energy-efficient, computers, particularly ones that employ “deep learning” and artificial neural nets—techniques capable of sophisticated learning and analysis, such as speech, object recognition and translation.
“This trick of the brain absolutely points to a way to design better computers,” says Sejnowski. “Using probabilistic transmission turns out to be as accurate and require much less energy for both computers and brains.”
Other authors on the paper were Cailey Bromer of the Salk Institute; Justin Kinney of the McGovern Institute for Brain Research; and Michael A. Chirillo and Jennifer N. Bourne of the University of Texas, Austin.
* Our memories and thoughts are the result of patterns of electrical and chemical activity in the brain. A key part of the activity happens when branches of neurons, much like electrical wire, interact at certain junctions, known as synapses. An output ‘wire’ (an axon) from one neuron connects to an input ‘wire’ (a dendrite) of a second neuron. Signals travel across the synapse as chemicals called neurotransmitters to tell the receiving neuron whether to convey an electrical signal to other neurons. Each neuron can have thousands of these synapses with thousands of other neurons. Synapses are still a mystery, though their dysfunction can cause a range of neurological diseases. Larger synapses — with more surface area and vesicles of neurotransmitters — are stronger, making them more likely to activate their surrounding neurons than medium or small synapses.
UPDATE 1/22/2016 “in the same ballpark as the World Wide Web” removed; appears to be inaccurate. The Internet Archive, a subset of the Web, currently stores 50 petabytes, for example.
Salk Institute | Salk scientists computationally reconstructed brain tissue in the hippocampus to study the sizes of connections (synapses). The larger the synapse, the more likely the neuron will send a signal to a neighboring neuron. The team found that there are actually 26 discrete sizes that can change over a span of a few minutes, meaning that the brain has a far great capacity at storing information than previous
Abstract of Nanoconnectomic upper bound on the variability of synaptic plasticity
Information in a computer is quantified by the number of bits that can be stored and recovered. An important question about the brain is how much information can be stored at a synapse through synaptic plasticity, which depends on the history of probabilistic synaptic activity. The strong correlation between size and efficacy of a synapse allowed us to estimate the variability of synaptic plasticity. In an EM reconstruction of hippocampal neuropil we found single axons making two or more synaptic contacts onto the same dendrites, having shared histories of presynaptic and postsynaptic activity. The spine heads and neck diameters, but not neck lengths, of these pairs were nearly identical in size. We found that there is a minimum of 26 distinguishable synaptic strengths, corresponding to storing 4.7 bits of information at each synapse. Because of stochastic variability of synaptic activation the observed precision requires averaging activity over several minutes.
A computer scientist and biologist propose to unify the theory of evolution with learning theories to explain the “amazing, apparently intelligent designs that evolution produces.”
The scientists — University of Southampton School of Electronics and Computer Science professor Richard Watson* and Eötvös Loránd University (Budapest) professor of biology Eörs Szathmáry* — say they’ve found that it’s possible for evolution to exhibit some of the same intelligent behaviors as learning systems — including neural networks.
Writing in an opinion paper published in the journal Trends in Ecology and Evolution, they use “formal analogies” and transfer specific models and results between the two theories in an attempt to solve several evolutionary puzzles.
The authors cite work by Pavlicev and colleagues** showing that selection on relational alleles (gene variants) increases phenotypic (organism trait) correlation if the traits are selected together and decreases correlation if they are selected antagonistically, which is characteristic of Hebbian learning, they note.
“This simple step from evolving traits to evolving correlations between traits is crucial; it moves the object of natural selection from fit phenotypes (which ultimately removes phenotypic variability altogether) to the control of phenotypic variability,” the researchers say.
Why evolution is not blind
“Learning theory is not just a different way of describing what Darwin already told us,” said Watson. “It expands what we think evolution is capable of. It shows that natural selection is sufficient to produce significant features of intelligent problem-solving.”
Conventionally, evolution, which depends on random variation, has been considered blind, or at least myopic, he notes. “But showing that evolving systems can learn from past experience means that evolution has the potential to anticipate what is needed to adapt to future environments in the same way that learning systems do.
“A system exhibits learning if its performance at some task improves with experience,” the authors note in the paper. “Reusing behaviors that have been successful in the past (reinforcement learning) is intuitively similar to the way selection increases the proportion of fit phenotypes [an organism's observable characteristics or traits] in a population. In fact, evolutionary processes and simple learning processes are formally equivalent.
“In particular, learning can be implemented by incrementally adjusting a probability distribution over behaviors (e.g., Bayesian learning or Bayesian updating). Or, if a behavior is represented by a vector of features or components, by adjusting the probability of using each individual component in proportion to its average reward in past behaviors (e.g., Multiplicative Weights Update Algorithm, MWUA).”
The evolution of connections in a Recurrent Gene Regulation Network (GRN) shows associative learning behaviors. When a Hopfield network is trained on a set of patterns with Hebbian learning, it forms an associative memory of the patterns in the training set. When subsequently stimulated with random excitation patterns, the activation dynamics of the trained network will spontaneously recall the patterns from the training set or generate new patterns that are generalizations of the training patterns. (A–D) A GRN is evolved to produce first one phenotype (set of characteristics or traits — Charles Darwin in this example) and then another (Donald Hebb) in an alternating manner. The resulting phenotype is not merely an average of the two phenotypic patterns that were selected in the past. Rather, different embryonic phenotypes (e.g., random initial conditions C and D) developed into different adult phenotypes (with this evolved GRN) and match either A or B. These two phenotypes can be produced from genotypes (DNA sequences) that are a single mutation apart. In a separate experiment, selection iterates over a set of target phenotypes (E–H). In addition to developing phenotypes that match patterns selected in the past (e.g., I), this GRN also generalizes to produce new phenotypes that were not selected for in the past but belong to a structurally similar class, for example, by creating novel combinations of evolved modules (e.g., developmental attractors exist for a phenotype with all four “loops” (J). This demonstrates a capability for evolution to exhibit phenotypic novelty in exactly the same sense that learning neural networks can generalize from past experience. (credit: Richard A. Watson and Eörs Szathmáry/Trends in Ecology and Evolution)
Unsupervised learning
An even more interesting process in evolution is unsupervised learning, where mechanisms do not depend on an external reward signal, the authors say in the paper:
By reinforcing correlations that are frequent, regardless of whether they are good, unsupervised correlation learning can produce system-level behaviors without system-level rewards. This can be implemented without centralized learning mechanisms. (Recent theoretical work shows that selection acting only to maximize individual growth rate, when applied to interspecific competition coefficients within an ecological community, produces unsupervised learning at the system level.)
This is an exciting possibility because it means that, despite not being a unit of selection, an ecological community might exhibit organizations that confer coordinated collective behaviors — for example, a distributed ecological memory that can recall multiple past ecological states. …
Taken together, correlation learning, unsupervised correlation learning, and deep correlation learning thus provide a formal way to understand how variation, selection, and inheritance, respectively, might be transformed over evolutionary time.
So if evolutionary theory can learn from the principles of cognitive science and deep learning, can cognitive science and deep learning learn from evolutionary theory?
** Watson, R.A. et al. (2014) The evolution of phenotypic correlations and ‘developmental memory.’ Evolution 68, 1124–1138 and Pavlicev, M.et al. (2011) Evolution of adaptive phenotypic variation patterns by direct selection for evolvability. Proc. R. Soc. B Biol. Sci. 278, 1903–1912
*** This book is available free on KurzweilAI, as noted.
Abstract of How Can Evolution Learn?
The theory of evolution links random variation and selection to incremental adaptation. In a different intellectual domain, learning theory links incremental adaptation (e.g., from positive and/or negative reinforcement) to intelligent behaviour. Specifically, learning theory explains how incremental adaptation can acquire knowledge from past experience and use it to direct future behaviours toward favourable outcomes. Until recently such cognitive learning seemed irrelevant to the ‘uninformed’ process of evolution. In our opinion, however, new results formally linking evolutionary processes to the principles of learning might provide solutions to several evolutionary puzzles – the evolution of evolvability, the evolution of ecological organisation, and evolutionary transitions in individuality. If so, the ability for evolution to learn might explain how it produces such apparently intelligent designs.
Human Computation Institute | Dr. Pietro Michelucci
“Human computation” — combining human and computer intelligence in crowd-powered systems — might be what we need to solve the “wicked” problems of the world, such as climate change and geopolitical conflict, say researchers from the Human Computation Institute (HCI) and Cornell University.
In an article published in the journal Science, the authors present a new vision of human computation that takes on hard problems that until recently have remained out of reach.
Humans surpass machines at many things, ranging from visual pattern recognition to creative abstraction. And with the help of computers, these cognitive abilities can be effectively combined into multidimensional collaborative networks that achieve what traditional problem-solving cannot, the authors say.
Microtasking
Microtasking: Crowdsourcing breaks large tasks down into microtasks, which can be things at which humans excel, like classifying images. The microtasks are delivered to a large crowd via a user-friendly interface, and the data are aggregated for further processing. (credit: Pietro Michelucci and Janis L. Dickinson/Science)
Most of today’s human-computation systems rely on “microtasking” — sending “micro-tasks” to many individuals and then stitching together the results. For example, 165,000 volunteers in EyeWire have analyzed thousands of images online to help build the world’s most complete map of human retinal neurons.
Another example is reCAPTCHA, a Web widget used by 100 million people a day when they transcribe distorted text into a box to prove they are human.
“Microtasking is well suited to problems that can be addressed by repeatedly applying the same simple process to each part of a larger data set, such as stitching together photographs contributed by residents to decide where to drop water during a forest fire,” the authors note.
But this microtasking approach alone cannot address the tough challenges we face today, say the authors. “A radically new approach is needed to solve ‘wicked problems’ — those that involve many interacting systems that are constantly changing, and whose solutions have unforeseen consequences, such as climate change, disease, and geopolitical conflict, which are dynamic, involve multiple, interacting systems, and have non-obvious secondary effects, such as political exploitation of a pandemic crisis.”
New human-computation technologies
New human-computation technologies: In creating problem-solving ecosystems, researchers are beginning to explore how to combine the cognitive processing of many human contributors with machine-based computing to build faithful models of the complex, interdependent systems that underlie the world’s most challenging problems. (credit: Pietro Michelucci and Janis L. Dickinson/Science)
The authors say new human computation technologies can help build flexible collaborative environments. Recent techniques provide real-time access to crowd-based inputs, where individual contributions can be processed by a computer and sent to the next person for improvement or analysis of a different kind.
This idea is already taking shape in several human-computation projects:
YardMap.org, launched by the Cornell in 2012, maps global conservation efforts. It allows participants to interact and build on each other’s work — something that crowdsourcing alone cannot achieve.
WeCureAlz.com accelerates Cornell-based Alzheimer’s disease research by combining two successful microtasking systems into an interactive analytic pipeline that builds blood-flow models of mouse brains. The stardust@home system, which was used to search for comet dust in one million images of aerogel, is being adapted to identify stalled blood vessels, which will then be pinpointed in the brain by a modified version of the EyeWire system.
“By enabling members of the general public to play some simple online game, we expect to reduce the time to treatment discovery from decades to just a few years,” says HCI director and lead author, Pietro Michelucci, PhD. “This gives an opportunity for anyone, including the tech-savvy generation of caregivers and early stage AD patients, to take the matter into their own hands.”
Abstract of The power of crowds
Human computation, a term introduced by Luis von Ahn, refers to distributed systems that combine the strengths of humans and computers to accomplish tasks that neither can do alone. The seminal example is reCAPTCHA, a Web widget used by 100 million people a day when they transcribe distorted text into a box to prove they are human. This free cognitive labor provides users with access to Web content and keeps websites safe from spam attacks, while feeding into a massive, crowd-powered transcription engine that has digitized 13 million articles from The New York Times archives. But perhaps the best known example of human computation is Wikipedia. Despite initial concerns about accuracy, it has become the key resource for all kinds of basic information. Information science has begun to build on these early successes, demonstrating the potential to evolve human computation systems that can model and address wicked problems (those that defy traditional problem-solving methods) at the intersection of economic, environmental, and sociopolitical systems.
Human brain connectome (credit: NIH Human Connectome Project)
Ever wonder how your brain creates your thoughts, based on everything that’s happening around you (and within you), and where these thoughts are actually located in the brain?
UMass Amherst computational neuroscientist Hava Siegelmann has, and she created a geometry-based method for doing just that. Her team did a massive data analysis of 20 years of functional magnetic resonance imaging (fMRI) data from tens of thousands of brain imaging experiments. The goal was to understand how abstract thought arises from brain structure, which could lead to better ways to identify and treat brain disease and even to new deep-learning artificial intelligence (AI) systems.
Details appear in an open-access article in the current issue of Nature Scientific Reports.
How abstract thoughts are formed
KurzweilAI has covered more than 200 research projects involving fMRI. Basically, fMRI detects changes in neural blood flow, which relates to specific brain activities (such as imagining what an object looks like, or talking). More blood flow means higher levels of neural activity in that specific brain region. While fMRI-based research has done an impressive job of relating specific brain areas with activities, surprisingly, “no one had ever tied together the tens of thousands of experiments performed over decades to show how the physical brain could give rise to abstract thought,” Siegelmann notes.
For this study, the researchers took a data-science approach. First, they defined a physiological directed network (a form of a graph with nodes and links) of the whole brain, starting at input areas and labeling each brain area with the distance (or “depth”) from sensory inputs. For example, in the drawing below, the visual cortex (in green) is located far away from the eyes (on the left) while the auditory cortex (in yellow) is relatively close to the ears (although routing via the thalamus makes this more complex).
OK, so what does that mean in terms of thinking? To find out, they processed a massive repository of fMRI data from about 17,000 experiments, representing about one fourth of the fMRI literature).
Regions of motor and sensory cortex (credit: Blausen.com staff/Blausen gallery 2014/Wikiversity)
“The idea was to project the active regions for a cognitive behavior onto the network depth and describe that cognitive behavior in terms of its depth distribution,” says Siegelmann. “We momentarily thought our research failed when we saw that each cognitive behavior showed activity through many network depths. Then we realized that cognition is far richer; it wasn’t the simple hierarchy that everyone was looking for. So, we developed our geometrical ‘slope’ algorithm.”
Ranking cognitive behaviors
The researchers summed all neural activity for a given behavior over all related fMRI experiments, then analyzed it using the slope algorithm. “With a slope identifier, behaviors could now be ordered by their relative depth activity, with no human intervention or bias,” she adds. They ranked slopes for all cognitive behaviors from the fMRI databases from negative to positive and found that they ordered from more tangible to highly abstract. An independent test of an additional 500 study participants supported the result.
She and colleagues found that cognitive function and abstract thought exist as a combination of many cortical sources ranging from those close to sensory cortices to far deeper from them along the brain connectome, or connection wiring diagram.
Generated by human-blind automated procedures, this diagram depicts an oversimplified graphical model of the information representation flow from sensory inputs (bottom) to abstract representations (top) in human cortex. Bottom layer of the pyramid included a sample representative description of the 20th percentile of behavioral elements closest to sensory inputs, the next layer up includes a sample description of behavioral elements from the 20–40th percentile…with the top layer containing a sample description of the behavioral elements distributed deepest in the cortical network, at the structural pinnacle of cognition. (credit: P. Taylor et al./Nature Scientific Reports)
The authors say their work demonstrates that all cognitive behaviors exist on a hierarchy, starting with the most tangible behaviors (such as finger tapping or pain), then to consciousness, and extending to the most abstract thoughts and activities such as naming. This hierarchy of abstraction is related to the connectome structure of the whole human brain — the connections between different regions of the brain — they add.
Creating a massively recurrent deep learning network
Siegelmann says this work will have great impact in computer science, especially in deep learning. “Deep learning is a computational system employing a multi-layered neural net, and is at the forefront of artificial intelligence (AI) learning algorithms,” she explains. “It bears similarity to the human brain in that higher layers are agglomerations of previous layers, and so provides more information in a single neuron.
“But the brain’s processing dynamic is far richer and less constrained because it has recurrent interconnection, sometimes called feedback loops. In current human-made deep learning networks that lack recurrent interconnections, a particular input cannot be related to other recent inputs, so they can’t be used for time-series prediction, control operations, or memory.”
Her lab is now creating a “massively recurrent deep learning network,” she says, for a more brain-like and superior learning AI, along with a new geometric data-science tool, which may find widespread use in other fields where massive data is difficult to view coherently due to data overlap.
New hope for biomarkers of brain disorders
Siegelmann believes this work will also have far-reaching effects for brain disorders. “Many brain disorders are implicated by non-standard processing or abnormal combination of sensory information,” she says. “Currently, many brain disorders lack a clear biological identifier, and are diagnosed by symptoms, such as confusion, memory loss and depression.
Our research suggests an entirely new method for analyzing brain abnormalities and is a source of new hope for developing biomarkers for more accurate and earlier diagnoses of psychiatric and neurological diseases.”
Siegelmann is director of the Biologically Inspired Neural and Dynamical Systems Laboratory at UMass Amherst and one of 16 recipients in 2015 of the National Science Foundation’s (NSF) Brain Research through Advancing Innovative Neurotechnologies (BRAIN) program initiated by President Obama to advance understanding of the brain. The work is supported by the U.S. Office of Naval Research.
Abstract of The global landscape of cognition: hierarchical aggregation as an organizational principle of human cortical networks and functions
Though widely hypothesized, limited evidence exists that human brain functions organize in global gradients of abstraction starting from sensory cortical inputs. Hierarchical representation is accepted in computational networks, and tentatively in visual neuroscience, yet no direct holistic demonstrations exist in vivo. Our methods developed network models enriched with tiered directionality, by including input locations, a critical feature for localizing representation in networks generally. Grouped primary sensory cortices defined network inputs, displaying global connectivity to fused inputs. Depth-oriented networks guided analyses of fMRI databases (~17,000 experiments;~1/4 of fMRI literature). Formally, we tested whether network depth predicted localization of abstract versus concrete behaviors over the whole set of studied brain regions. For our results, new cortical graph metrics, termednetwork-depth, ranked all databased cognitive function activations by network-depth. Thus, we objectively sorted stratified landscapes of cognition, starting from grouped sensory inputs in parallel, progressing deeper into cortex. This exposed escalating amalgamation of function or abstraction with increasing network-depth, globally. Nearly 500 new participants confirmed our results. In conclusion, data-driven analyses defined a hierarchically ordered connectome, revealing a related continuum of cognitive function. Progressive functional abstraction over network depth may be a fundamental feature of brains, and is observed in artificial networks.
