Do our brains use the same kind of deep-learning algorithms used in AI?

This is an illustration of a multi-compartment neural network model for deep learning. Left: Reconstruction of pyramidal neurons from mouse primary visual cortex, the most prevalent cell type in the cortex. The tree-like form separates “roots,” where bottoms of cortical neurons are located just where they need to be to receive signals about sensory input, from “branches” at the top, which are well placed to receive feedback error signals. Right: Illustration of simplified pyramidal neuron models. (credit: CIFAR)

Deep-learning researchers have found that certain neurons in the brain have shape and electrical properties that appear to be well-suited for “deep learning” — the kind of machine-intelligence used in beating humans at Go and Chess.

Canadian Institute For Advanced Research (CIFAR) Fellow Blake Richards and his colleagues — Jordan Guerguiev at the University of Toronto, Scarborough, and Timothy Lillicrap at Google DeepMind — developed an algorithm that simulates how a deep-learning network could work in our brains. It represents a biologically realistic way by which real brains could do deep learning.*

The finding is detailed in a study published December 5th in the open-access journal eLife. (The paper is highly technical; Adam Shai of Stanford University and Matthew E. Larkum of Humboldt University, Germany wrote a more accessible paper summarizing the ideas, published in the same eLife issue.)

Seeing the trees and the forest

Image of a neuron recorded in Blake Richard’s lab (credit: Blake Richards)

“Most of these neurons are shaped like trees, with ‘roots’ deep in the brain and ‘branches’ close to the surface,” says Richards. “What’s interesting is that these roots receive a different set of inputs than the branches that are way up at the top of the tree.” That allows these functions to have the required separation.

Using this knowledge of the neurons’ structure, the researchers built a computer model using the same shapes, with received signals in specific sections. It turns out that these sections allowed simulated neurons in different layers to collaborate — achieving deep learning.

“It’s just a set of simulations so it can’t tell us exactly what our brains are doing, but it does suggest enough to warrant further experimental examination if our own brains may use the same sort of algorithms that they use in AI,” Richards says.

“No one has tested our predictions yet,” he told KurzweilAI. “But, there’s a new preprint that builds on what we were proposing in a nice way from Walter Senn‘s group, and which includes some results on unsupervised learning (Yoshua [Bengio] mentions this work in his talk).

How the brain achieves deep learning

The tree-like pyramidal neocortex neurons are only one of many types of cells in the brain. Richards says future research should model different brain cells and examine how they interact together to achieve deep learning. In the long term, he hopes researchers can overcome major challenges, such as how to learn through experience without receiving feedback or to solve the “credit assignment problem.”**

Deep learning has brought about machines that can “see” the world more like humans can, and recognize language. But does the brain actually learn this way? The answer has the potential to create more powerful artificial intelligence and unlock the mysteries of human intelligence, he believes.

“What we might see in the next decade or so is a real virtuous cycle of research between neuroscience and AI, where neuroscience discoveries help us to develop new AI and AI can help us interpret and understand our experimental data in neuroscience,” Richards says.

Perhaps this kind of research could one day also address future ethical and other human-machine-collaboration issues — including merger, as Elon Musk and Ray Kurzweil have proposed, to achieve a “soft takeoff” in the emergence of superintelligence.

* This research idea goes back to AI pioneers Geoffrey Hinton, a CIFAR Distinguished Fellow and founder of the Learning in Machines & Brains program, and program Co-Director Yoshua Bengio, who was one of the main motivations for founding the program. These researchers sought not only to develop artificial intelligence, but also to understand how the human brain learns, says Richards.

In the early 2000s, Richards and Lillicrap took a course with Hinton at the University of Toronto and were convinced deep learning models were capturing “something real” about how human brains work. At the time, there were several challenges to testing that idea. Firstly, it wasn’t clear that deep learning could achieve human-level skill. Secondly, the algorithms violated biological facts proven by neuroscientists.

The paper builds on research from Bengio’s lab on a more biologically plausible way to train neural nets and an algorithm developed by Lillicrap that further relaxes some of the rules for training neural nets. The paper also incorporates research from Matthew Larkam on the structure of neurons in the neocortex.

By combining neurological insights with existing algorithms, Richards’ team was able to create a better and more realistic algorithm for simulating learning in the brain.

The study was funded by the Natural Sciences and Engineering Research Council of Canada (NSERC), a Google Faculty Research Award, and CIFAR.

** In the paper, the authors note that a large gap exists between deep learning in AI and our current understanding of learning and memory in neuroscience. “In particular, unlike deep learning researchers, neuroscientists do not yet have a solution to the ‘credit assignment problem’ (Rumelhart et al., 1986; Lillicrap et al., 2016; Bengio et al., 2015). Learning to optimize some behavioral or cognitive function requires a method for assigning ‘credit’ (or ‘blame’) to neurons for their contribution to the final behavioral output (LeCun et al., 2015; Bengio et al., 2015). The credit assignment problem refers to the fact that assigning credit in multi-layer networks is difficult, since the behavioral impact of neurons in early layers of a network depends on the downstream synaptic connections.” The authors go on to suggest a solution.


How to train a robot to do complex abstract thinking

Robot inspects cooler, ponders next step (credit: Intelligent Robot Lab / Brown University)

Robots are great at following programmed steps. But asking a robot to “move the green bottle from the cooler to the cupboard” would require it to have abstract representations of these things and actions, plus knowledge of its surroundings.

(“Hmm, which of those millions of pixels is a ‘cooler,’ whatever than means? How do I get inside it and also the ‘cupboard’? …”)

To help robots answer these kinds of questions and plan complex multi-step tasks, robots can construct two kinds of abstract representations of the world around them, say Brown University and MIT researchers:

  • “Procedural abstractions”: bundling all the low-level movements composed into higher-level skills (such as opening a door). Most of those robots doing fancy athletic tricks are explicitly programmed with such procedural abstractions, say the researchers.
  • “Perceptual abstractions”: making sense out of the millions of confusing pixels in the real world.

Building truly intelligent robots

According to George Konidaris, Ph.D., an assistant professor of computer science at Brown and the lead author of the new study, there’s been less progress in perceptual abstraction — the focus of the new research.

To explore this, the researchers trained a robot they called “Anathema” (aka “Ana”). They started by teaching Ana “procedural abstractions” in a room containing a cupboard, a cooler, a switch that controls a light inside the cupboard, and a bottle that could be left in either the cooler or the cupboard. They gave Ana a set of high-level motor skills for manipulating the objects in the room, such as opening and closing both the cooler and the cupboard, flipping the switch, and picking up a bottle.

Ana was also able to learn a very abstract description of the visual environment that contained only what was necessary for her to be able to perform a particular skill. Once armed with these learned abstract procedures and perceptions, the researchers gave Ana a challenge: “Take the bottle from the cooler and put it in the cupboard.”

Ana’s dynamic concept of a “cooler,” based on configurations of pixels in open and closed positions. (credit: Intelligent Robot Lab / Brown University)

Accepting the challenge, Ana navigated to the cooler. She had learned the configuration of pixels in her visual field associated with the cooler lid being closed (the only way to open it). She had also learned how to open it: stand in front of it and don’t do anything (because she needed both hands to open the lid).

She opened the cooler and sighted the bottle. But she didn’t pick it up. Not yet.

She realized that if she had the bottle in her gripper, she wouldn’t be able to open the cupboard — that requires both hands. Instead, she went directly to the cupboard.

There, she saw that the light switch was in the “on” position, and instantly realized that opening the cupboard would block the switch. So she turned the switch off before opening the cupboard. Finally, she returned to the cooler, retrieved the bottle, and placed it in the cupboard.

She had developed the entire plan in about four milliseconds.

“She learned these abstractions on her own”

Once a robot has high-level motor skills, it can automatically construct a compatible high-level symbolic representation of the world by making sense of its pixelated surroundings, according to Konidaris. “We didn’t provide Ana with any of the abstract representations she needed to plan for the task,” he said. “She learned those abstractions on her own, and once she had them, planning was easy.”

Her entire knowledge and skill set was represented in a text file just 126 lines long.

Konidaris says the research provides an important theoretical building block for applying artificial intelligence to robotics. “We believe that allowing our robots to plan and learn in the abstract rather than the concrete will be fundamental to building truly intelligent robots,” he said. “Many problems are often quite simple, if you think about them in the right way.”

Source: Journal of Artificial Intelligence Research (open-access). Funded by DARPA and MIT’s Intelligence Initiative.

IRL Lab | Learning Symbolic Representations for High-Level Robot Planning

AI algorithm with ‘social skills’ teaches humans how to collaborate

(credit: Iyad Rahwan)

An international team has developed an AI algorithm with social skills that has outperformed humans in the ability to cooperate with people and machines in playing a variety of two-player games.

The researchers, led by Iyad Rahwan, PhD, an MIT Associate Professor of Media Arts and Sciences, tested humans and the algorithm, called S# (“S sharp”), in three types of interactions: machine-machine, human-machine, and human-human. In most instances, machines programmed with S# outperformed humans in finding compromises that benefit both parties.

“Two humans, if they were honest with each other and loyal, would have done as well as two machines,” said lead author BYU computer science professor Jacob Crandall. “As it is, about half of the humans lied at some point. So essentially, this particular algorithm is learning that moral characteristics are better [since it’s programmed to not lie] and it also learns to maintain cooperation once it emerges.”

“The end goal is that we understand the mathematics behind cooperation with people and what attributes artificial intelligence needs to develop social skills,” said Crandall. “AI needs to be able to respond to us and articulate what it’s doing. It has to be able to interact with other people.”

How casual talk by AI helps humans be more cooperative

One important finding: colloquial phrases (called “cheap talk” in the study) doubled the amount of cooperation. In tests, if human participants cooperated with the machine, the machine might respond with a “Sweet. We are getting rich!” or “I accept your last proposal.” If the participants tried to betray the machine or back out of a deal with them, they might be met with a trash-talking “Curse you!”, “You will pay for that!” or even an “In your face!”

And when machines used cheap talk, their human counterparts were often unable to tell whether they were playing a human or machine — a sort of mini “Turing test.”

The research findings, Crandall hopes, could have long-term implications for human relationships. “In society, relationships break down all the time,” he said. “People that were friends for years all of a sudden become enemies. Because the machine is often actually better at reaching these compromises than we are, it can potentially teach us how to do this better.”

The research is described in an open-access paper in Nature Communications.

A human-machine collaborative chatbot system 

An actual conversation on Evorus, combining multiple chatbots and workers. (credit: T. Huang et al.)

In a related study, Carnegie Mellon University (CMU) researchers have created a new collaborative chatbot called Evorus that goes beyond Siri, Alexa, and Cortana by adding humans in the loop.

Evorus combines a chatbot called Chorus with inputs by paid crowd workers at Amazon Mechanical Turk, who answer questions from users and vote on the best answer. Evorus keeps track of the questions asked and answered and, over time, begins to suggest these answers for subsequent questions. It can also use multiple chatbots, such as vote bots, Yelp Bot (restaurants) and Weather Bot to provide enhanced information.

Humans are simultaneously training the system’s AI, making it gradually less dependent on people, says Jeff Bigham, associate professor in the CMU Human-Computer Interaction Institute.

The hope is that as the system grows, the AI will be able to handle an increasing percentage of questions, while the number of crowd workers necessary to respond to “long tail” questions will remain relatively constant.

Keeping humans in the loop also reduces the risk that malicious users will manipulate the conversational agent inappropriately, as occurred when Microsoft briefly deployed its Tay chatbot in 2016, noted co-developer Ting-Hao Huang, a Ph.D. student in the Language Technologies Institute (LTI).

The preliminary system is available for download and use by anyone willing to be part of the research effort. It is deployed via Google Hangouts, which allows for voice input as well as access from computers, phones, and smartwatches. The software architecture can also accept automated question-answering components developed by third parties.

A open-access research paper on Evorus, available online, will be presented at CHI 2018, the Conference on Human Factors in Computing Systems in Montreal, April 21–26, 2018.

Abstract of Cooperating with machines

Since Alan Turing envisioned artificial intelligence, technical progress has often been measured by the ability to defeat humans in zero-sum encounters (e.g., Chess, Poker, or Go). Less attention has been given to scenarios in which human–machine cooperation is beneficial but non-trivial, such as scenarios in which human and machine preferences are neither fully aligned nor fully in conflict. Cooperation does not require sheer computational power, but instead is facilitated by intuition, cultural norms, emotions, signals, and pre-evolved dispositions. Here, we develop an algorithm that combines a state-of-the-art reinforcement-learning algorithm with mechanisms for signaling. We show that this algorithm can cooperate with people and other algorithms at levels that rival human cooperation in a variety of two-player repeated stochastic games. These results indicate that general human–machine cooperation is achievable using a non-trivial, but ultimately simple, set of algorithmic mechanisms.

Abstract of A Crowd-powered Conversational Assistant Built to Automate Itself Over Time

Crowd-powered conversational assistants have been shown to be more robust than automated systems, but do so at the cost of higher response latency and monetary costs. A promising direction is to combine the two approaches for high quality, low latency, and low cost solutions. In this paper, we introduce Evorus, a crowd-powered conversational assistant built to automate itself over time by (i) allowing new chatbots to be easily integrated to automate more scenarios, (ii) reusing prior crowd answers, and (iii) learning to automatically approve response candidates. Our 5-month-long deployment with 80 participants and 281 conversations shows that Evorus can automate itself without compromising conversation quality. Crowd-AI architectures have long been proposed as a way to reduce cost and latency for crowd-powered systems; Evorus demonstrates how automation can be introduced successfully in a deployed system. Its architecture allows future researchers to make further innovation on the underlying automated components in the context of a deployed open domain dialog system.

Ultra-thin ‘atomistor’ synapse-like memory storage device paves way for faster, smaller, smarter computer chips

Illustration of single-atom-layer “atomristors” — the thinnest-ever memory-storage device (credit: Cockrell School of Engineering, The University of Texas at Austin)

A team of electrical engineers at The University of Texas at Austin and scientists at Peking University has developed a one-atom-thick 2D “atomristor” memory storage device that may lead to faster, smaller, smarter computer chips.

The atomristor (atomic memristor) improves upon memristor (memory resistor) memory storage technology by using atomically thin nanomaterials (atomic sheets). (Combining memory and logic functions, similar to the synapses of biological brains, memristors “remember” their previous state after being turned off.)

Schematic of atomristor memory sandwich based on molybdenum sulfide (MoS2) in a form of a single-layer atomic sheet grown on gold foil. (Blue: Mo; yellow: S) (credit: Ruijing Ge et al./Nano Letters)

Memory storage and transistors have, to date, been separate components on a microchip. Atomristors combine both functions on a single, more-efficient device. They use metallic atomic sheets (such as graphene or gold) as electrodes and semiconducting atomic sheets (such as molybdenum sulfide) as the active layer. The entire memory cell is a two-layer sandwich only ~1.5 nanometers thick.

“The sheer density of memory storage that can be made possible by layering these synthetic atomic sheets onto each other, coupled with integrated transistor design, means we can potentially make computers that learn and remember the same way our brains do,” said Deji Akinwande, associate professor in the Cockrell School of Engineering’s Department of Electrical and Computer Engineering.

“This discovery has real commercialization value, as it won’t disrupt existing technologies,” Akinwande said. “Rather, it has been designed to complement and integrate with the silicon chips already in use in modern tech devices.”

The research is described in an open-access paper in the January American Chemical Society journal Nano Letters.

Longer battery life in cell phones

For nonvolatile operation (preserving data after power is turned off), the new design also “offers a substantial advantage over conventional flash memory, which occupies far larger space. In addition, the thinness allows for faster and more efficient electric current flow,” the researchers note in the paper.

The research team also discovered another unique application for the atomristor technology: Atomristors are the smallest radio-frequency (RF) memory switches to be demonstrated, with no DC battery consumption, which could ultimately lead to longer battery life for cell phones and other battery-powered devices.*

Funding for the UT Austin team’s work was provided by the National Science Foundation and the Presidential Early Career Award for Scientists and Engineers, awarded to Akinwande in 2015.

* “Contemporary switches are realized with transistor or microelectromechanical devices, both of which are volatile, with the latter also requiring large switching voltages [which are not ideal] for mobile technologies,” the researchers note in the paper. Atomristors instead allow for nonvolatile low-power radio-frequency (RF) switches with “low voltage operation, small form-factor, fast switching speed, and low-temperature integration compatible with silicon or flexible substrates.”

Abstract of Atomristor: Nonvolatile Resistance Switching in Atomic Sheets of Transition Metal Dichalcogenides

Recently, two-dimensional (2D) atomic sheets have inspired new ideas in nanoscience including topologically protected charge transport,1,2 spatially separated excitons,3 and strongly anisotropic heat transport.4 Here, we report the intriguing observation of stable nonvolatile resistance switching (NVRS) in single-layer atomic sheets sandwiched between metal electrodes. NVRS is observed in the prototypical semiconducting (MX2, M = Mo, W; and X = S, Se) transitional metal dichalcogenides (TMDs),5 which alludes to the universality of this phenomenon in TMD monolayers and offers forming-free switching. This observation of NVRS phenomenon, widely attributed to ionic diffusion, filament, and interfacial redox in bulk oxides and electrolytes,6−9 inspires new studies on defects, ion transport, and energetics at the sharp interfaces between atomically thin sheets and conducting electrodes. Our findings overturn the contemporary thinking that nonvolatile switching is not scalable to subnanometre owing to leakage currents.10 Emerging device concepts in nonvolatile flexible memory fabrics, and brain-inspired (neuromorphic) computing could benefit substantially from the wide 2D materials design space. A new major application, zero-static power radio frequency (RF) switching, is demonstrated with a monolayer switch operating to 50 GHz.

An artificial synapse for future miniaturized portable ‘brain-on-a-chip’ devices

Biological synapse structure (credit: Thomas Splettstoesser/CC)

MIT engineers have designed a new artificial synapse made from silicon germanium that can precisely control the strength of an electric current flowing across it.

In simulations, the researchers found that the chip and its synapses could be used to recognize samples of handwriting with 95 percent accuracy. The engineers say the new design, published today (Jan. 22) in the journal Nature Materials, is a major step toward building portable, low-power neuromorphic chips for use in pattern recognition and other machine-learning tasks.

Controlling the flow of ions: the challenge

Researchers in the emerging field of “neuromorphic computing” have attempted to design computer chips that work like the human brain. The idea is to apply a voltage across layers that would cause ions (electrically charged atoms) to move in a switching medium (synapse-like space) to create conductive filaments in a manner that’s similar to how the “weight” (connection strength) of a synapse changes.

There are more than 100 trillion synapses (in a typical human brain) that mediate neuron signaling in the brain, strengthening some neural connections while pruning (weakening) others — a process that enables the brain to recognize patterns, remember facts, and carry out other learning tasks, all at lightning speeds.

Instead of carrying out computations based on binary, on/off signaling, like current digital chips, the elements of a “brain on a chip” would work in an analog fashion, exchanging a gradient of signals, or “weights” — much like neurons that activate in various ways (depending on the type and number of ions that flow across a synapse).

But it’s been difficult to control the flow of ions in existing synapse designs. These have multiple paths that make it difficult to predict where ions will make it through, according to research team leader Jeehwan Kim, PhD, an assistant professor in the departments of Mechanical Engineering and Materials Science and Engineering, a principal investigator in MIT’s Research Laboratory of Electronics and Microsystems Technology Laboratories.

“Once you apply some voltage to represent some data with your artificial neuron, you have to erase and be able to write it again in the exact same way,” Kim says. “But in an amorphous solid, when you write again, the ions go in different directions because there are lots of defects. This stream is changing, and it’s hard to control. That’s the biggest problem — nonuniformity of the artificial synapse.”

Epitaxial random access memory (epiRAM)

(Left) Cross-sectional transmission electron microscope image of 60 nm silicon-germanium (SiGe) crystal grown on a silicon substrate (diagonal white lines represent candidate dislocations). Scale bar: 25 nm. (Right) Cross-sectional scanning electron microscope image of an epiRAM device with titanium (Ti)–gold (Au) and silver (Ag)–palladium (Pd) layers. Scale bar: 100 nm. (credit: Shinhyun Choi et al./Nature Materials)

So instead of using amorphous materials as an artificial synapse, Kim and his colleagues created an new “epitaxial random access memory” (epiRAM) design.

They started with a wafer of silicon. They then grew a similar pattern of silicon germanium — a material used commonly in transistors — on top of the silicon wafer. Silicon germanium’s lattice is slightly larger than that of silicon, and Kim found that together, the two perfectly mismatched materials could form a funnel-like dislocation, creating a single path through which ions can predictably flow.*

This is the most uniform device we could achieve, which is the key to demonstrating artificial neural networks,” Kim says.

Testing the ability to recognize samples of handwriting

As a test, Kim and his team explored how the epiRAM device would perform if it were to carry out an actual learning task: recognizing samples of handwriting — which researchers consider to be a practical test for neuromorphic chips. Such chips would consist of artificial “neurons” connected to other “neurons” via filament-based artificial “synapses.”

Image-recognition simulation. (Left) A 3-layer multilayer-perception neural network with black and white input signal for each layer in algorithm level. The inner product (summation) of input neuron signal vector and first synapse array vector is transferred after activation and binarization as input vectors of second synapse arrays. (Right) Circuit block diagram of hardware implementation showing a synapse layer composed of epiRAM crossbar arrays and the peripheral circuit. (credit: Shinhyun Choi et al./Nature Materials)

They ran a computer simulation of an artificial neural network consisting of three sheets of neural layers connected via two layers of artificial synapses, based on measurements from their actual neuromorphic chip. They fed into their simulation tens of thousands of samples from the MNIST handwritten recognition dataset**, commonly used by neuromorphic designers.

They found that their neural network device recognized handwritten samples 95.1 percent of the time — close to the 97 percent accuracy of existing software algorithms running on large computers.

A chip to replace a supercomputer

The team is now in the process of fabricating a real working neuromorphic chip that can carry out handwriting-recognition tasks. Looking beyond handwriting, Kim says the team’s artificial synapse design will enable much smaller, portable neural network devices that can perform complex computations that are currently only possible with large supercomputers.

“Ultimately, we want a chip as big as a fingernail to replace one big supercomputer,” Kim says. “This opens a stepping stone to produce real artificial intelligence hardware.”

This research was supported in part by the National Science Foundation. Co-authors included researchers at Arizona State University.

* They applied voltage to each synapse and found that all synapses exhibited about the same current, or flow of ions, with about a 4 percent variation between synapses — a much more uniform performance compared with synapses made from amorphous material. They also tested a single synapse over multiple trials, applying the same voltage over 700 cycles, and found the synapse exhibited the same current, with just 1 percent variation from cycle to cycle.

** The MNIST (Modified National Institute of Standards and Technology database) is a large database of handwritten digits that is commonly used for training various image processing systems and for training and testing in the field of machine learning. It contains 60,000 training images and 10,000 testing images. 

Abstract of SiGe epitaxial memory for neuromorphic computing with reproducible high performance based on engineered dislocations

Although several types of architecture combining memory cells and transistors have been used to demonstrate artificial synaptic arrays, they usually present limited scalability and high power consumption. Transistor-free analog switching devices may overcome these limitations, yet the typical switching process they rely on—formation of filaments in an amorphous medium—is not easily controlled and hence hampers the spatial and temporal reproducibility of the performance. Here, we demonstrate analog resistive switching devices that possess desired characteristics for neuromorphic computing networks with minimal performance variations using a single-crystalline SiGe layer epitaxially grown on Si as a switching medium. Such epitaxial random access memories utilize threading dislocations in SiGe to confine metal filaments in a defined, one-dimensional channel. This confinement results in drastically enhanced switching uniformity and long retention/high endurance with a high analog on/off ratio. Simulations using the MNIST handwritten recognition data set prove that epitaxial random access memories can operate with an online learning accuracy of 95.1%.

Amazon’s store of the future opens

(credit: Amazon)

Amazon’s first Amazon Go store opened today in Seattle, automating most of the purchase, checkout, and payment steps associated with a retail transaction and replacing cash registers, cashiers, credit cards, self-checkout kiosks, RFID chips — and lines — with hundreds of small cameras, computer vision, deep-learning algorithms, and sensor fusion.

Just walk in (as long as you have the Amazon Go app and an account), scan a QR code at the turnstile, grab, and go.

Meanwhile, the shutdown of the dysfunctional U.S. government continues.* Hmm, what if we created Government Go?

If you visit the store (2131 7th Ave — 7 a.m. to 9 p.m. PT Monday to Friday), let us know about your experience and thoughts in the comments below.

* January 22 at 6:11 PM EST: House votes to end government shutdown, sending legislation to Trump — Washington Post

Deep neural network models score higher than humans in reading and comprehension test

(credit: Alibaba Group)

Microsoft has developed a deep neural network that scored higher than humans on exact scores in a Stanford University reading and comprehension test Stanford Question Answering Dataset (SQuAD).

Microsoft achieved 82.650 on Jan. 3; Alibaba Group Holding Ltd. came in at second place at 82.440 on Jan. 5. The best human score so far is 82.304.

“SQuAD is a new reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage,” according to the Stanford NLP Group. “With 100,000+ question-answer pairs on 500+ articles, SQuAD is significantly larger than previous reading comprehension datasets.”

“The Chinese e-commerce titan has joined the likes of Tencent Holdings Ltd. and Baidu Inc. in a race to develop AI that can enrich social media feeds, target ads and services or even aid in autonomous driving, Bloomberg notes. “Beijing has endorsed the technology in a national-level plan that calls for the country to become the industry leader 2030.”

Read more: China’s Plan for World Domination in AI (Bloomberg)

Deep neural network models score higher than humans in reading and comprehension test

(credit: Alibaba Group)

Microsoft and Alibaba have developed deep neural network models that scored higher than humans in a Stanford University reading and comprehension test, Stanford Question Answering Dataset (SQuAD).

Microsoft achieved 82.650 on the ExactMatch (EM) metric* on Jan. 3, and Alibaba Group Holding Ltd. scored 82.440 on Jan. 5. The best human score so far is 82.304.

“SQuAD is a new reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles, where the answer to every question is a segment of text, or span, from the corresponding reading passage,” according to the Stanford NLP Group. “With 100,000+ question-answer pairs on 500+ articles, SQuAD is significantly larger than previous reading comprehension datasets.”

“A strong start to 2018 with the first model (SLQA+) to exceed human-level performance on @stanfordnlp SQuAD’s EM metric!,” said Pranav Rajpurkar, a Ph.D. student in the Stanford Machine Learning Group and lead author of a paper in Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing on SQuAD (available on open-access ArXiv). “Next challenge: the F1 metric*, where humans still lead by ~2.5 points!” (Alibaba’s SLQA+ scored 88.607 on the F1 metric and Microsoft’s r-net+ scored 88.493.)

However, challenging the “comprehension” description, Gary Marcus, PhD, a Professor of Psychology and Neural Science at NYU, notes in a tweet that “the SQUAD test shows that machines can highlight relevant passages in text, not that they understand those passages.”

“The Chinese e-commerce titan has joined the likes of Tencent Holdings Ltd. and Baidu Inc. in a race to develop AI that can enrich social media feeds, target ads and services or even aid in autonomous driving, Bloomberg notes. “Beijing has endorsed the technology in a national-level plan that calls for the country to become the industry leader 2030.”

Read more: China’s Plan for World Domination in AI (Bloomberg)

*”The ExactMatch metric measures the percentage of predictions that match any one of the ground truth answers exactly. The F1 score metric measures the average overlap between the prediction and ground truth answer.” – Pranav Rajpurkar et al., ArXiv

How to grow functioning human muscles from stem cells

A cross section of a muscle fiber grown from induced pluripotent stem cells, showing muscle cells (green), cell nuclei (blue), and the surrounding support matrix for the cells (credit: Duke University)

Biomedical engineers at Duke University have grown the first functioning human skeletal muscle from human induced pluripotent stem cells (iPSCs). (Pluripotent stem cells are important in regenerative medicine because they can generate any type of cell in the body and can propagate indefinitely; the induced version can be generated from adult cells instead of embryos.)

The engineers say the new technique is promising for cellular therapies, drug discovery, and studying rare diseases. “When a child’s muscles are already withering away from something like Duchenne muscular dystrophy, it would not be ethical to take muscle samples from them and do further damage,” explained Nenad Bursac, professor of biomedical engineering at Duke University and senior author of an open-access paper on the research published Tuesday, January 9, in Nature Communications.

How to grow a muscle

In the study, the researchers started with human induced pluripotent stem cells. These are cells taken from adult non-muscle tissues, such as skin or blood, and reprogrammed to revert to a primordial state. The pluripotent stem cells are then grown while being flooded with a molecule called Pax7 — which signals the cells to start becoming muscle.

After two to four weeks of 3-D culture, the resulting muscle cells form muscle fibers that contract and react to external stimuli such as electrical pulses and biochemical signals — mimicking neuronal inputs just like native muscle tissue. The researchers also implanted the newly grown muscle fibers into adult mice. The muscles survived and functions for at least three weeks, while progressively integrating into the native tissue through vascularization (growing blood vessels).

A stained cross section of the new muscle fibers, showing muscle cells (red), receptors for neuronal input (green), and cell nuclei (blue) (credit: Duke University)

Once the cells were well on their way to becoming muscle, the researchers stopped providing the Pax7 signaling molecule and started giving the cells the support and nourishment they needed to fully mature. (At this point in the research, the resulting muscle is not as strong as native muscle tissue, and also falls short of the muscle grown in a previous study*, which started from muscle biopsies.)

However, the pluripotent stem cell-derived muscle fibers develop reservoirs of “satellite-like cells” that are necessary for normal adult muscles to repair damage, while the muscle from the previous study had much fewer of these cells. The stem cell method is also capable of growing many more cells from a smaller starting batch than the previous biopsy method.

“With this technique, we can just take a small sample of non-muscle tissue, like skin or blood, revert the obtained cells to a pluripotent state, and eventually grow an endless amount of functioning muscle fibers to test,” said Bursac.

The researchers could also, in theory, fix genetic malfunctions in the induced pluripotent stem cells derived from a patient, he added. Then they could grow small patches of completely healthy muscle. This could not heal or replace an entire body’s worth of diseased muscle, but it could be used in tandem with more widely targeted genetic therapies or to heal more localized problems.

The researchers are now refining their technique to grow more robust muscles and beginning work to develop new models of rare muscle diseases. This work was supported by the National Institutes of Health.

Duke Engineering | Human Muscle Grown from Skin Cells

Muscles for future microscale robot exoskeletons

Meanwhile, physicists at Cornell University are exploring ways to create muscles for future microscale robot exoskeletons — rapidly changing their shape upon sensing chemical or thermal changes in their environment. The new designs are compatible with semiconductor manufacturing, making them useful for future microscale robotics.

The microscale robot exoskeleton muscles move using a motor called a bimorph. (A bimorph is an assembly of two materials — in this case, graphene and glass — that bends when driven by a stimulus like heat, a chemical reaction or an applied voltage.) The shape change happens because, in the case of heat, two materials with different thermal responses expand by different amounts over the same temperature change. The bimorph bends to relieve some of this strain, allowing one layer to stretch out longer than the other. By adding rigid flat panels that cannot be bent by bimorphs, the researchers localize bending to take place only in specific places, creating folds. With this concept, they are able to make a variety of folding structures ranging from tetrahedra (triangular pyramids) to cubes. The bimorphs also fold in response to chemical stimuli by driving large ions into the glass, causing it to expand. (credit: Marc Z. Miskin et al./PNAS)

Their work is outlined in a paper published Jan. 2 in Proceedings of the National Academy of Sciences.

* The advance builds on work published in 2015, when the Duke engineers grew the first functioning human muscle tissue from cells obtained from muscle biopsies. In that research, Bursac and his team started with small samples of human cells obtained from muscle biopsies, called “myoblasts,” that had already progressed beyond the stem cell stage but hadn’t yet become mature muscle fibers. The engineers grew these myoblasts by many folds and then put them into a supportive 3-D scaffolding filled with a nourishing gel that allowed them to form aligned and functioning human muscle fibers.

Abstract of Engineering human pluripotent stem cells into a functional skeletal muscle tissue

The generation of functional skeletal muscle tissues from human pluripotent stem cells (hPSCs) has not been reported. Here, we derive induced myogenic progenitor cells (iMPCs) via transient overexpression of Pax7 in paraxial mesoderm cells differentiated from hPSCs. In 2D culture, iMPCs readily differentiate into spontaneously contracting multinucleated myotubes and a pool of satellite-like cells endogenously expressing Pax7. Under optimized 3D culture conditions, iMPCs derived from multiple hPSC lines reproducibly form functional skeletal muscle tissues (iSKM bundles) containing aligned multi-nucleated myotubes that exhibit positive force–frequency relationship and robust calcium transients in response to electrical or acetylcholine stimulation. During 1-month culture, the iSKM bundles undergo increased structural and molecular maturation, hypertrophy, and force generation. When implanted into dorsal window chamber or hindlimb muscle in immunocompromised mice, the iSKM bundles survive, progressively vascularize, and maintain functionality. iSKM bundles hold promise as a microphysiological platform for human muscle disease modeling and drug development.

Abstract of Graphene-based bimorphs for micron-sized, autonomous origami machines

Origami-inspired fabrication presents an attractive platform for miniaturizing machines: thinner layers of folding material lead to smaller devices, provided that key functional aspects, such as conductivity, stiffness, and flexibility, are persevered. Here, we show origami fabrication at its ultimate limit by using 2D atomic membranes as a folding material. As a prototype, we bond graphene sheets to nanometer-thick layers of glass to make ultrathin bimorph actuators that bend to micrometer radii of curvature in response to small strain differentials. These strains are two orders of magnitude lower than the fracture threshold for the device, thus maintaining conductivity across the structure. By patterning 2-<mml:math><mml:mi>

Will artificial intelligence become conscious?

(Credit: EPFL/Blue Brain Project)

By Subhash Kak, Regents Professor of Electrical and Computer Engineering, Oklahoma State University

Forget about today’s modest incremental advances in artificial intelligence, such as the increasing abilities of cars to drive themselves. Waiting in the wings might be a groundbreaking development: a machine that is aware of itself and its surroundings, and that could take in and process massive amounts of data in real time. It could be sent on dangerous missions, into space or combat. In addition to driving people around, it might be able to cook, clean, do laundry — and even keep humans company when other people aren’t nearby.

A particularly advanced set of machines could replace humans at literally all jobs. That would save humanity from workaday drudgery, but it would also shake many societal foundations. A life of no work and only play may turn out to be a dystopia.

Conscious machines would also raise troubling legal and ethical problems. Would a conscious machine be a “person” under law and be liable if its actions hurt someone, or if something goes wrong? To think of a more frightening scenario, might these machines rebel against humans and wish to eliminate us altogether? If yes, they represent the culmination of evolution.

As a professor of electrical engineering and computer science who works in machine learning and quantum theory, I can say that researchers are divided on whether these sorts of hyperaware machines will ever exist. There’s also debate about whether machines could or should be called “conscious” in the way we think of humans, and even some animals, as conscious. Some of the questions have to do with technology; others have to do with what consciousness actually is.

Is awareness enough?

Most computer scientists think that consciousness is a characteristic that will emerge as technology develops. Some believe that consciousness involves accepting new information, storing and retrieving old information and cognitive processing of it all into perceptions and actions. If that’s right, then one day machines will indeed be the ultimate consciousness. They’ll be able to gather more information than a human, store more than many libraries, access vast databases in milliseconds and compute all of it into decisions more complex, and yet more logical, than any person ever could.

On the other hand, there are physicists and philosophers who say there’s something more about human behavior that cannot be computed by a machine. Creativity, for example, and the sense of freedom people possess don’t appear to come from logic or calculations.

Yet these are not the only views of what consciousness is, or whether machines could ever achieve it.

Quantum views

Another viewpoint on consciousness comes from quantum theory, which is the deepest theory of physics. According to the orthodox Copenhagen Interpretation, consciousness and the physical world are complementary aspects of the same reality. When a person observes, or experiments on, some aspect of the physical world, that person’s conscious interaction causes discernible change. Since it takes consciousness as a given and no attempt is made to derive it from physics, the Copenhagen Interpretation may be called the “big-C” view of consciousness, where it is a thing that exists by itself – although it requires brains to become real. This view was popular with the pioneers of quantum theory such as Niels Bohr, Werner Heisenberg and Erwin Schrödinger.

The interaction between consciousness and matter leads to paradoxes that remain unresolved after 80 years of debate. A well-known example of this is the paradox of Schrödinger’s cat, in which a cat is placed in a situation that results in it being equally likely to survive or die – and the act of observation itself is what makes the outcome certain.

The opposing view is that consciousness emerges from biology, just as biology itself emerges from chemistry which, in turn, emerges from physics. We call this less expansive concept of consciousness “little-C.” It agrees with the neuroscientists’ view that the processes of the mind are identical to states and processes of the brain. It also agrees with a more recent interpretation of quantum theory motivated by an attempt to rid it of paradoxes, the Many Worlds Interpretation, in which observers are a part of the mathematics of physics.

Philosophers of science believe that these modern quantum physics views of consciousness have parallels in ancient philosophy. Big-C is like the theory of mind in Vedanta – in which consciousness is the fundamental basis of reality, on par with the physical universe.

Little-C, in contrast, is quite similar to Buddhism. Although the Buddha chose not to address the question of the nature of consciousness, his followers declared that mind and consciousness arise out of emptiness or nothingness.

Big-C and scientific discovery

Scientists are also exploring whether consciousness is always a computational process. Some scholars have argued that the creative moment is not at the end of a deliberate computation. For instance, dreams or visions are supposed to have inspired Elias Howe‘s 1845 design of the modern sewing machine, and August Kekulé’s discovery of the structure of benzene in 1862.

A dramatic piece of evidence in favor of big-C consciousness existing all on its own is the life of self-taught Indian mathematician Srinivasa Ramanujan, who died in 1920 at the age of 32. His notebook, which was lost and forgotten for about 50 years and published only in 1988, contains several thousand formulas, without proof in different areas of mathematics, that were well ahead of their time. Furthermore, the methods by which he found the formulas remain elusive. He himself claimed that they were revealed to him by a goddess while he was asleep.

The concept of big-C consciousness raises the questions of how it is related to matter, and how matter and mind mutually influence each other. Consciousness alone cannot make physical changes to the world, but perhaps it can change the probabilities in the evolution of quantum processes. The act of observation can freeze and even influence atoms’ movements, as Cornell physicists proved in 2015. This may very well be an explanation of how matter and mind interact.

Mind and self-organizing systems

It is possible that the phenomenon of consciousness requires a self-organizing system, like the brain’s physical structure. If so, then current machines will come up short.

Scholars don’t know if adaptive self-organizing machines can be designed to be as sophisticated as the human brain; we lack a mathematical theory of computation for systems like that. Perhaps it’s true that only biological machines can be sufficiently creative and flexible. But then that suggests people should – or soon will – start working on engineering new biological structures that are, or could become, conscious.

Reprinted with permission from The Conversation