Tag Archives: neural network

Automated science writing?

It seems that automated science writing is not ready—yet. Still, an April 18, 2019 news item on ScienceDaily suggests that progress is being made,

The work of a science writer, including this one, includes reading journal papers filled with specialized technical terminology, and figuring out how to explain their contents in language that readers without a scientific background can understand.

Now, a team of scientists at MIT [Massachusetts Institute of Technology] and elsewhere has developed a neural network, a form of artificial intelligence (AI), that can do much the same thing, at least to a limited extent: It can read scientific papers and render a plain-English summary in a sentence or two.

An April 17, 2019 MIT news release, which originated the news item, delves into the research and its implications,

Even in this limited form, such a neural network could be useful for helping editors, writers, and scientists [emphasis mine] scan a large number of papers to get a preliminary sense of what they’re about. But the approach the team developed could also find applications in a variety of other areas besides language processing, including machine translation and speech recognition.

The work is described in the journal Transactions of the Association for Computational Linguistics, in a paper by Rumen Dangovski and Li Jing, both MIT graduate students; Marin Soljačić, a professor of physics at MIT; Preslav Nakov, a principal scientist at the Qatar Computing Research Institute, HBKU; and Mićo Tatalović, a former Knight Science Journalism fellow at MIT and a former editor at New Scientist magazine.

From AI for physics to natural language

The work came about as a result of an unrelated project, which involved developing new artificial intelligence approaches based on neural networks, aimed at tackling certain thorny problems in physics. However, the researchers soon realized that the same approach could be used to address other difficult computational problems, including natural language processing, in ways that might outperform existing neural network systems.

“We have been doing various kinds of work in AI for a few years now,” Soljačić says. “We use AI to help with our research, basically to do physics better. And as we got to be  more familiar with AI, we would notice that every once in a while there is an opportunity to add to the field of AI because of something that we know from physics — a certain mathematical construct or a certain law in physics. We noticed that hey, if we use that, it could actually help with this or that particular AI algorithm.”

This approach could be useful in a variety of specific kinds of tasks, he says, but not all. “We can’t say this is useful for all of AI, but there are instances where we can use an insight from physics to improve on a given AI algorithm.”

Neural networks in general are an attempt to mimic the way humans learn certain new things: The computer examines many different examples and “learns” what the key underlying patterns are. Such systems are widely used for pattern recognition, such as learning to identify objects depicted in photos.

But neural networks in general have difficulty correlating information from a long string of data, such as is required in interpreting a research paper. Various tricks have been used to improve this capability, including techniques known as long short-term memory (LSTM) and gated recurrent units (GRU), but these still fall well short of what’s needed for real natural-language processing, the researchers say.

The team came up with an alternative system, which instead of being based on the multiplication of matrices, as most conventional neural networks are, is based on vectors rotating in a multidimensional space. The key concept is something they call a rotational unit of memory (RUM).

Essentially, the system represents each word in the text by a vector in multidimensional space — a line of a certain length pointing in a particular direction. Each subsequent word swings this vector in some direction, represented in a theoretical space that can ultimately have thousands of dimensions. At the end of the process, the final vector or set of vectors is translated back into its corresponding string of words.

“RUM helps neural networks to do two things very well,” Nakov says. “It helps them to remember better, and it enables them to recall information more accurately.”

After developing the RUM system to help with certain tough physics problems such as the behavior of light in complex engineered materials, “we realized one of the places where we thought this approach could be useful would be natural language processing,” says Soljačić,  recalling a conversation with Tatalović, who noted that such a tool would be useful for his work as an editor trying to decide which papers to write about. Tatalović was at the time exploring AI in science journalism as his Knight fellowship project.

“And so we tried a few natural language processing tasks on it,” Soljačić says. “One that we tried was summarizing articles, and that seems to be working quite well.”

The proof is in the reading

As an example, they fed the same research paper through a conventional LSTM-based neural network and through their RUM-based system. The resulting summaries were dramatically different.

The LSTM system yielded this highly repetitive and fairly technical summary: “Baylisascariasis,” kills mice, has endangered the allegheny woodrat and has caused disease like blindness or severe consequences. This infection, termed “baylisascariasis,” kills mice, has endangered the allegheny woodrat and has caused disease like blindness or severe consequences. This infection, termed “baylisascariasis,” kills mice, has endangered the allegheny woodrat.

Based on the same paper, the RUM system produced a much more readable summary, and one that did not include the needless repetition of phrases: Urban raccoons may infect people more than previously assumed. 7 percent of surveyed individuals tested positive for raccoon roundworm antibodies. Over 90 percent of raccoons in Santa Barbara play host to this parasite.

Already, the RUM-based system has been expanded so it can “read” through entire research papers, not just the abstracts, to produce a summary of their contents. The researchers have even tried using the system on their own research paper describing these findings — the paper that this news story is attempting to summarize.

Here is the new neural network’s summary: Researchers have developed a new representation process on the rotational unit of RUM, a recurrent memory that can be used to solve a broad spectrum of the neural revolution in natural language processing.

It may not be elegant prose, but it does at least hit the key points of information.

Çağlar Gülçehre, a research scientist at the British AI company Deepmind Technologies, who was not involved in this work, says this research tackles an important problem in neural networks, having to do with relating pieces of information that are widely separated in time or space. “This problem has been a very fundamental issue in AI due to the necessity to do reasoning over long time-delays in sequence-prediction tasks,” he says. “Although I do not think this paper completely solves this problem, it shows promising results on the long-term dependency tasks such as question-answering, text summarization, and associative recall.”

Gülçehre adds, “Since the experiments conducted and model proposed in this paper are released as open-source on Github, as a result many researchers will be interested in trying it on their own tasks. … To be more specific, potentially the approach proposed in this paper can have very high impact on the fields of natural language processing and reinforcement learning, where the long-term dependencies are very crucial.”

The research received support from the Army Research Office, the National Science Foundation, the MIT-SenseTime Alliance on Artificial Intelligence, and the Semiconductor Research Corporation. The team also had help from the Science Daily website, whose articles were used in training some of the AI models in this research.

As usual, this ‘automated writing system’ is framed as a ‘helper’ not an usurper of anyone’s job. However, its potential for changing the nature of the work is there. About five years ago I featured another ‘automated writing’ story in a July 16, 2014 posting titled: ‘Writing and AI or is a robot writing this blog?’ You may have been reading ‘automated’ news stories for years. At the time, the focus was on sports and business.

Getting back to 2019 and science writing, here’s a link to and a citation for the paper,

Rotational Unit of Memory: A Novel Representation Unit for RNNs with Scalable Applications by Rumen Dangovski, Li Jing, Preslav Nakov, Mićo Tatalović and Marin Soljačić. Transactions of the Association for Computational Linguistics Volume 07, 2019 pp.121-138 DOI: https://doi.org/10.1162/tacl_a_00258 Posted Online 2019

© 2019 Association for Computational Linguistics. Distributed under a CC-BY 4.0 license.

This paper is open access.

Robot radiologists (artificially intelligent doctors)

Mutaz Musa, a physician at New York Presbyterian Hospital/Weill Cornell (Department of Emergency Medicine) and software developer in New York City, has penned an eyeopening opinion piece about artificial intelligence (or robots if you prefer) and the field of radiology. From a June 25, 2018 opinion piece for The Scientist (Note: Links have been removed),

Although artificial intelligence has raised fears of job loss for many, we doctors have thus far enjoyed a smug sense of security. There are signs, however, that the first wave of AI-driven redundancies among doctors is fast approaching. And radiologists seem to be first on the chopping block.

Andrew Ng, founder of online learning platform Coursera and former CTO of “China’s Google,” Baidu, recently announced the development of CheXNet, a convolutional neural net capable of recognizing pneumonia and other thoracic pathologies on chest X-rays better than human radiologists. Earlier this year, a Hungarian group developed a similar system for detecting and classifying features of breast cancer in mammograms. In 2017, Adelaide University researchers published details of a bot capable of matching human radiologist performance in detecting hip fractures. And, of course, Google achieved superhuman proficiency in detecting diabetic retinopathy in fundus photographs, a task outside the scope of most radiologists.

Beyond single, two-dimensional radiographs, a team at Oxford University developed a system for detecting spinal disease from MRI data with a performance equivalent to a human radiologist. Meanwhile, researchers at the University of California, Los Angeles, reported detecting pathology on head CT scans with an error rate more than 20 times lower than a human radiologist.

Although these particular projects are still in the research phase and far from perfect—for instance, often pitting their machines against a limited number of radiologists—the pace of progress alone is telling.

Others have already taken their algorithms out of the lab and into the marketplace. Enlitic, founded by Aussie serial entrepreneur and University of San Francisco researcher Jeremy Howard, is a Bay-Area startup that offers automated X-ray and chest CAT scan interpretation services. Enlitic’s systems putatively can judge the malignancy of nodules up to 50 percent more accurately than a panel of radiologists and identify fractures so small they’d typically be missed by the human eye. One of Enlitic’s largest investors, Capitol Health, owns a network of diagnostic imaging centers throughout Australia, anticipating the broad rollout of this technology. Another Bay-Area startup, Arterys, offers cloud-based medical imaging diagnostics. Arterys’s services extend beyond plain films to cardiac MRIs and CAT scans of the chest and abdomen. And there are many others.

Musa has offered a compelling argument with lots of links to supporting evidence.

[downloaded from https://www.the-scientist.com/news-opinion/opinion–rise-of-the-robot-radiologists-64356]

And evidence keeps mounting, I just stumbled across this June 30, 2018 news item on Xinhuanet.com,

An artificial intelligence (AI) system scored 2:0 against elite human physicians Saturday in two rounds of competitions in diagnosing brain tumors and predicting hematoma expansion in Beijing.

The BioMind AI system, developed by the Artificial Intelligence Research Centre for Neurological Disorders at the Beijing Tiantan Hospital and a research team from the Capital Medical University, made correct diagnoses in 87 percent of 225 cases in about 15 minutes, while a team of 15 senior doctors only achieved 66-percent accuracy.

The AI also gave correct predictions in 83 percent of brain hematoma expansion cases, outperforming the 63-percent accuracy among a group of physicians from renowned hospitals across the country.

The outcomes for human physicians were quite normal and even better than the average accuracy in ordinary hospitals, said Gao Peiyi, head of the radiology department at Tiantan Hospital, a leading institution on neurology and neurosurgery.

To train the AI, developers fed it tens of thousands of images of nervous system-related diseases that the Tiantan Hospital has archived over the past 10 years, making it capable of diagnosing common neurological diseases such as meningioma and glioma with an accuracy rate of over 90 percent, comparable to that of a senior doctor.

All the cases were real and contributed by the hospital, but never used as training material for the AI, according to the organizer.

Wang Yongjun, executive vice president of the Tiantan Hospital, said that he personally did not care very much about who won, because the contest was never intended to pit humans against technology but to help doctors learn and improve [emphasis mine] through interactions with technology.

“I hope through this competition, doctors can experience the power of artificial intelligence. This is especially so for some doctors who are skeptical about artificial intelligence. I hope they can further understand AI and eliminate their fears toward it,” said Wang.

Dr. Lin Yi who participated and lost in the second round, said that she welcomes AI, as it is not a threat but a “friend.” [emphasis mine]

AI will not only reduce the workload but also push doctors to keep learning and improve their skills, said Lin.

Bian Xiuwu, an academician with the Chinese Academy of Science and a member of the competition’s jury, said there has never been an absolute standard correct answer in diagnosing developing diseases, and the AI would only serve as an assistant to doctors in giving preliminary results. [emphasis mine]

Dr. Paul Parizel, former president of the European Society of Radiology and another member of the jury, also agreed that AI will not replace doctors, but will instead function similar to how GPS does for drivers. [emphasis mine]

Dr. Gauden Galea, representative of the World Health Organization in China, said AI is an exciting tool for healthcare but still in the primitive stages.

Based on the size of its population and the huge volume of accessible digital medical data, China has a unique advantage in developing medical AI, according to Galea.

China has introduced a series of plans in developing AI applications in recent years.

In 2017, the State Council issued a development plan on the new generation of Artificial Intelligence and the Ministry of Industry and Information Technology also issued the “Three-Year Action Plan for Promoting the Development of a New Generation of Artificial Intelligence (2018-2020).”

The Action Plan proposed developing medical image-assisted diagnostic systems to support medicine in various fields.

I note the reference to cars and global positioning systems (GPS) and their role as ‘helpers’;, it seems no one at the ‘AI and radiology’ competition has heard of driverless cars. Here’s Musa on those reassuring comments abut how the technology won’t replace experts but rather augment their skills,

To be sure, these services frame themselves as “support products” that “make doctors faster,” rather than replacements that make doctors redundant. This language may reflect a reserved view of the technology, though it likely also represents a marketing strategy keen to avoid threatening or antagonizing incumbents. After all, many of the customers themselves, for now, are radiologists.

Radiology isn’t the only area where experts might find themselves displaced.

Eye experts

It seems inroads have been made by artificial intelligence systems (AI) into the diagnosis of eye diseases. It got the ‘Fast Company’ treatment (exciting new tech, learn all about it) as can be seen further down in this posting. First, here’s a more restrained announcement, from an August 14, 2018 news item on phys.org (Note: A link has been removed),

An artificial intelligence (AI) system, which can recommend the correct referral decision for more than 50 eye diseases, as accurately as experts has been developed by Moorfields Eye Hospital NHS Foundation Trust, DeepMind Health and UCL [University College London].

The breakthrough research, published online by Nature Medicine, describes how machine-learning technology has been successfully trained on thousands of historic de-personalised eye scans to identify features of eye disease and recommend how patients should be referred for care.

Researchers hope the technology could one day transform the way professionals carry out eye tests, allowing them to spot conditions earlier and prioritise patients with the most serious eye diseases before irreversible damage sets in.

An August 13, 2018 UCL press release, which originated the news item, describes the research and the reasons behind it in more detail,

More than 285 million people worldwide live with some form of sight loss, including more than two million people in the UK. Eye diseases remain one of the biggest causes of sight loss, and many can be prevented with early detection and treatment.

Dr Pearse Keane, NIHR Clinician Scientist at the UCL Institute of Ophthalmology and consultant ophthalmologist at Moorfields Eye Hospital NHS Foundation Trust said: “The number of eye scans we’re performing is growing at a pace much faster than human experts are able to interpret them. There is a risk that this may cause delays in the diagnosis and treatment of sight-threatening diseases, which can be devastating for patients.”

“The AI technology we’re developing is designed to prioritise patients who need to be seen and treated urgently by a doctor or eye care professional. If we can diagnose and treat eye conditions early, it gives us the best chance of saving people’s sight. With further research it could lead to greater consistency and quality of care for patients with eye problems in the future.”

The study, launched in 2016, brought together leading NHS eye health professionals and scientists from UCL and the National Institute for Health Research (NIHR) with some of the UK’s top technologists at DeepMind to investigate whether AI technology could help improve the care of patients with sight-threatening diseases, such as age-related macular degeneration and diabetic eye disease.

Using two types of neural network – mathematical systems for identifying patterns in images or data – the AI system quickly learnt to identify 10 features of eye disease from highly complex optical coherence tomography (OCT) scans. The system was then able to recommend a referral decision based on the most urgent conditions detected.

To establish whether the AI system was making correct referrals, clinicians also viewed the same OCT scans and made their own referral decisions. The study concluded that AI was able to make the right referral recommendation more than 94% of the time, matching the performance of expert clinicians.

The AI has been developed with two unique features which maximise its potential use in eye care. Firstly, the system can provide information that helps explain to eye care professionals how it arrives at its recommendations. This information includes visuals of the features of eye disease it has identified on the OCT scan and the level of confidence the system has in its recommendations, in the form of a percentage. This functionality is crucial in helping clinicians scrutinise the technology’s recommendations and check its accuracy before deciding the type of care and treatment a patient receives.

Secondly, the AI system can be easily applied to different types of eye scanner, not just the specific model on which it was trained. This could significantly increase the number of people who benefit from this technology and future-proof it, so it can still be used even as OCT scanners are upgraded or replaced over time.

The next step is for the research to go through clinical trials to explore how this technology might improve patient care in practice, and regulatory approval before it can be used in hospitals and other clinical settings.

If clinical trials are successful in demonstrating that the technology can be used safely and effectively, Moorfields will be able to use an eventual, regulatory-approved product for free, across all 30 of their UK hospitals and community clinics, for an initial period of five years.

The work that has gone into this project will also help accelerate wider NHS research for many years to come. For example, DeepMind has invested significant resources to clean, curate and label Moorfields’ de-identified research dataset to create one of the most advanced eye research databases in the world.

Moorfields owns this database as a non-commercial public asset, which is already forming the basis of nine separate medical research studies. In addition, Moorfields can also use DeepMind’s trained AI model for future non-commercial research efforts, which could help advance medical research even further.

Mustafa Suleyman, Co-founder and Head of Applied AI at DeepMind Health, said: “We set up DeepMind Health because we believe artificial intelligence can help solve some of society’s biggest health challenges, like avoidable sight loss, which affects millions of people across the globe. These incredibly exciting results take us one step closer to that goal and could, in time, transform the diagnosis, treatment and management of patients with sight threatening eye conditions, not just at Moorfields, but around the world.”

Professor Sir Peng Tee Khaw, director of the NIHR Biomedical Research Centre at Moorfields Eye Hospital NHS Foundation Trust and UCL Institute of Ophthalmology said: “The results of this pioneering research with DeepMind are very exciting and demonstrate the potential sight-saving impact AI could have for patients. I am in no doubt that AI has a vital role to play in the future of healthcare, particularly when it comes to training and helping medical professionals so that patients benefit from vital treatment earlier than might previously have been possible. This shows the transformative research than can be carried out in the UK combining world leading industry and NIHR/NHS hospital/university partnerships.”

Matt Hancock, Health and Social Care Secretary, said: “This is hugely exciting and exactly the type of technology which will benefit the NHS in the long term and improve patient care – that’s why we fund over a billion pounds a year in health research as part of our long term plan for the NHS.”

Here’s a link to and a citation for the study,

Clinically applicable deep learning for diagnosis and referral in retinal disease by Jeffrey De Fauw, Joseph R. Ledsam, Bernardino Romera-Paredes, Stanislav Nikolov, Nenad Tomasev, Sam Blackwell, Harry Askham, Xavier Glorot, Brendan O’Donoghue, Daniel Visentin, George van den Driessche, Balaji Lakshminarayanan, Clemens Meyer, Faith Mackinder, Simon Bouton, Kareem Ayoub, Reena Chopra, Dominic King, Alan Karthikesalingam, Cían O. Hughes, Rosalind Raine, Julian Hughes, Dawn A. Sim, Catherine Egan, Adnan Tufail, Hugh Montgomery, Demis Hassabis, Geraint Rees, Trevor Back, Peng T. Khaw, Mustafa Suleyman, Julien Cornebise, Pearse A. Keane, & Olaf Ronneberger. Nature Medicine (2018) DOI: https://doi.org/10.1038/s41591-018-0107-6 Published 13 August 2018

This paper is behind a paywall.

And now, Melissa Locker’s August 15, 2018 article for Fast Company (Note: Links have been removed),

In a paper published in Nature Medicine on Monday, Google’s DeepMind subsidiary, UCL, and researchers at Moorfields Eye Hospital showed off their new AI system. The researchers used deep learning to create algorithm-driven software that can identify common patterns in data culled from dozens of common eye diseases from 3D scans. The result is an AI that can identify more than 50 diseases with incredible accuracy and can then refer patients to a specialist. Even more important, though, is that the AI can explain why a diagnosis was made, indicating which part of the scan prompted the outcome. It’s an important step in both medicine and in making AIs slightly more human

The editor or writer has even highlighted the sentence about the system’s accuracy—not just good but incredible!

I will be publishing something soon [my August 21, 2018 posting] which highlights some of the questions one might want to ask about AI and medicine before diving headfirst into this brave new world of medicine.

Announcing the ‘memtransistor’

Yet another advance toward ‘brainlike’ computing (how many times have I written this or a variation thereof in the last 10 years? See: Dexter Johnson’s take on the situation at the end of this post): Northwestern University announced their latest memristor research in a February 21, 2018 news item on Nanowerk,

Computer algorithms might be performing brain-like functions, such as facial recognition and language translation, but the computers themselves have yet to operate like brains.

“Computers have separate processing and memory storage units, whereas the brain uses neurons to perform both functions,” said Northwestern University’s Mark C. Hersam. “Neural networks can achieve complicated computation with significantly lower energy consumption compared to a digital computer.”

A February 21, 2018 Northwestern University news release (also on EurekAlert), which originated the news item, provides more information about the latest work from this team,

In recent years, researchers have searched for ways to make computers more neuromorphic, or brain-like, in order to perform increasingly complicated tasks with high efficiency. Now Hersam, a Walter P. Murphy Professor of Materials Science and Engineering in Northwestern’s McCormick School of Engineering, and his team are bringing the world closer to realizing this goal.

The research team has developed a novel device called a “memtransistor,” which operates much like a neuron by performing both memory and information processing. With combined characteristics of a memristor and transistor, the memtransistor also encompasses multiple terminals that operate more similarly to a neural network.

Supported by the National Institute of Standards and Technology and the National Science Foundation, the research was published online today, February 22 [2018], in Nature. Vinod K. Sangwan and Hong-Sub Lee, postdoctoral fellows advised by Hersam, served as the paper’s co-first authors.

The memtransistor builds upon work published in 2015, in which Hersam, Sangwan, and their collaborators used single-layer molybdenum disulfide (MoS2) to create a three-terminal, gate-tunable memristor for fast, reliable digital memory storage. Memristor, which is short for “memory resistors,” are resistors in a current that “remember” the voltage previously applied to them. Typical memristors are two-terminal electronic devices, which can only control one voltage channel. By transforming it into a three-terminal device, Hersam paved the way for memristors to be used in more complex electronic circuits and systems, such as neuromorphic computing.

To develop the memtransistor, Hersam’s team again used atomically thin MoS2 with well-defined grain boundaries, which influence the flow of current. Similar to the way fibers are arranged in wood, atoms are arranged into ordered domains – called “grains” – within a material. When a large voltage is applied, the grain boundaries facilitate atomic motion, causing a change in resistance.

“Because molybdenum disulfide is atomically thin, it is easily influenced by applied electric fields,” Hersam explained. “This property allows us to make a transistor. The memristor characteristics come from the fact that the defects in the material are relatively mobile, especially in the presence of grain boundaries.”

But unlike his previous memristor, which used individual, small flakes of MoS2, Hersam’s memtransistor makes use of a continuous film of polycrystalline MoS2 that comprises a large number of smaller flakes. This enabled the research team to scale up the device from one flake to many devices across an entire wafer.

“When length of the device is larger than the individual grain size, you are guaranteed to have grain boundaries in every device across the wafer,” Hersam said. “Thus, we see reproducible, gate-tunable memristive responses across large arrays of devices.”

After fabricating memtransistors uniformly across an entire wafer, Hersam’s team added additional electrical contacts. Typical transistors and Hersam’s previously developed memristor each have three terminals. In their new paper, however, the team realized a seven-terminal device, in which one terminal controls the current among the other six terminals.

“This is even more similar to neurons in the brain,” Hersam said, “because in the brain, we don’t usually have one neuron connected to only one other neuron. Instead, one neuron is connected to multiple other neurons to form a network. Our device structure allows multiple contacts, which is similar to the multiple synapses in neurons.”

Next, Hersam and his team are working to make the memtransistor faster and smaller. Hersam also plans to continue scaling up the device for manufacturing purposes.

“We believe that the memtransistor can be a foundational circuit element for new forms of neuromorphic computing,” he said. “However, making dozens of devices, as we have done in our paper, is different than making a billion, which is done with conventional transistor technology today. Thus far, we do not see any fundamental barriers that will prevent further scale up of our approach.”

The researchers have made this illustration available,

Caption: This is the memtransistor symbol overlaid on an artistic rendering of a hypothetical circuit layout in the shape of a brain. Credit; Hersam Research Group

Here’s a link to and a citation for the paper,

Multi-terminal memtransistors from polycrystalline monolayer molybdenum disulfide by Vinod K. Sangwan, Hong-Sub Lee, Hadallia Bergeron, Itamar Balla, Megan E. Beck, Kan-Sheng Chen, & Mark C. Hersam. Nature volume 554, pages 500–504 (22 February 2018 doi:10.1038/nature25747 Published online: 21 February 2018

This paper is behind a paywall.

The team’s earlier work referenced in the news release was featured here in an April 10, 2015 posting.

Dexter Johnson

From a Feb. 23, 2018 posting by Dexter Johnson on the Nanoclast blog (on the IEEE [Institute of Electrical and Electronics Engineers] website),

While this all seems promising, one of the big shortcomings in neuromorphic computing has been that it doesn’t mimic the brain in a very important way. In the brain, for every neuron there are a thousand synapses—the electrical signal sent between the neurons of the brain. This poses a problem because a transistor only has a single terminal, hardly an accommodating architecture for multiplying signals.

Now researchers at Northwestern University, led by Mark Hersam, have developed a new device that combines memristors—two-terminal non-volatile memory devices based on resistance switching—with transistors to create what Hersam and his colleagues have dubbed a “memtransistor” that performs both memory storage and information processing.

This most recent research builds on work that Hersam and his team conducted back in 2015 in which the researchers developed a three-terminal, gate-tunable memristor that operated like a kind of synapse.

While this work was recognized as mimicking the low-power computing of the human brain, critics didn’t really believe that it was acting like a neuron since it could only transmit a signal from one artificial neuron to another. This was far short of a human brain that is capable of making tens of thousands of such connections.

“Traditional memristors are two-terminal devices, whereas our memtransistors combine the non-volatility of a two-terminal memristor with the gate-tunability of a three-terminal transistor,” said Hersam to IEEE Spectrum. “Our device design accommodates additional terminals, which mimic the multiple synapses in neurons.”

Hersam believes that these unique attributes of these multi-terminal memtransistors are likely to present a range of new opportunities for non-volatile memory and neuromorphic computing.

If you have the time and the interest, Dexter’s post provides more context,

New breed of memristors?

This new ‘breed’ of memristor (a component in brain-like/neuromorphic computing) is a kind of thin film. First, here’s an explanation of neuromorphic computing from the Finnish researchers looking into a new kind of memristor, from a January 10, 2018 news item on Nanowerk,

The internet of things [IOT] is coming, that much we know. But still it won’t; not until we have components and chips that can handle the explosion of data that comes with IoT. In 2020, there will already be 50 billion industrial internet sensors in place all around us. A single autonomous device – a smart watch, a cleaning robot, or a driverless car – can produce gigabytes of data each day, whereas an airbus may have over 10 000 sensors in one wing alone.

Two hurdles need to be overcome. First, current transistors in computer chips must be miniaturized to the size of only few nanometres; the problem is they won’t work anymore then. Second, analysing and storing unprecedented amounts of data will require equally huge amounts of energy. Sayani Majumdar, Academy Fellow at Aalto University, along with her colleagues, is designing technology to tackle both issues.

Majumdar has with her colleagues designed and fabricated the basic building blocks of future components in what are called “neuromorphic” computers inspired by the human brain. It’s a field of research on which the largest ICT companies in the world and also the EU are investing heavily. Still, no one has yet come up with a nano-scale hardware architecture that could be scaled to industrial manufacture and use.

An Aalto University January 10, 2018 press release, which originated the news item, provides more detail about the work,

“The technology and design of neuromorphic computing is advancing more rapidly than its rival revolution, quantum computing. There is already wide speculation both in academia and company R&D about ways to inscribe heavy computing capabilities in the hardware of smart phones, tablets and laptops. The key is to achieve the extreme energy-efficiency of a biological brain and mimic the way neural networks process information through electric impulses,” explains Majumdar.

Basic components for computers that work like the brain

In their recent article in Advanced Functional Materials, Majumdar and her team show how they have fabricated a new breed of “ferroelectric tunnel junctions”, that is, few-nanometre-thick ferroelectric thin films sandwiched between two electrodes. They have abilities beyond existing technologies and bode well for energy-efficient and stable neuromorphic computing.

The junctions work in low voltages of less than five volts and with a variety of electrode materials – including silicon used in chips in most of our electronics. They also can retain data for more than 10 years without power and be manufactured in normal conditions.

Tunnel junctions have up to this point mostly been made of metal oxides and require 700 degree Celsius temperatures and high vacuums to manufacture. Ferroelectric materials also contain lead which makes them – and all our computers – a serious environmental hazard.

“Our junctions are made out of organic hydro-carbon materials and they would reduce the amount of toxic heavy metal waste in electronics. We can also make thousands of junctions a day in room temperature without them suffering from the water or oxygen in the air”, explains Majumdar.

What makes ferroelectric thin film components great for neuromorphic computers is their ability to switch between not only binary states – 0 and 1 – but a large number of intermediate states as well. This allows them to ‘memorise’ information not unlike the brain: to store it for a long time with minute amounts of energy and to retain the information they have once received – even after being switched off and on again.

We are no longer talking of transistors, but ‘memristors’. They are ideal for computation similar to that in biological brains.  Take for example the Mars 2020 Rover about to go chart the composition of another planet. For the Rover to work and process data on its own using only a single solar panel as an energy source, the unsupervised algorithms in it will need to use an artificial brain in the hardware.

“What we are striving for now, is to integrate millions of our tunnel junction memristors into a network on a one square centimetre area. We can expect to pack so many in such a small space because we have now achieved a record-high difference in the current between on and off-states in the junctions and that provides functional stability. The memristors could then perform complex tasks like image and pattern recognition and make decisions autonomously,” says Majumdar.

The probe-station device (the full instrument, left, and a closer view of the device connection, right) which measures the electrical responses of the basic components for computers mimicking the human brain. The tunnel junctions are on a thin film on the substrate plate. Photo: Tapio Reinekoski

Here’s a link to and a citation for the paper,

Electrode Dependence of Tunneling Electroresistance and Switching Stability in Organic Ferroelectric P(VDF-TrFE)-Based Tunnel Junctions by Sayani Majumdar, Binbin Chen, Qi Hang Qin, Himadri S. Majumdar, and Sebastiaan van Dijken. Advanced Functional Materials Vol. 28 Issue 2 DOI: 10.1002/adfm.201703273 Version of Record online: 27 NOV 2017

© 2017 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim

This paper is behind a paywall.

High-performance, low-energy artificial synapse for neural network computing

This artificial synapse is apparently an improvement on the standard memristor-based artificial synapse but that doesn’t become clear until reading the abstract for the paper. First, there’s a Feb. 20, 2017 Stanford University news release by Taylor Kubota (dated Feb. 21, 2017 on EurekAlert), Note: Links have been removed,

For all the improvements in computer technology over the years, we still struggle to recreate the low-energy, elegant processing of the human brain. Now, researchers at Stanford University and Sandia National Laboratories have made an advance that could help computers mimic one piece of the brain’s efficient design – an artificial version of the space over which neurons communicate, called a synapse.

“It works like a real synapse but it’s an organic electronic device that can be engineered,” said Alberto Salleo, associate professor of materials science and engineering at Stanford and senior author of the paper. “It’s an entirely new family of devices because this type of architecture has not been shown before. For many key metrics, it also performs better than anything that’s been done before with inorganics.”

The new artificial synapse, reported in the Feb. 20 issue of Nature Materials, mimics the way synapses in the brain learn through the signals that cross them. This is a significant energy savings over traditional computing, which involves separately processing information and then storing it into memory. Here, the processing creates the memory.

This synapse may one day be part of a more brain-like computer, which could be especially beneficial for computing that works with visual and auditory signals. Examples of this are seen in voice-controlled interfaces and driverless cars. Past efforts in this field have produced high-performance neural networks supported by artificially intelligent algorithms but these are still distant imitators of the brain that depend on energy-consuming traditional computer hardware.

Building a brain

When we learn, electrical signals are sent between neurons in our brain. The most energy is needed the first time a synapse is traversed. Every time afterward, the connection requires less energy. This is how synapses efficiently facilitate both learning something new and remembering what we’ve learned. The artificial synapse, unlike most other versions of brain-like computing, also fulfills these two tasks simultaneously, and does so with substantial energy savings.

“Deep learning algorithms are very powerful but they rely on processors to calculate and simulate the electrical states and store them somewhere else, which is inefficient in terms of energy and time,” said Yoeri van de Burgt, former postdoctoral scholar in the Salleo lab and lead author of the paper. “Instead of simulating a neural network, our work is trying to make a neural network.”

The artificial synapse is based off a battery design. It consists of two thin, flexible films with three terminals, connected by an electrolyte of salty water. The device works as a transistor, with one of the terminals controlling the flow of electricity between the other two.

Like a neural path in a brain being reinforced through learning, the researchers program the artificial synapse by discharging and recharging it repeatedly. Through this training, they have been able to predict within 1 percent of uncertainly what voltage will be required to get the synapse to a specific electrical state and, once there, it remains at that state. In other words, unlike a common computer, where you save your work to the hard drive before you turn it off, the artificial synapse can recall its programming without any additional actions or parts.

Testing a network of artificial synapses

Only one artificial synapse has been produced but researchers at Sandia used 15,000 measurements from experiments on that synapse to simulate how an array of them would work in a neural network. They tested the simulated network’s ability to recognize handwriting of digits 0 through 9. Tested on three datasets, the simulated array was able to identify the handwritten digits with an accuracy between 93 to 97 percent.

Although this task would be relatively simple for a person, traditional computers have a difficult time interpreting visual and auditory signals.

“More and more, the kinds of tasks that we expect our computing devices to do require computing that mimics the brain because using traditional computing to perform these tasks is becoming really power hungry,” said A. Alec Talin, distinguished member of technical staff at Sandia National Laboratories in Livermore, California, and senior author of the paper. “We’ve demonstrated a device that’s ideal for running these type of algorithms and that consumes a lot less power.”

This device is extremely well suited for the kind of signal identification and classification that traditional computers struggle to perform. Whereas digital transistors can be in only two states, such as 0 and 1, the researchers successfully programmed 500 states in the artificial synapse, which is useful for neuron-type computation models. In switching from one state to another they used about one-tenth as much energy as a state-of-the-art computing system needs in order to move data from the processing unit to the memory.

This, however, means they are still using about 10,000 times as much energy as the minimum a biological synapse needs in order to fire. The researchers are hopeful that they can attain neuron-level energy efficiency once they test the artificial synapse in smaller devices.

Organic potential

Every part of the device is made of inexpensive organic materials. These aren’t found in nature but they are largely composed of hydrogen and carbon and are compatible with the brain’s chemistry. Cells have been grown on these materials and they have even been used to make artificial pumps for neural transmitters. The voltages applied to train the artificial synapse are also the same as those that move through human neurons.

All this means it’s possible that the artificial synapse could communicate with live neurons, leading to improved brain-machine interfaces. The softness and flexibility of the device also lends itself to being used in biological environments. Before any applications to biology, however, the team plans to build an actual array of artificial synapses for further research and testing.

Additional Stanford co-authors of this work include co-lead author Ewout Lubberman, also of the University of Groningen in the Netherlands, Scott T. Keene and Grégorio C. Faria, also of Universidade de São Paulo, in Brazil. Sandia National Laboratories co-authors include Elliot J. Fuller and Sapan Agarwal in Livermore and Matthew J. Marinella in Albuquerque, New Mexico. Salleo is an affiliate of the Stanford Precourt Institute for Energy and the Stanford Neurosciences Institute. Van de Burgt is now an assistant professor in microsystems and an affiliate of the Institute for Complex Molecular Studies (ICMS) at Eindhoven University of Technology in the Netherlands.

This research was funded by the National Science Foundation, the Keck Faculty Scholar Funds, the Neurofab at Stanford, the Stanford Graduate Fellowship, Sandia’s Laboratory-Directed Research and Development Program, the U.S. Department of Energy, the Holland Scholarship, the University of Groningen Scholarship for Excellent Students, the Hendrik Muller National Fund, the Schuurman Schimmel-van Outeren Foundation, the Foundation of Renswoude (The Hague and Delft), the Marco Polo Fund, the Instituto Nacional de Ciência e Tecnologia/Instituto Nacional de Eletrônica Orgânica in Brazil, the Fundação de Amparo à Pesquisa do Estado de São Paulo and the Brazilian National Council.

Here’s an abstract for the researchers’ paper (link to paper provided after abstract) and it’s where you’ll find the memristor connection explained,

The brain is capable of massively parallel information processing while consuming only ~1–100fJ per synaptic event1, 2. Inspired by the efficiency of the brain, CMOS-based neural architectures3 and memristors4, 5 are being developed for pattern recognition and machine learning. However, the volatility, design complexity and high supply voltages for CMOS architectures, and the stochastic and energy-costly switching of memristors complicate the path to achieve the interconnectivity, information density, and energy efficiency of the brain using either approach. Here we describe an electrochemical neuromorphic organic device (ENODe) operating with a fundamentally different mechanism from existing memristors. ENODe switches at low voltage and energy (<10pJ for 103μm2 devices), displays >500 distinct, non-volatile conductance states within a ~1V range, and achieves high classification accuracy when implemented in neural network simulations. Plastic ENODes are also fabricated on flexible substrates enabling the integration of neuromorphic functionality in stretchable electronic systems6, 7. Mechanical flexibility makes ENODes compatible with three-dimensional architectures, opening a path towards extreme interconnectivity comparable to the human brain.

Here’s a link to and a citation for the paper,

A non-volatile organic electrochemical device as a low-voltage artificial synapse for neuromorphic computing by Yoeri van de Burgt, Ewout Lubberman, Elliot J. Fuller, Scott T. Keene, Grégorio C. Faria, Sapan Agarwal, Matthew J. Marinella, A. Alec Talin, & Alberto Salleo. Nature Materials (2017) doi:10.1038/nmat4856 Published online 20 February 2017

This paper is behind a paywall.

ETA March 8, 2017 10:28 PST: You may find this this piece on ferroelectricity and neuromorphic engineering of interest (March 7, 2017 posting titled: Ferroelectric roadmap to neuromorphic computing).

Resistive memory from University of California Riverside (replacing flash memory in mobile devices) and Boise State University (neuron chips)

Today, (Aug. 19, 2 013)I have two items on memristors. First, Dexter Johnson provides some context for understanding why a University of California Riverside research team’s approach to creating memristors is exciting some interest in his Aug. 17, 2013 posting (Nanoclast blog on the IEEE [Institute of Electrical and Electronics Engineers] website), Note: Links have been removed,

The heralding of the memristor, or resistive memory, and the long-anticipated demise of flash memory have both been tracking on opposite trajectories with resistive memory expected to displace flash ever since the memristor was first discovered by Stanley Williams’ group at Hewlett Packard in 2008.

The memristor has been on a rapid development track ever since and has been promised to be commercially available as early as 2014, enabling 10 times greater embedded memory for mobile devices than currently available.

The obsolescence of flash memory at the hands of the latest nanotechnology has been predicted for longer than the commercial introduction of the memristor. But just at the moment it appears it’s going to reach its limits in storage capacity along comes a new way to push its capabilities to new heights, sometimes thanks to a nanomaterial like graphene.

In addition to the graphene promise, Dexter goes on to discuss another development,  which could push memory capabilities and which is mentioned in an Aug. 14, 2013 news item on ScienceDaily (and elsewhere),

A team at the University of California, Riverside Bourns College of Engineering has developed a novel way to build what many see as the next generation memory storage devices for portable electronic devices including smart phones, tablets, laptops and digital cameras.

The device is based on the principles of resistive memory [memristor], which can be used to create memory cells that are smaller, operate at a higher speed and offer more storage capacity than flash memory cells, the current industry standard. Terabytes, not gigbytes, will be the norm with resistive memory.

The key advancement in the UC Riverside research is the creation of a zinc oxide nano-island on silicon. It eliminates the need for a second element called a selector device, which is often a diode.

The Aug. 13, 2013 University of California Riverside news release by Sean Nealon, which originated the news item, further describes the limitations of flash memory and reinforces the importance of being able to eliminate a component (selector device),

Flash memory has been the standard in the electronics industry for decades. But, as flash continues to get smaller and users want higher storage capacity, it appears to reaching the end of its lifespan, Liu [Jianlin Liu, a professor of electrical engineering] said.

With that in mind, resistive memory is receiving significant attention from academia and the electronics industry because it has a simple structure, high-density integration, fast operation and long endurance.

Researchers have also found that resistive memory can be scaled down in the sub 10-nanometer scale. (A nanometer is one-billionth of a meter.) Current flash memory devices are roughly using a feature size twice as large.

Resistive memory usually has a metal-oxide-metal structure in connection with a selector device. The UC Riverside team has demonstrated a novel alternative way by forming self-assembled zinc oxide nano-islands on silicon. Using a conductive atomic force microscope, the researchers observed three operation modes from the same device structure, essentially eliminating the need for a separate selector device.

Here’s a link to and a citation for the researchers’ published paper,

Multimode Resistive Switching in Single ZnO Nanoisland System by Jing Qi, Mario Olmedo, Jian-Guo Zheng, & Jianlin Liu. Scientific Reports 3, Article number: 2405 doi:10.1038/srep02405 Published 12 August 2013

This study is open access.

Meanwhile, Boise State University (Idaho, US) is celebrating a new project, CIF: Small: Realizing Chip-scale Bio-inspired Spiking Neural Networks with Monolithically Integrated Nano-scale Memristors, which was announced in an Aug. 17, 2013 news item on Azonano,

Electrical and computer engineering faculty Elisa Barney Smith, Kris Campbell and Vishal Saxena are joining forces on a project titled “CIF: Small: Realizing Chip-scale Bio-inspired Spiking Neural Networks with Monolithically Integrated Nano-scale Memristors.”

Team members are experts in machine learning (artificial intelligence), integrated circuit design and memristor devices. Funded by a three-year, $500,000 National Science Foundation grant, they have taken on the challenge of developing a new kind of computing architecture that works more like a brain than a traditional digital computer.

“By mimicking the brain’s billions of interconnections and pattern recognition capabilities, we may ultimately introduce a new paradigm in speed and power, and potentially enable systems that include the ability to learn, adapt and respond to their environment,” said Barney Smith, who is the principal investigator on the grant.

The Aug. 14, 2013 Boise State University news release by Kathleen Tuck, which originated the news item, describes the team’s focus on mimicking the brain’s capabilities ,

One of the first memristors was built in Campbell’s Boise State lab, which has the distinction of being one of only five or six labs worldwide that are up to the task.

The team’s research builds on recent work from scientists who have derived mathematical algorithms to explain the electrical interaction between brain synapses and neurons.

“By employing these models in combination with a new device technology that exhibits similar electrical response to the neural synapses, we will design entirely new computing chips that mimic how the brain processes information,” said Barney Smith.

Even better, these new chips will consume power at an order of magnitude lower than current computing processors, despite the fact that they match existing chips in physical dimensions. This will open the door for ultra low-power electronics intended for applications with scarce energy resources, such as in space, environmental sensors or biomedical implants.

Once the team has successfully built an artificial neural network, they will look to engage neurobiologists in parallel to what they are doing now. A proposal for that could be written in the coming year.

Barney Smith said they hope to send the first of the new neuron chips out for fabrication within weeks.

With the possibility that HP Labs will make its ‘memristor chips‘ commercially available in 2014 and neuron chips fabricated for the Boise State University researchers within weeks of this Aug. 19, 2013 date, it seems that memristors have been developed at a lightning fast pace. It’s been a fascinating process to observe.