Posted on

The top 10 emerging technologies of 2016

(credit: WEF)

The World Economic Forum’s annual list of this year’s breakthrough technologies, published today, includes “socially aware” openAI, grid-scale energy storage, perovskite solar cells, and other technologies with the potential to “transform industries, improve lives, and safeguard the planet.” The WEF’s specific interest is to “close gaps in investment and regulation.”

“Horizon scanning for emerging technologies is crucial to staying abreast of developments that can radically transform our world, enabling timely expert analysis in preparation for these disruptors. The global community needs to come together and agree on common principles if our society is to reap the benefits and hedge the risks of these technologies,” said Bernard Meyerson, PhD, Chief Innovation Officer of IBM and Chair of the WEF’s Meta-Council on Emerging Technologies.

The list also provides an opportunity to debate human, societal, economic or environmental risks and concerns that the technologies may pose — prior to widespread adoption.

One of the criteria used by council members during their deliberations was the likelihood that 2016 represents a tipping point in the deployment of each technology. So the list includes some technologies that have been known for a number of years, but are only now reaching a level of maturity where their impact can be meaningfully felt.

The top 10 technologies that make this year’s list are:

  1. Nanosensors and the Internet of Nanothings  — With the Internet of Things expected to comprise 30 billion connected devices by 2020, one of the most exciting areas of focus today is now on nanosensors capable of circulating in the human body or being embedded in construction materials. They could use DNA and proteins to recognize specific chemical targets, store a few bits of information, and then report their status by changing color or emitting some other easily detectable signal.
  2. Next-Generation Batteries — One of the greatest obstacles holding renewable energy back is matching supply with demand, but recent advances in energy storage using sodium, aluminum, and zinc based batteries makes mini-grids feasible that can provide clean, reliable, around-the-clock energy sources to entire villages.
  3. The Blockchain — With venture investment related to the online currency Bitcoin exceeding $1 billion in 2015 alone, the economic and social impact of blockchain’s potential to fundamentally change the way markets and governments work is only now emerging.
  4. 2D Materials — Plummeting production costs mean that 2D materials like graphene are emerging in a wide range of applications, from air and water filters to new generations of wearables and batteries.
  5. Autonomous Vehicles — The potential of self-driving vehicles for saving lives, cutting pollution, boosting economies, and improving quality of life for the elderly and other segments of society has led to rapid deployment of key technology forerunners along the way to full autonomy.
  6. Organs-on-chips — Miniature models of human organs could revolutionize medical research and drug discovery by allowing researchers to see biological mechanism behaviors in ways never before possible.
  7. Perovskite Solar Cells — This new photovoltaic material offers three improvements over the classic silicon solar cell: it is easier to make, can be used virtually anywhere and, to date, keeps on generating power more efficiently.
  8. Open AI Ecosystem — Shared advances in natural language processing and social awareness algorithms, coupled with an unprecedented availability of data, will soon allow smart digital assistants to help with a vast range of tasks, from keeping track of one’s finances and health to advising on wardrobe choice.
  9. Optogenetics — Recent developments mean light can now be delivered deeper into brain tissue, something that could lead to better treatment for people with brain disorders.
  10. Systems Metabolic Engineering — Advances in synthetic biology, systems biology, and evolutionary engineering mean that the list of building block chemicals that can be manufactured better and more cheaply by using plants rather than fossil fuels is growing every year.

To compile this list, the World Economic Forum’s Meta-Council on Emerging Technologies, a panel of global experts, “drew on the collective expertise of the Forum’s communities to identify the most important recent technological trends. By doing so, the Meta-Council aims to raise awareness of their potential and contribute to closing gaps in investment, regulation and public understanding that so often thwart progress.”

You can read 10 expert views on these technologies here or download the series as a PDF.

Posted on

Real-time robot-motion planning

New computer processor allows for fast, energy-efficient robot motion planning in cluttered environments (credit: Duke Robotics)

Duke University researchers have designed a new computer processor that’s optimized for robot motion planning (for example, for quickly picking up and accurately moving an object in a cluttered environment while evading obstacles). The new processor can plan an optimal motion path up to 10,000 times faster than existing systems while using a small fraction of the required power.

The new processor is fast enough to plan and operate in real time, and power-efficient enough to be used in large-scale manufacturing environments with thousands of robots, according to George Konidaris, assistant professor of computer science and electrical and computer engineering at Duke.

“When you think about a car assembly line, the entire environment is carefully controlled so that the robots can blindly repeat the same movements over and over again,” said Konidaris. “The car parts are in exactly the same place every time, and the robots are contained within cages so that humans don’t wander past.”

But for uncontrolled environments (such as homes), robot motion planning has to be a lot smarter and able to learn in real time. That would save the time and expense of custom-engineering the environment around the robot, said Konidaris, who presented the new work yesterday (June 20) at a conference called Robotics: Science and Systems in Ann Arbor, Mich.


Duke Robotics | Robotic Motion Planning

Collision detection in real time

Most existing approaches for robot motion planning rely on general-purpose CPUs or computationally faster but more power-hungry graphics processors (GPUs). Instead, the Duke team specifically designed a new processor for motion planning.

“While a general-purpose CPU is good at many tasks, it cannot compete with a processor specially designed for just a single task,” said Daniel Sorin, professor of electrical and computer engineering and computer science at Duke.

Konidaris and Sorin’s team designed their new processor to perform collision detection — the most time-consuming aspect of motion planning — requiring thousands of collision checks in parallel. “We streamlined our design and focused our hardware and power budgets on just the specific tasks that matter for motion planning,” Sorin said.

The key was to use an FPGA (field-programmable gate array) integrated circuit, which can be configured by a designer for customized uses.

The robot-motion processor selects the set of voxels swept by the robot arm and this set is used to build specialized circuits in an FPGA integrated circuit to detect collisions and optimize motions in real time during operation (credit: Duke Robotics)

The technology works by breaking down the arm’s operating space into thousands of 3D volumes called voxels (volume pixels). The algorithm then determines whether or not an object is present in one of the voxels contained within pre-programmed motion paths. Thanks to the specially designed hardware, the technology can check thousands of motion paths simultaneously, and then stitch together the shortest motion path possible using the “safe” options remaining.

Game-changer

“The state of the art prior to our work used high-performance, commodity graphics processors that consume 200 to 300 watts,” said Konidaris. “And even then, it was taking hundreds of milliseconds, or even as much as a second, to find a motion plan. We’re at less than a millisecond, and less than 10 watts. Even if we weren’t faster, the power savings alone will add up in factories with thousands, or even millions, of robots.”

Konidaris also notes that the technology opens up new ways to use motion planning. “Previously, planning was done once per movement, because it was so slow,” he said, “but now it is fast enough that it could be used as a component of a more complex planning algorithm, perhaps one that sequences several simpler motions or plans ahead to reason about the movement of several objects.”

The new processor’s speed and power efficiency could create many opportunities for automation. So Konidaris, Sorin and their students have formed a spinoff company, Realtime Robotics, to commercialize the technology. “Real-time motion planning could really be a game-changer for robotics,” said Konidaris.

This research was supported by the Defense Advanced Research Projects Agency and the National Institutes of Health.

Abstract of Robot Motion Planning on a Chip

We describe a process that constructs robot-specific circuitry for motion planning, capable of generating motion plans approximately three orders of magnitude faster than existing methods. Our method is based on building collision detection circuits for a probabilistic roadmap. Collision detection for the roadmap edges is completely parallelized, so that the time to determine which edges are in collision is independent of the number of edges. We demonstrate planning using a 6-degree- of-freedom robot arm in less than 1 millisecond.

Posted on

First self-driving ‘cognitive’ vehicle uses IBM Watson Internet of Things

Olli (credit: Local Motors)

Local Motors, creator of the world’s first 3D-printed cars, has developed the first self-driving “cognitive” vehicle, using IBM Watson Internet of Things (IoT) for Automotive.

The vehicle, dubbed “Olli,” can carry up to 12 people. It uses IBM Watson and other systems to improve the passenger experience and allow natural interaction with the vehicle. Olli will be used on public roads locally in Washington DC and later this year in Miami-Dade County.

Olli is the first vehicle to use the cloud-based cognitive computing capability of IBM Watson IoT to analyze and learn from high volumes of transportation data, produced by more than 30 sensors embedded throughout the vehicle. Sensors will be added and adjusted continually as passenger needs and local preferences are identified.

Four Watson developer APIs — Speech to Text, Natural Language Classifier, Entity Extraction and Text to Speech — will enable passengers to interact conversationally with Olli while traveling from point A to point B, discussing topics about how the vehicle works, where they are going, and why Olli is making specific driving decisions.

Watson empowers Olli to understand and respond to passengers’ questions as they enter the vehicle, such as destinations (“Olli, can you take me downtown?”) or specific vehicle functions (“how does this feature work?” or even “are we there yet?”). Passengers can also ask for recommendations on local destinations such as popular restaurants or historical sites based on analysis of personal preferences.

“Cognitive computing provides incredible opportunities to create unparalleled, customized experiences for customers, taking advantage of the massive amounts of streaming data from all devices connected to the Internet of Things, including an automobile’s myriad sensors and systems,” said Harriet Green, General Manager, IBM Watson Internet of Things, Commerce & Education.

Miami-Dade County and Las Vegas are also exploring a pilot program in which several autonomous vehicles would be used to transport people around Miami and Las Vegas.


IBM Internet of Things | Local Motors Debuts “Olli,” the First Self-driving Vehicle to Tap the Power of IBM Watson


IBM Internet of Things | Harnessing vehicle safety data with analytics

Posted on

Could deep-learning systems radically transform drug discovery?

(credit: Insilico Medicine)

Scientists at Insilico Medicine have developed a new drug-discovery engine that they say is capable of predicting therapeutic use, toxicity, and adverse effects of thousands of molecules, and they plan to reveal it at the Re-Work Machine Intelligence Summit in Berlin, June 29–30.

Drug discovery takes decades, with high failure rates. Among the reasons: irreproducible experiments with poor choice of animal models and inability to translate the results from animal models directly to humans, the wide variety of diseases, and communication difficulties between scientists, managers, venture capitalists, pharmaceutical companies and regulators. And perhaps the biggest problem: the slow-paced, bureaucratic culture in the pharmaceutical industry, the researchers note.

Radically transforming pharmas with AI

Insilico Medicine says it aims to address these reasons by developing “multimodal deep-learned and parametric biomarkers,” as well as multiple drug-scoring pipelines for drug discovery and drug repurposing, and hypothesis and lead generation.

“At Insilico, we want to radically transform the pharmaceutical industry and double the number of drugs on the market, using artificial intelligence and deep understanding of pharmaceutical R&D processes,” said Polina Mamoshina*, senior research scientist at Insilico Medicine, Inc.

“We decided to start with nutraceuticals and cosmetics, but soon we will be announcing our cancer immunology concomitant drug discovery engine to boost the response rates to checkpoint inhibitors in immuno-oncology.”

“Using our drug discovery engine, we made thousands of hypotheses and narrowed these down to 800 strong molecule-disease predictions, with efficacy, toxicity, adverse effects, bioavailability and many other parameters,” said Alex Aliper, president of Insilico Medicine, Inc.

“We added many drug scoring mechanisms that further validate the initial predictions and put together a team of analysts to research and evaluate individual molecules. We are now partnering with various institutions to validate these predictions in vitro and in vivo.”

As KurzweilAI reported, earlier this month, Insilico Medicine signed an exclusive agreement with Life Extension, a major nutraceutical product vendor, to collaboratively develop a set of geroprotectors — natural products that mimic the healthy young state in multiple old tissues. The goal is to increase the rejuvenation rate of the body and slow down, or even reverse, the aging process.

Polina Mamoshina was the lead author on the paper, “Applications of Deep Learning in Biomedicine” in Molecular Pharmaceutics and contributed to another publication, “Deep learning applications for predicting pharmacological properties of drugs and drug repurposing using transcriptomic data” also in Molecular Pharmaceutics. The later paper received the Editors’ Choice Award from the American Chemical Society. She also co-authored a paper, “Deep biomarkers of human aging: Application of deep neural networks to biomarker development” in Aging, one of the highest-impact journals in aging research.

Posted on

Hierarchies exist in the brain because of lower connection costs, research shows

The Evolutionary Origins of Hierarchy: Evolution with performance-only selection results in non-hierarchical and non-modular networks, which take longer to adapt to new environments. However, evolving networks with a connection cost creates hierarchical and functionally modular networks that can solve the overall problem by recursively solving its sub-problems. These networks also adapt to new environments faster. (credit: Henok Mengistu et al./PLOS Comp. Bio)

New research suggests why the human brain and other biological networks exhibit a hierarchical structure, and the study may improve attempts to create artificial intelligence.

The study, by researchers from the University of Wyoming and the French Institute for Research in Computer Science and Automation (INRIA, in France), demonstrates that the evolution of hierarchy — a simple system of ranking — in biological networks may arise because of the costs associated with network connections.

This study also supports Ray Kurzweil’s theory of the hierarchical structure of the neocortex, presented in his 2012 book, How to Create a Mind.

The human brain has separate areas for vision, motor control, and tactile processing, for example, and each of these areas consist of sub-regions that govern different parts of the body.

Evolutionary pressure to reduce the number and cost of connections

The research findings suggest that hierarchy evolves not because it produces more efficient networks, but instead because hierarchically wired networks have fewer connections. That’s because connections in biological networks are expensive — they have to be built, maintained, etc. — so there’s an evolutionary pressure to reduce the number of connections.

In addition to shedding light on the emergence of hierarchy across the many domains in which it appears, these findings may also accelerate future research into evolving more complex, intelligent computational brains in the fields of artificial intelligence and robotics.

The research, led by Henok S. Mengistu, is described in an open-access paper in PLOS Computational Biology. The researchers also simulated the evolution of computational brain models, known as artificial neural networks, both with and without a cost for network connections. They found that hierarchical structures emerge much more frequently when a cost for connections is present.

Aside from explaining why biological networks are hierarchical, the research might also explain why many man-made systems such as the Internet and road systems are also hierarchical. “The next step is to harness and combine this knowledge to evolve large-scale, structurally organized networks in the hopes of creating better artificial intelligence and increasing our understanding of the evolution of animal intelligence, including our own,” according to the researchers.

Abstract of The Evolutionary Origins of Hierarchy

Hierarchical organization—the recursive composition of sub-modules—is ubiquitous in biological networks, including neural, metabolic, ecological, and genetic regulatory networks, and in human-made systems, such as large organizations and the Internet. To date, most research on hierarchy in networks has been limited to quantifying this property. However, an open, important question in evolutionary biology is why hierarchical organization evolves in the first place. It has recently been shown that modularity evolves because of the presence of a cost for network connections. Here we investigate whether such connection costs also tend to cause a hierarchical organization of such modules. In computational simulations, we find that networks without a connection cost do not evolve to be hierarchical, even when the task has a hierarchical structure. However, with a connection cost, networks evolve to be both modular and hierarchical, and these networks exhibit higher overall performance and evolvability (i.e. faster adaptation to new environments). Additional analyses confirm that hierarchy independently improves adaptability after controlling for modularity. Overall, our results suggest that the same force–the cost of connections–promotes the evolution of both hierarchy and modularity, and that these properties are important drivers of network performance and adaptability. In addition to shedding light on the emergence of hierarchy across the many domains in which it appears, these findings will also accelerate future research into evolving more complex, intelligent computational brains in the fields of artificial intelligence and robotics.

Posted on

Soft, safe robot actuators inspired by human bicep muscles

VAMPs are functionally modeled after the human bicep, similar to the biological muscle in terms of response time and efficiency. (credit: Wyss Institute at Harvard University)

If robots are going work around humans, they will have to be softer and safer. A Harvard team has designed a new actuator with that in mind. Its movements are similar to those of a human bicep muscle, using vacuum power to automate soft rubber beams. Like real muscles, the actuators are soft, shock-absorbing, and pose no danger, according to the researchers.

The work is led by George Whitesides, Ph.D., a Core Faculty member at Harvard’s Wyss Institute for Biologically Inspired Engineering, the Woodford L. and Ann A. Flowers University Professor of Chemistry and Chemical Biology in Harvard University’s Faculty of Arts and Sciences (FAS), and a Director of the Kavli Institute for Bionano Science and Technology at Harvard University.

Whitesides’ team took an unconventional approach to its design, relying on vacuum to decrease the actuator’s volume and cause it to buckle. While conventional engineering would consider bucking to be a mechanical instability and a point of failure, in this case the team leveraged this instability to develop VAMPs (vacuum-actuated muscle-inspired pneumatic structures). Previous soft actuators rely on pressurized systems that expand in volume, but VAMPs mimic true muscle because they contract, which makes them useful in confined spaces and for a variety of purposes.

In this image, VAMPs are shown actuated and cut open in cross section. The honeycomb cross section shows the inner chambers that collapse when vacuum is applied. (credit: Wyss Institute at Harvard University)

The actuator has soft elastomeric rubber beams filled with small, hollow chambers of air like a honeycomb. By applying vacuum, the chambers collapse and the entire actuator contracts, generating movement. The internal honeycomb structure can be custom tailored to enable linear, twisting, bending, or combinatorial motions.

The team envisions that robots built with VAMPs could be used to assist the disabled or elderly, to serve food, deliver goods, and perform other tasks related to the service industry. Soft robots could also make industrial production lines safer and faster, and quality control easier to manage by enabling human operators to work in the same space.

Fail-safe design

VAMPs are designed to prevent failure — even when damaged with a 2mm hole, the team showed that VAMPs will still function. In the event that major damage is caused to the system, it fails safely. “It can’t explode, so it’s intrinsically safe,” said Whitesides. Whereas other actuators powered by electricity or combustion could cause damage to humans or their surroundings, loss of vacuum pressure in VAMPs would simply render the actuator motionless.

“These self-healing, bioinspired actuators bring us another step closer to being able to build entirely soft-bodied robots, which may help to bridge the gap between humans and robots and open entirely new application areas in medicine and beyond,” said Wyss Founding Director Donald Ingber, M.D., Ph.D., who is also the Judah Folkman Professor of Vascular Biology at Harvard Medical School and the Boston Children’s Hospital Vascular Biology Program, as well as Professor of Bioengineering at Harvard’s John A. Paulson School of Engineering and Applied Sciences (SEAS).

The work was reported June 1 in the journal Advanced Materials Technologies.

Harvard’s Office of Technology Development has filed patents on this and related inventions, and the soft actuator technology has been licensed to Soft Robotics, Inc., a startup launched in 2013 and cofounded by Whitesides. The company is developing robotic grasping systems toward initial applications including picking and packing in unstructured environments — for example, handling fruits and vegetables in produce distribution warehouses. Longer term, this technology can be leveraged to develop products for biomedical applications.

Abstract of Buckling Pneumatic Linear Actuators Inspired by Muscle

The mechanical features of biological muscles are difficult to reproduce completely in synthetic systems. A new class of soft pneumatic structures (vacuum-actuated muscle-inspired pneumatic structures) is described that combines actuation by negative pressure (vacuum), with cooperative buckling of beams fabricated in a slab of elastomer, to achieve motion and demonstrate many features that are similar to that of mammalian muscle.

Posted on

How to make opaque AI decisionmaking accountable

Machine-learning algorithms are increasingly used in making important decisions about our lives — such as credit approval, medical diagnoses, and in job applications — but exactly how they work usually remains a mystery. Now Carnegie Mellon University researchers may devised an effective way to improve transparency and head off confusion or possibly legal issues.

CMU’s new Quantitative Input Influence (QII) testing tools can generate “transparency reports” that provide the relative weight of each factor in the final decision, claims Anupam Datta, associate professor of computer science and electrical and computer engineering.

Testing for discrimination

These reports could also be used proactively by an organization to see if an artificial intelligence system is working as desired, or by a regulatory agency to determine whether a decision-making system inappropriately discriminated, based on factors like race and gender.

To achieve that, the QII measures considers correlated inputs while measuring influence. For example, consider a system that assists in hiring decisions for a moving company, in which two inputs, gender and the ability to lift heavy weights, are positively correlated with each other and with hiring decisions.

Yet transparency into whether the system actually uses weightlifting ability or gender in making its decisions has substantive implications for determining if it is engaging in discrimination. In this example, the company could keep the weightlifting ability fixed, vary gender, and check whether there is a difference in the decision.

CMU researchers are careful to state in an open-access report on QII (presented at the IEEE Symposium on Security and Privacy, May 23–25, in San Jose, Calif.), that “QII does not suggest any normative definition of fairness. Instead, we view QII as a diagnostic tool to aid fine-grained fairness determinations.”

Is your AI biased?

The Ford Foundation published a controversial blog post last November stating that “while we’re lead to believe that data doesn’t lie — and therefore, that algorithms that analyze the data can’t be prejudiced — that isn’t always true. The origin of the prejudice is not necessarily embedded in the algorithm itself. Rather, it is in the models used to process massive amounts of available data and the adaptive nature of the algorithm. As an adaptive algorithm is used, it can learn societal biases it observes.

“As Professor Alvaro Bedoya, executive director of the Center on Privacy and Technology at Georgetown University, explains, ‘any algorithm worth its salt’ will learn from the external process of bias or discriminatory behavior. To illustrate this, Professor Bedoya points to a hypothetical recruitment program that uses an algorithm written to help companies screen potential hires. If the hiring managers using the program only select younger applicants, the algorithm will learn to screen out older applicants the next time around.”

Influence variables

The QII measures also quantify the joint influence of a set of inputs (such as age and income) on outcomes, and the marginal influence of each input within the set. Since a single input may be part of multiple influential sets, the average marginal influence of the input is computed using “principled game-theoretic aggregation” measures that were previously applied to measure influence in revenue division and voting.

Examples of outcomes from transparency reports for two job applicants. Left: “Mr. X” is deemed to be a low income individual, an income classifier learned from the data. This result may be surprising to him: he reports high capital gains ($14k), and only 2.1% of people with capital gains higher than $10k are reported as low income. In fact, he might be led to believe that his classification may be a result of his ethnicity or country of origin. Examining his transparency report in the figure, however, we find that the most influential features that led to his negative classification were Marital Status, Relationship and Education. Right: “Mr. Y” has even higher capital gains than Mr. X. Mr. Y is a 27-year-old, with only Preschool education, and is engaged in fishing. Examination of the transparency report reveals that the most influential factor for negative classification for Mr. Y is his Occupation. Interestingly, his low level of education is not considered very important by this classifier. (credit: Anupam Datta et al./2016 P IEEE S SECUR PRIV)

“To get a sense of these influence measures, consider the U.S. presidential election,” said Yair Zick, a post-doctoral researcher in the CMU Computer Science Department. “California and Texas have influence because they have many voters, whereas Pennsylvania and Ohio have power because they are often swing states. The influence aggregation measures we employ account for both kinds of power.”

The researchers tested their approach against some standard machine-learning algorithms that they used to train decision-making systems on real data sets. They found that the QII provided better explanations than standard associative measures for a host of scenarios they considered, including sample applications for predictive policing and income prediction.

Privacy concerns

But transparency reports could also potentially compromise privacy, so in the paper, the researchers also explore the transparency-privacy tradeoff and prove that a number of useful transparency reports can be made differentially private with very little addition of noise.

QII is not yet available, but the CMU researchers are seeking collaboration with industrial partners so that they can employ QII at scale on operational machine-learning systems.

Abstract of Algorithmic Transparency via Quantitative Input Influence: Theory and Experiments with Learning Systems

Algorithmic systems that employ machine learning play an increasing role in making substantive decisions in modern society, ranging from online personalization to insurance and credit decisions to predictive policing. But their decision-making processes are often opaque—it is difficult to explain why a certain decision was made. We develop a formal foundation to improve the transparency of such decision-making systems. Specifically, we introduce a family of Quantitative Input Influence (QII) measures that capture the degree of influence of inputs on outputs of systems. These measures provide a foundation for the design of transparency reports that accompany system decisions (e.g., explaining a specific credit decision) and for testing tools useful for internal and external oversight (e.g., to detect algorithmic discrimination). Distinctively, our causal QII measures carefully account for correlated inputs while measuring influence. They support a general class of transparency queries and can, in particular, explain decisions about individuals (e.g., a loan decision) and groups (e.g., disparate impact based on gender). Finally, since single inputs may not always have high influence, the QII measures also quantify the joint influence of a set of inputs (e.g., age and income) on outcomes (e.g. loan decisions) and the marginal influence of individual inputs within such a set (e.g., income). Since a single input may be part of multiple influential sets, the average marginal influence of the input is computed using principled aggregation measures, such as the Shapley value, previously applied to measure influence in voting. Further, since transparency reports could compromise privacy, we explore the transparency-privacy tradeoff and prove that a number of useful transparency reports can be made differentially private with very little addition of noise. Our empirical validation with standard machine learning algorithms demonstrates that QII measures are a useful transparency mechanism when black box access to the learning system is available. In particular, they provide better explanations than standard associative measures for a host of scenarios that we consider. Further, we show that in the situations we consider, QII is efficiently approximable and can be made differentially private while preserving accuracy.

Posted on

Deep learning applied to drug discovery and repurposing

Deep neural networks for drug discovery (credit: Insilico Medicine, Inc.)

Scientists from Insilico Medicine, Inc. have trained deep neural networks (DNNs) to predict the potential therapeutic uses of 678 drugs, using gene-expression data obtained from high-throughput experiments on human cell lines from Broad Institute’s LINCS databases and NIH MeSH databases.

The supervised deep-learning drug-discovery engine used the properties of small molecules, transcriptional data, and literature to predict efficacy, toxicity, tissue-specificity, and heterogeneity of response.

“We used LINCS data from Broad Institute to determine the effects on cell lines before and after incubation with compounds, co-author and research scientist Polina Mamoshina explained to KurzweilIAI.

“We used gene expression data of total mRNA from cell lines extracted and measured before incubation with compound X and after incubation with compound X to identify the response on a molecular level. The goal is to understand how gene expression (the transcriptome) will change after drug uptake. It is a differential value, so we need a reference (molecular state before incubation) to compare.”

The research is described in a paper in the upcoming issue of the journal Molecular Pharmaceutics.

Helping pharmas accelerate R&D

Alex Zhavoronkov, PhD, Insilico Medicine CEO, who coordinated the study, said the initial goal of their research was to help pharmaceutical companies significantly accelerate their R&D and increase the number of approved drugs. “In the process we came up with more than 800 strong hypotheses in oncology, cardiovascular, metabolic, and CNS spaces and started basic validation,” he said.

The team measured the “differential signaling pathway activation score for a large number of pathways to reduce the dimensionality of the data while retaining biological relevance.” They then used those scores to train the deep neural networks.*

“This study is a proof of concept that DNNs can be used to annotate drugs using transcriptional response signatures, but we took this concept to the next level,” said Alex Aliper, president of research, Insilico Medicine, Inc., lead author of the study.

Via Pharma.AI, a newly formed subsidiary of Insilico Medicine, “we developed a pipeline for in silico drug discovery — which has the potential to substantially accelerate the preclinical stage for almost any therapeutic — and came up with a broad list of predictions, with multiple in silico validation steps that, if validated in vitro and in vivo, can almost double the number of drugs in clinical practice.”

Despite the commercial orientation of the companies, the authors agreed not to file for intellectual property on these methods and to publish the proof of concept.

Deep-learning age biomarkers

According to Mamoshina, earlier this month, Insilico Medicine scientists published the first deep-learned biomarker of human age — aiming to predict the health status of the patient — in a paper titled “Deep biomarkers of human aging: Application of deep neural networks to biomarker development” by Putin et al, in Aging; and an overview of recent advances in deep learning in a paper titled “Applications of Deep Learning in Biomedicine” by Mamoshina et al., also in Molecular Pharmaceutics.

Insilico Medicine is located in the Emerging Technology Centers at Johns Hopkins University in Baltimore, Maryland, in collaboration with Datalytic Solutions and Mind Research Network.

* In this study, scientists used the perturbation samples of 678 drugs across A549, MCF-7 and PC-3 cell lines from the Library of Integrated Network-Based Cellular Signatures (LINCS) project developed by the National Institutes of Health (NIH) and linked those to 12 therapeutic use categories derived from MeSH (Medical Subject Headings) developed and maintained by the National Library of Medicine (NLM) of the NIH.

To train the DNN, scientists utilized both gene level transcriptomic data and transcriptomic data processed using a pathway activation scoring algorithm, for a pooled dataset of samples perturbed with different concentrations of the drug for 6 and 24 hours. Cross-validation experiments showed that DNNs achieve 54.6% accuracy in correctly predicting one out of 12 therapeutic classes for each drug.

One peculiar finding of this experiment was that a large number of drugs misclassified by the DNNs had dual use, suggesting possible application of DNN confusion matrices in drug repurposing.


FutureTechnologies Media Group | Video presentation Insilico medicine

Abstract of Deep learning applications for predicting pharmacological properties of drugs and drug repurposing using transcriptomic data

Deep learning is rapidly advancing many areas of science and technology with multiple success stories in image, text, voice and video recognition, robotics and autonomous driving. In this paper we demonstrate how deep neural networks (DNN) trained on large transcriptional response data sets can classify various drugs to therapeutic categories solely based on their transcriptional profiles. We used the perturbation samples of 678 drugs across A549, MCF-7 and PC-3 cell lines from the LINCS project and linked those to 12 therapeutic use categories derived from MeSH. To train the DNN, we utilized both gene level transcriptomic data and transcriptomic data processed using a pathway activation scoring algorithm, for a pooled dataset of samples perturbed with different concentrations of the drug for 6 and 24 hours. When applied to normalized gene expression data for “landmark genes,” DNN showed cross-validation mean F1 scores of 0.397, 0.285 and 0.234 on 3-, 5- and 12-category classification problems, respectively. At the pathway level DNN performed best with cross-validation mean F1 scores of 0.701, 0.596 and 0.546 on the same tasks. In both gene and pathway level classification, DNN convincingly outperformed support vector machine (SVM) model on every multiclass classification problem. For the first time we demonstrate a deep learning neural net trained on transcriptomic data to recognize pharmacological properties of multiple drugs across different biological systems and conditions. We also propose using deep neural net confusion matrices for drug repositioning. This work is a proof of principle for applying deep learning to drug discovery and development.

Posted on

Robots learn to cut through clutter

New software developed by Carnegie Mellon University helps mobile robots deal efficiently with clutter, whether it is in the back of a refrigerator or on the surface of the moon. (credit: Carnegie Mellon University Personal Robotics Lab)

Carnegie Mellon University roboticists have developed an algorithm that helps robots cope with a cluttered world.

Robots are adept at picking up an object in a specified place (such as in a factory assembly line) and putting it down at another specified place (known as “pick-and-place,” or P&P, processes). But homes and other planets, for example, are a special challenge for robots.

When a person reaches for a milk carton in a refrigerator, he doesn’t necessarily move every other item out of the way. Rather, a person might move an item or two, while shoving others out of the way as the carton is pulled out.

Robot creativity

Robot employs a “push and shove” method (credit: Jennifer E. King et al./Proceedings of IEEE International Conference on Robotics and Automation)

In tests, the new “push and shove” algorithm helped a robot deal efficiently with clutter, but surprisingly, it also revealed the robot’s creativity in solving problems.

“It was exploiting sort of superhuman capabilities,” Siddhartha Srinivasa, associate professor of robotics, said of his lab’s two-armed mobile robot, the Home Exploring Robot Butler, or HERB. “The robot’s wrist has a 270-degree range, which led to behaviors we didn’t expect. Sometimes, we’re blinded by our own anthropomorphism.”

In one case, the robot used the crook of its arm to cradle an object to be moved. “We never taught it that,” Srinivasa said.

K-Rex rover prototype (credit: NASA)

The new algorithm was also tested on NASA’s KRex robot, which is being designed to traverse the lunar surface. While HERB focused on clutter typical of a home, KRex used the software to find traversable paths across an obstacle-filled landscape while pushing an object.

A “rearrangement planner” automatically finds a balance between the two strategies (pick-and-place vs. push-and-shove), Srinivasa said, based on the robot’s progress on its task. The robot is programmed to understand the basic physics of its world, so it has some idea of what can be pushed, lifted, or stepped on. And it can be taught to pay attention to items that might be valuable or delicate.

They researchers presented their work last week (May 19) at the IEEE International Conference on Robotics and Automation in Stockholm, Sweden. NASA, the National Science Foundation, Toyota Motor Engineering and Manufacturing, and the Office of Naval Research supported this research.

Abstract of Rearrangement Planning Using Object-Centric and Robot-Centric Action Spaces

This paper addresses the problem of rearrangement planning, i.e. to find a feasible trajectory for a robot that must interact with multiple objects in order to achieve a goal. We propose a planner to solve the rearrangement planning problem by considering two different types of actions: robot-centric and object-centric. Object-centric actions guide the planner to perform specific actions on specific objects. Robot-centric actions move the robot without object relevant intent, easily allowing simultaneous object contact and whole arm interaction. We formulate a hybrid planner that uses both action types. We evaluate the planner on tasks for a mobile robot and a household manipulator.

Posted on

Using animal training techniques to teach robots household chores

Virtual environments for training a robot dog (credit: Washington State University)

Researchers at Washington State University are using ideas from animal training to help non-expert users teach robots how to do desired tasks.

As robots become more pervasive in society, humans will want them to do chores like cleaning house or cooking. But to get a robot started on a task, people who aren’t computer programmers will have to give it instructions. “So we needed to provide a way for everyone to train robots, without programming,” said Matthew Taylor, Allred Distinguished Professor in the WSU School of Electrical Engineering and Computer Science.

User feedback improves robot performance

With Bei Peng, a doctoral student in computer science, and collaborators at Brown University and North Carolina State University, Taylor designed a computer program that lets humans without programming expertise teach a virtual robot that resembles a dog in WSU’s Intelligent Robot Learning Laboratory.

For the study, the researchers varied the speed at which their virtual dog reacted. As when somebody is teaching a new skill to a real animal, the slower movements let the trainer know that the virtual dog was unsure of how to behave, so trainers could provide clearer guidance to help the robot learn better.

The researchers have begun working with physical robots as well as virtual ones. They also hope to eventually also use the program to help people learn to be more effective animal trainers.

The researchers recently presented their work at the international Autonomous Agents and Multiagent Systems conference, a scientific gathering for agents and robotics research. Funding for the project came from a National Science Foundation grant.


Bei Peng/WSU | Dog Training — AAMAS 2016

Abstract of A Need for Speed: Adapting Agent Action Speed to Improve Task Learning from Non-Expert Humans

As robots become pervasive in human environments, it is important to enable users to effectively convey new skills without programming. Most existing work on Interactive Reinforcement Learning focuses on interpreting and incorporating non-expert human feedback to speed up learning; we aim to design a better representation of the learning agent that is able to elicit more natural and effective communication between the human trainer and the learner, while treating human feedback as discrete communication that depends probabilistically on the trainer’s target policy. This work entails a user study where participants train a virtual agent to accomplish tasks by giving reward and/or punishment in a variety of simulated environments. We present results from 60 participants to show how a learner can ground natural language commands and adapt its action execution speed to learn more efficiently from human trainers. The agent’s action execution speed can be successfully modulated to encourage more explicit feedback from a human trainer in areas of the state space where there is high uncertainty. Our results show that our novel adaptive speed agent dominates different fixed speed agents on several measures of performance. Additionally, we investigate the impact of instructions on user performance and user preference in training conditions.