Posted on

Deep neural networks that identify shapes nearly as well as humans

Self-driving car vs. pedestrians (credit: Google)

Deep neural networks (DNNs) are capable of learning to identify shapes, so “we’re on the right track in developing machines with a visual system and vocabulary as flexible and versatile as ours,” say KU Leuven researchers.

“For the first time, a dramatic increase in performance has been observed on object and scene categorization tasks, quickly reaching performance levels rivaling humans,” they note in an open-access paper in PLOS Computational Biology.

Categorization accuracy for models created by three DNNs (CaffeNet, VGG-19, and GoggLeNet) for three types of images (color, grayscaled, silhouette). For each type, mean human performance is indicated by a gray horizontal line, with the gray surrounding band depicting 95% confidence intervals. Error bars (vertical black lines) depict 95% confidence intervals. (credit: J. Kubilius et al./PLoS Comput Biol)

The researchers found that when trained for generic object recognition from natural photographs, several different DNNs developed visual representations that relate closely to human perceptual shape judgments, even though they were never explicitly trained for shape processing.

However, “We’re not there just yet,” say the researchers. “Even if machines will at some point be equipped with a visual system as powerful as ours, self-driving cars would still make occasional mistakes —- although, unlike human drivers, they wouldn’t be distracted because they’re tired or busy texting. However, even in those rare instances when self-driving cars would err, their decisions would be at least as reasonable as ours.”

Abstract of Deep Neural Networks as a Computational Model for Human Shape Sensitivity

Theories of object recognition agree that shape is of primordial importance, but there is no consensus about how shape might be represented, and so far attempts to implement a model of shape perception that would work with realistic stimuli have largely failed. Recent studies suggest that state-of-the-art convolutional ‘deep’ neural networks (DNNs) capture important aspects of human object perception. We hypothesized that these successes might be partially related to a human-like representation of object shape. Here we demonstrate that sensitivity for shape features, characteristic to human and primate vision, emerges in DNNs when trained for generic object recognition from natural photographs. We show that these models explain human shape judgments for several benchmark behavioral and neural stimulus sets on which earlier models mostly failed. In particular, although never explicitly trained for such stimuli, DNNs develop acute sensitivity to minute variations in shape and to non-accidental properties that have long been implicated to form the basis for object recognition. Even more strikingly, when tested with a challenging stimulus set in which shape and category membership are dissociated, the most complex model architectures capture human shape sensitivity as well as some aspects of the category structure that emerges from human judgments. As a whole, these results indicate that convolutional neural networks not only learn physically correct representations of object categories but also develop perceptually accurate representational spaces of shapes. An even more complete model of human object representations might be in sight by training deep architectures for multiple tasks, which is so characteristic in human development.

Posted on

Scientists turn skin cells into heart and brain cells using only drugs — no stem cells required

Neurons created from chemically induced neural stem cells. The cells were created from skin cells that were reprogrammed into neural stem cells using a cocktail of only nine chemicals. This is the first time cellular reprogramming has been accomplished without adding external genes to the cells. (credit: Mingliang Zhang, PhD, Gladstone Institutes)

Scientists at the Gladstone Institutes have used chemicals to transform skin cells into heart cells and brain cells, instead of adding external genes — making this accomplishment a breakthrough, according to the scientists.

The research lays the groundwork for one day being able to regenerate lost or damaged cells directly with pharmaceutical drugs — a more efficient and reliable method to reprogram cells and one that avoids medical concerns surrounding genetic engineering.

Instead, in two studies published in an open-access paper in Science and in Cell Stem Cell, the team of scientists at the Roddenberry Center for Stem Cell Biology and Medicine at Gladstone used chemical cocktails to gradually coax skin cells to change into organ-specific stem-cell-like cells and ultimately into heart or brain cells.

“This method brings us closer to being able to generate new cells at the site of injury in patients,” said Gladstone senior investigator Sheng Ding, PhD, the senior author on both studies. “Our hope is to one day treat diseases like heart failure or Parkinson’s disease with drugs that help the heart and brain regenerate damaged areas from their own existing tissue cells. This process is much closer to the natural regeneration that happens in animals like newts and salamanders, which has long fascinated us.”

Chemically Repaired Hearts

A human heart cell that was chemically reprogrammed from a human skin cell (credit: Nan Cao/Gladstone Institutes)

Transplanted adult heart cells do not survive or integrate properly into the heart and few stem cells can be coaxed into becoming heart cells.

Instead, in the Science study, the researchers used a cocktail of nine chemicals to change human skin cells into beating heart cells. By trial and error, they found the best combination of chemicals to begin the process by changing the cells into a state resembling multipotent stem cells (cells that can turn into many different types of cells in a particular organ). A second cocktail of chemicals and growth factors then helped transition the cells to become heart muscle cells.

With this method, more than 97% of the cells began beating, a characteristic of fully developed, healthy heart cells. The cells also responded appropriately to hormones, and molecularly, they resembled heart muscle cells, not skin cells. What’s more, when the cells were transplanted into a mouse heart early in the process, they developed into healthy-looking heart muscle cells within the organ.

“The ultimate goal in treating heart failure is a robust, reliable way for the heart to create new muscle cells,” said Srivastava, co-senior author on the Science paper. “Reprogramming a patient’s own cells could provide the safest and most efficient way to regenerate dying or diseased heart muscle.”

Rejuvenating the brain with neural stem cell-like cells

In the second study, authored by Gladstone postdoctoral scholar Mingliang Zhang, PhD, and published in Cell Stem Cell, the scientists created neural stem-cell-like cells from mouse skin cells using a similar approach.

The chemical cocktail again consisted of nine molecules, some of which overlapped with those used in the first study. Over ten days, the cocktail changed the identity of the cells, until all of the skin-cell genes were turned off and the genes of the neural stem-cell-like cells were gradually turned on.

When transplanted into mice, the neural stem-cell-like cells spontaneously developed into the three basic types of brain cells: neurons, oligodendrocytes, and astrocytes. The neural stem-cell-like cells were also able to self-replicate, making them ideal for treating neurodegenerative diseases or brain injury.

With their improved safety, these neural stem-cell-like cells could one day be used for cell replacement therapy in neurodegenerative diseases like Parkinson’s disease and Alzheimer’s disease, according to co-senior author Yadong Huang, MD, PhD, a senior investigator at Gladstone. “In the future, we could even imagine treating patients with a drug cocktail that acts on the brain or spinal cord, rejuvenating cells in the brain in real time.”


Gladstone Institutes | Chemically Reprogrammed Beating Heart Cell

Abstract of Conversion of human fibroblasts into functional cardiomyocytes by small molecules

Reprogramming somatic fibroblasts into alternative lineages would provide a promising source of cells for regenerative therapy. However, transdifferentiating human cells to specific homogeneous, functional cell types is challenging. Here we show that cardiomyocyte-like cells can be generated by treating human fibroblasts with a combination of nine compounds (9C). The chemically induced cardiomyocyte-like cells (ciCMs) uniformly contracted and resembled human cardiomyocytes in their transcriptome, epigenetic, and electrophysiological properties. 9C treatment of human fibroblasts resulted in a more open-chromatin conformation at key heart developmental genes, enabling their promoters/enhancers to bind effectors of major cardiogenic signals. When transplanted into infarcted mouse hearts, 9C-treated fibroblasts were efficiently converted to ciCMs. This pharmacological approach for lineage-specific reprogramming may have many important therapeutic implications after further optimization to generate mature cardiac cells.

Abstract of Pharmacological Reprogramming of Fibroblasts into Neural Stem Cells by Signaling-Directed Transcriptional Activation

Cellular reprogramming using chemically defined conditions, without genetic manipulation, is a promising approach for generating clinically relevant cell types for regenerative medicine and drug discovery. However, small-molecule approaches for inducing lineage-specific stem cells from somatic cells across lineage boundaries have been challenging. Here, we report highly efficient reprogramming of mouse fibroblasts into induced neural stem cell-like cells (ciNSLCs) using a cocktail of nine components (M9). The resulting ciNSLCs closely resemble primary neural stem cells molecularly and functionally. Transcriptome analysis revealed that M9 induces a gradual and specific conversion of fibroblasts toward a neural fate. During reprogramming specific transcription factors such as Elk1 and Gli2 that are downstream of M9-induced signaling pathways bind and activate endogenous master neural genes to specify neural identity. Our study provides an effective chemical approach for generating neural stem cells from mouse fibroblasts and reveals mechanistic insights into underlying reprogramming processes.

Posted on

Ultrasound allows for transmitting HD video through animal tissues

HD video transmission through human tissue from implanted medical devices via in-body ultrasonic communications: beef liver and pork loin were used to represent the density and moisture content found in human tissue (credit: UIUC)(credit: UIUC)

University of Illinois at Urbana-Champaign engineers have demonstrated real-time video-rate (>30Mbps) “meat comm” data transmission through tissue, which could mean in-body ultrasonic communications may be possible for implanted medical devices, including hi-def video.

For example, a patient could swallow a miniaturized HD video camera that could stream live to an external screen, with the orientation of the device controlled wirelessly and externally by a physician, according to Andrew Singer, the Fox Family Professor in the Department of Electrical and Computer Engineering at Illinois,

“To our knowledge, this is the first time anyone has ever sent such high data rates through animal tissue,” Singer added. “These data rates are sufficient to allow real-time streaming of high definition video, enough to watch Netflix, for example, and to operate and control small devices within the body.”

Ingestible cameras and other devices

Potential biomedical uses include ingestible cameras for imaging the digestive track, as well as lower-bandwidth devices such as implanted pacemakers and defibrillators, glucose monitors and insulin pumps, intracranial pressure sensors, and epilepsy control.

Currently, most implanted medical devices use RF electromagnetic waves to communicate through the body. The Federal Communications Commission (FCC) regulates the bandwidths that can be used for RF electromagnetic wave propagation available to implanted medical devices. For example, the Medical Device Radiocommunication Service (MDRS) designates frequencies of operation ranging from 401–406 MHz (where these is high absorption). The corresponding maximum bandwidth allowed is 300 kHz and a maximum of 50 kb/s.

The main limitation for using RF electromagnetic waves in the body is loss of signal that occurs because of attenuation in the body. That requires higher power, which can cause tissue damage from heating due to absorption.

“For underwater applications, radio-frequency (RF) electromagnetic communications has long since been supplanted by acoustic communication,” Singer noted. “Acoustic or ultrasonic communication is the preferred communication means underwater because sound (pressure) waves exhibit dramatically lower losses than RF and can propagate tremendous distances for signals of modest bandwidth.”

The study was reported in an open-access paper on arXiv.org. The researchers have received a provisional patent application on the high-definition ultrasonic technology. They will be presenting their findings at the 17th IEEE International Workshop on Signal Processing  Advances in Wireless Communications, this July in Edinburgh, UK.

Abstract of Mbps Experimental Acoustic Through-Tissue Communications: MEAT-COMMS

Methods for digital, phase-coherent acoustic communication date to at least the work of Stojanjovic, et al [20], and the added robustness afforded by improved phase tracking and compensation of Johnson, et al [21]. This work explores the use of such methods for communications through tissue for potential biomedical applications, using the tremendous bandwidth available in commercial medical ultrasound transducers. While long-range ocean acoustic experiments have been at rates of under 100kbps, typically on the order of 1- 10kbps, data rates in excess of 120Mb/s have been achieved over cm-scale distances in ultrasonic testbeds [19]. This paper describes experimental transmission of digital communication signals through samples of real pork tissue and beef liver, achieving data rates of 20-30Mbps, demonstrating the possibility of real-time video-rate data transmission through tissue for inbody ultrasonic communications with implanted medical devices.

Posted on

Just 1 minute of intense exercise produces health benefits similar to 50 minutes of moderate exercise

Intense trumps moderate (credit: iStock)

Researchers at McMaster University have found that a single minute of very intense exercise within a 10-minute session produces health benefits similar to those from 50 minutes of moderate-intensity continuous exercise.

Brief bursts of intense exercise are remarkably effective, a very time-efficient workout strategy, according to Martin Gibala, a professor of kinesiology at McMaster and lead author on the study, published online in an open-access paper in the journal PLOS ONE

Gibala and associates compared their “sprint interval training” (SIT) protocol to moderate-intensity continuous training (MICT), which is recommended in current public-health guidelines. They examined key health indicators, including insulin sensitivity (a measure of how the body regulates blood sugar) and cardiorespiratory fitness.

Quick intense vs. longer moderate

The ”sprint interval training” (SIT) protocol in the experiment involved three intermittent 20-second “all-out” cycle sprints interspersed with two minutes of continuous low-intensity exercise for recovery. MICT (the current exercise guideline) involves 45 minutes of continuous cycling at ~70% maximal heart rate. Both protocols involve a two-minute warm-up and three-minute cool-down.

In the experiment, a total of 27 sedentary men were recruited and assigned to perform three weekly sessions of either intense or moderate training for 12 weeks, or to a control group that did not exercise.

After 12 weeks of training, the results were remarkably similar, even though the MICT protocol involved five times as much exercise and a five-fold greater time commitment. Specifically, the researchers found a strikingly similar 19% improvement in cardiorespiratory fitness as determined by peak oxygen uptake (VO2 peak), which compares favorably with the typical change reported after several months of traditional endurance training (MICT).

“Most people cite ‘lack of time’ as the main reason for not being active,” says Gibala. “Our study shows that an interval-based approach can be more efficient — you can get health and fitness benefits comparable to the traditional approach, in less time. The basic principles apply to many forms of exercise. Climbing a few flights of stairs on your lunch hour can provide a quick and effective workout. The health benefits are significant.”

This project was supported by an operating grant from the Natural Sciences and Engineering Research Council, and an internally-sponsored research grant from McMaster University to MJG.


McMaster | Gibala on HIIT

Abstract of Twelve Weeks of Sprint Interval Training Improves Indices of Cardiometabolic Health Similar to Traditional Endurance Training despite a Five-Fold Lower Exercise Volume and Time Commitment

Aims: We investigated whether sprint interval training (SIT) was a time-efficient exercise strategy to improve insulin sensitivity and other indices of cardiometabolic health to the same extent as traditional moderate-intensity continuous training (MICT). SIT involved 1 minute of intense exercise within a 10-minute time commitment, whereas MICT involved 50 minutes of continuous exercise per session.

Methods: Sedentary men (27±8y; BMI = 26±6kg/m2) performed three weekly sessions of SIT (n = 9) or MICT (n = 10) for 12 weeks or served as non-training controls (n = 6). SIT involved 3×20-second ‘all-out’ cycle sprints (~500W) interspersed with 2 minutes of cycling at 50W, whereas MICT involved 45 minutes of continuous cycling at ~70% maximal heart rate (~110W). Both protocols involved a 2-minute warm-up and 3-minute cool-down at 50W.

Results: Peak oxygen uptake increased after training by 19% in both groups (SIT: 32±7 to 38±8; MICT: 34±6 to 40±8ml/kg/min; p<0.001 for both). Insulin sensitivity index (CSI), determined by intravenous glucose tolerance tests performed before and 72 hours after training, increased similarly after SIT (4.9±2.5 to 7.5±4.7, p = 0.002) and MICT (5.0±3.3 to 6.7±5.0 x 10−4 min-1[μU/mL]-1, p = 0.013) (p<0.05). Skeletal muscle mitochondrial content also increased similarly after SIT and MICT, as primarily reflected by the maximal activity of citrate synthase (CS; P<0.001). The corresponding changes in the control group were small for VO2peak (p = 0.99), CSI (p = 0.63) and CS (p = 0.97).

Conclusions: Twelve weeks of brief intense interval exercise improved indices of cardiometabolic health to the same extent as traditional endurance training in sedentary men, despite a five-fold lower exercise volume and time commitment.

Posted on

Do you trust robots?

Would you trust this robot? (credit: Rethink Robotics)

Trust in robots is a critical component in safety that requires study, says MIT Professor Emeritus Thomas B. Sheridan in an open-access study published in Human Factors journal.

For decades, he has studied humans and automation and in each case, he noted significant human factors challenges — particularly concerning safety. He looked at self-driving cars and highly automated transit systems; routine tasks such as the delivery of packages in Amazon warehouses; devices that handle tasks in hazardous or inaccessible environments, such as the Fukushima nuclear plant; and robots that engage in social interaction (Barbies).

For example, no human driver, he claims, will stay alert to take over control of a self-driving car quickly enough should the automation fail. Nor does self-driving car technology consider the value of social interaction between drivers such as eye contact and hand signals. And would airline passengers be happy if computerized monitoring replaced the second pilot?

Designing a robot to move an elderly person in and out of bed would potentially reduce back injuries among human caregivers, but questions abound as to what physical form that robot should take, and hospital patients may be alienated by robots delivering their food trays. The ability of robots to learn from human feedback is an area that demands human factors research, as is understanding how people of different ages and abilities best learn from robots.

Sheridan also challenges the human factors community to address the inevitable trade-offs: the possibility of robots providing jobs rather than taking them away, robots as assistants that can enhance human self-worth instead of diminishing it, and the role of robots to improve rather than jeopardize security.

Abstract of Human–Robot Interaction: Status and Challenges

Objective: The current status of human–robot interaction (HRI) is reviewed, and key current research challenges for the human factors community are described.

Background: Robots have evolved from continuous human-controlled master–slave servomechanisms for handling nuclear waste to a broad range of robots incorporating artificial intelligence for many applications and under human supervisory control.

Methods: This mini-review describes HRI developments in four application areas and what are the challenges for human factors research.

Results: In addition to a plethora of research papers, evidence of success is manifest in live demonstrations of robot capability under various forms of human control.

Conclusions: HRI is a rapidly evolving field. Specialized robots under human teleoperation have proven successful in hazardous environments and medical application, as have specialized telerobots under human supervisory control for space and repetitive industrial tasks. Research in areas of self-driving cars, intimate collaboration with humans in manipulation tasks, human control of humanoid robots for hazardous environments, and social interaction with robots is at initial stages. The efficacy of humanoid general-purpose robots has yet to be proven.

Applications: HRI is now applied in almost all robot tasks, including manufacturing, space, aviation, undersea, surgery, rehabilitation, agriculture, education, package fetch and delivery, policing, and military operations.

Posted on

Public beta of toolkit for developing machine learning for robots and games released

Make a three-dimensional bipedal robot walk forward as fast as possible, without falling over (credit: OpenAI Gym)

OpenAI (a non-profit AI research company sponsored by Elon Musk and others) has released the public beta of OpenAI Gym, a toolkit for developing and comparing algorithms for reinforcement learning (RL), a type of machine learning.

OpenAI Gym consists of a growing suite of environments (from simulated robots to Atari games), and a site for comparing and reproducing results. OpenAI Gym is compatible with algorithms written in any framework, such as Tensorflow and Theano. The environments are initially written in Python (other languages planned).

If you’d like to dive in right away, you can work through a tutorial, and you help out while learning by reproducing a result.

What is reinforcement learning?

Reinforcement learning (RL) is the subfield of machine learning concerned with decision making and motor control. It studies how an agent can learn how to achieve goals in a complex, uncertain environment. It’s exciting for two reasons, according to OpenAI’s Greg Brockman and John Schulman:

  • RL is very general, encompassing all problems that involve making a sequence of decisions: for example, controlling a robot’s motors so that it’s able to run and jump, making business decisions like pricing and inventory management, or playing video games and board games. RL can even be applied to supervised learning problems with sequential or structured outputs.
  • RL algorithms have started to achieve good results in many difficult environments. RL has a long history, but until recent advances in deep learning, it required lots of problem-specific engineering. DeepMind’s Atari results, BRETT from Pieter Abbeel’s group, and AlphaGo all used deep RL algorithms, which did not make too many assumptions about their environment, and thus can be applied in other settings.

However, RL research is also slowed down by two factors:

  • The need for better benchmarks. In supervised (human-managed) learning, progress has been driven by large labeled datasets like ImageNet. In RL, the closest equivalent would be a large and diverse collection of environments. However, the existing open-source collections of RL environments don’t have enough variety, and they are often difficult to even set up and use.
  • Lack of standardization of environments used in publications. Subtle differences in the problem definition, such as the reward function or the set of actions, can drastically alter a task’s difficulty. This issue makes it difficult to reproduce published research and compare results from different papers.

OpenAI Gym is an attempt to fix both problems.

Partners include:

More information, including enviroments (Atari games, 2D and 3D robots, and toy text, for example) is available here.

“During the public beta, we’re looking for feedback on how to make this into an even better tool for research,” says the OpenAI team. “If you’d like to help, you can try your hand at improving the state-of-the-art on each environment, reproducing other people’s results, or even implementing your own environments. Also please join us in the community chat!


John Schulman | hopper

Posted on

Artificial protein controls first self-assembly of C60 fullerenes

Buckminsterfullerene (C60), aka fullerene and buckyball (credit: St Stev via Foter.com / CC BY-NC-ND)

A Dartmouth College scientist and his collaborators* have created the first high-resolution co-assembly between a protein and buckminsterfullerene (C60), aka fullerene and buckyball (a sphere-like molecule composed of 60 carbon atoms and shaped like a soccer ball).

“This is a proof-of-principle study demonstrating that proteins can be used as effective vehicles for organizing nanomaterials by design,” says senior author Gevorg Grigoryan, an assistant professor of computer science at Dartmouth and senior author of a study discussed in an open-access paper in the journal in Nature Communications.

Proteins organize and orchestrate essentially all molecular processes in our cells. The goal of the new study was to create a new artificial protein that can direct the self-assembly of fullerene into ordered superstructures.

COP, a stable tetramer (a polymer derived from four identical single molecule) in isolation, interacts with C60 (fullerene) molecules via a surface-binding site and further self-assembles into a co-crystalline array called C60Sol–COP (credit: Kook-Han Kim et al./Nature Communications)

Grigoryan and his colleagues show that that their artificial protein organizes a fullerene into a lattice called C60Sol–COP. COP, a protein that is a stable tetramer (a polymer derived from four identical single molecules), interacted with fullerene molecules via a surface-binding site and further self-assembled into an ordered crystalline superstructure. Interestingly, the superstructure exhibits high charge conductance, whereas both the protein-alone crystal and amorphous C60 are electrically insulating.

Grigoryan says that if we learn to do the programmable self-assembly of precisely organized molecular building blocks more generally, it will lead to a range of new materials with properties such as higher strength, lighter weight, and greater chemical reactivity, resulting in a host of applications, from medicine to energy and electronics.

Fullerenes are currently used in nanotechnology because of their high heat resistance and electrical superconductivity (when doped), but the molecule has been difficult to organize in useful ways.

* The study also included researchers from Dartmouth College, Sungkyunkwan University, the New Jersey Institute of Technology, the National Institute of Science Education and Research, the University of California-San Francisco, the University of Pennsylvania, and the Institute for Basic Science.

Abstract of Protein-directed self-assembly of a fullerene crystal

Learning to engineer self-assembly would enable the precise organization of molecules by design to create matter with tailored properties. Here we demonstrate that proteins can direct the self-assembly of buckminsterfullerene (C60) into ordered superstructures. A previously engineered tetrameric helical bundle binds C60 in solution, rendering it water soluble. Two tetramers associate with one C60, promoting further organization revealed in a 1.67-Å crystal structure. Fullerene groups occupy periodic lattice sites, sandwiched between two Tyr residues from adjacent tetramers. Strikingly, the assembly exhibits high charge conductance, whereas both the protein-alone crystal and amorphous C60 are electrically insulating. The affinity of C60 for its crystal-binding site is estimated to be in the nanomolar range, with lattices of known protein crystals geometrically compatible with incorporating the motif. Taken together, these findings suggest a new means of organizing fullerene molecules into a rich variety of lattices to generate new properties by design.

Posted on

Micro-needle insertion into hippocampus stimulates brain regeneration in animal model of AD

Micro-needle insertion results. Left panel is the control side (no lesion placement) and the right panel (the lesioned side) showed increased expression of DCX+ cells (neuronal precursors) in the SGZ (a brain region in the hippocampus where adult neurogenesis occurs) (scale bar= 20 μm). (credit: Shijie Song et al./Cell Transplantation)

Sticking a needle into the hippocampus of mice modeled with Alzheimer’s disease (AD) improved performance on memory tasks, stimulated regenerative activity, and reduced β-amyloid plaques (a hallmark of AD). This area was chosen because the early and primary damage by AD appears to take place in the hippocampus.

Until recently, many diseases of the central nervous system could not be treated by this method because of inaccessibility of the brain to micro-needles, said the researchers.

“Because Alzheimer’s disease is increasing in prevalence, new intervention strategies are becoming invaluable,” said Dr. Shinn-Zong Lin, professor of Neurosurgery at China Medical University Hospital in TaiChung, Taiwan and Co-Editor-in-Chief for Cell Transplantation. “Since the host’s microenvironment can be inhospitable to transplanted cells and pharmacological interventions in diseased conditions, strategies to increase the regenerative capacity of the patient’s own body may be another viable option. Future studies should strive to include a larger sample size in order to validate this approach.”

The study will be published in a future issue of Cell Transplantation and is currently available open-access as an unedited, early epub.

Abstract of Transient Micro-needle Insertion into Hippocampus Triggers Neurogenesis and Decreases Amyloid Burden in a Mouse Model of Alzheimer’s Disease

Targeted micro-lesions of the hippocampus have been reported to enhance neurogenesis in the sub-granular zone (SGZ). The potential therapeutic impact of transient insertion of a micro-needle was investigated in a mouse model of Alzheimer’s disease (AD). Here we tested the hypothesis that transient micro-injury to the brain elicits cellular responses that mediate beneficial regenerative processes. Brief stereotaxic insertion and removal of a micro-needle into the right hippocampus of 14 month old APP/PS1 mice brain resulted in a) stimulation of hippocampal neurogenesis and b) reduction of beta-amyloid plaque number in the CA-1 region. This treatment also resulted in a trend towards improved performance in the radial arm water maze (RAWM). Further studies of fundamental cellular mechanisms of the brain’s response to micro-injury will be useful for investigation of potential neuro-protective and deleterious effects of targeted micro-lesions and deep brain stimulation in Alzheimer Disease (AD).

Posted on

System predicts 85 percent of cyber attacks using input from human experts

AI2 combs through data and detects suspicious activity using unsupervised machine-learning. It then presents this activity to human analysts, who confirm which events are actual attacks, and incorporate that feedback into its models for the next set of data. (credit: Kalyan Veeramachaneni/MIT CSAIL)

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the machine-learning startup PatternEx have developed an AI platform called AI2 that predicts cyber-attacks significantly better than existing systems by continuously incorporating input from human experts (AI2 refers to merging AI with “analyst intuition”:  rules created by living experts).

The team showed that AI2 can detect 85 percent of attacks —about three times better than previous benchmarks — while also reducing the number of false positives by a factor of 5. The system was tested on 3.6 billion pieces of data known as “log lines,” which were generated by millions of users over a period of three months.

To predict attacks, AI2 combs through data and detects suspicious activity by clustering the data into meaningful patterns using unsupervised (automatic, no human help) machine learning. It then presents this activity to human analysts, who confirm which events are actual attacks, and incorporates that feedback into its models for the next set of data.

“You can think about the system as a virtual analyst,” says CSAIL research scientist Kalyan Veeramachaneni, who developed AI2 with Ignacio Arnaldo, a chief data scientist at PatternEx and a former CSAIL postdoc. “It continuously generates new models that it can refine in as little as a few hours, meaning it can improve its detection rates significantly and rapidly.”

Creating cybersecurity systems that merge human- and computer-based approaches is tricky, partly because of the challenge of manually labeling cybersecurity data for the algorithms. For example, let’s say you want to develop a computer-vision algorithm that can identify objects with high accuracy. Labeling data for that is simple: Just enlist a few human volunteers to label photos as either “objects” or “non-objects,” and feed that data into the algorithm.

But for a cybersecurity task, the average person on a crowdsourcing site like Amazon Mechanical Turk simply doesn’t have the skillset to apply labels like “DDOS” or “exfiltration attacks,” says Veeramachaneni. “You need security experts.” That opens up another problem: Experts are busy and expensive, so an effective machine-learning system has to be able to improve itself without overwhelming its human overlords.

Merging methods

AI2’s secret weapon is that it fuses together three different unsupervised-learning methods, and then shows the top events to analysts for them to label. It then builds a supervised model that it can constantly refine through what the team calls a “continuous active learning system.”

Specifically, on day one of its training, AI2 picks the 200 most abnormal events and gives them to the expert. As it improves over time, it identifies more and more of the events as actual attacks, meaning that in a matter of days, the analyst may only be looking at 30 or 40 events a day.

“This paper brings together the strengths of analyst intuition and machine learning, and ultimately drives down both false positives and false negatives,” says Nitesh Chawla, the Frank M. Freimann Professor of Computer Science at the University of Notre Dame. “This research has the potential to become a line of defense against attacks such as fraud, service abuse and account takeover, which are major challenges faced by consumer-facing systems.”

The team says that AI2 can scale to billions of log lines per day, transforming the pieces of data on a minute-by-minute basis into different “features”, or discrete types of behavior that are eventually deemed “normal” or “abnormal.”

“The more attacks the system detects, the more analyst feedback it receives, which, in turn, improves the accuracy of future predictions,” Veeramachaneni says. “That human-machine interaction creates a beautiful, cascading effect.”

Veeramachaneni presented a paper about the system at last week’s IEEE International Conference on Big Data Security in New York City.


MITCSAIL | AI2: an AI-driven predictive cybersecurity platform

Abstract of AI2 : Training a big data machine to defend

We present an analyst-in-the-loop security system, where analyst intuition is put together with stateof-the-art machine learning to build an end-to-end active learning system. The system has four key features: a big data behavioral analytics platform, an ensemble of outlier detection methods, a mechanism to obtain feedback from security analysts, and a supervised learning module. When these four components are run in conjunction on a daily basis and are compared to an unsupervised outlier detection method, detection rate improves by an average of 3.41×, and false positives are reduced fivefold. We validate our system with a real-world data set consisting of 3.6 billion log lines. These results show that our system is capable of learning to defend against unseen attacks.

Posted on

Machine learning rivals human skills in cancer detection

Two announcements yesterday (April 21) suggest that deep learning algorithms rival human skills in detecting cancer from ultrasound images and in identifying cancer in pathology reports.

Samsung Medison RS80A ultrasound imaging system (credit: Samsung)

Samsung Medison, a global medical equipment company and an affiliate of Samsung Electronics, has just updated its RS80A ultrasound imaging system with a deep learning algorithm for breast-lesion analysis.

The “S-Detect for Breast” feature uses big data collected from breast-exam cases and recommends whether the selected lesion is benign or malignant. It’s used in in lesion segmentation, characteristic analysis, and assessment processes, providing “more accurate results.”

“We saw a high level of conformity from analyzing and detecting lesion in various cases by using the S-Detect,” said professor Han Boo Kyung, a radiologist at Samsung Medical Center.

“Users can reduce taking unnecessary biopsies and doctors-in-training will likely have more reliable support in accurately detecting malignant and suspicious lesions.”

Deep learning is better than humans in extracting meaning from cancer pathology reports

Meanwhile, researchers from the Regenstrief Institute and Indiana University School of Informatics and Computing at Indiana University-Purdue University Indianapolis say they’ve found that open-source machine learning tools are as good as — or better than — humans in extracting crucial meaning from free-text (unstructured) pathology reports and detecting cancer cases. The computer tools are also faster and less resource-intensive.

(U.S. states require cancer cases to be reported to statewide cancer registries for disease tracking, identification of at-risk populations, and recognition of unusual trends or clusters. This free-text information can be difficult for health officials to interpret, which can further delay health department action, when action is needed.)

“We think that its no longer necessary for humans to spend time reviewing text reports to determine if cancer is present or not,” said study senior author Shaun Grannis*, M.D., M.S., interim director of the Regenstrief Center of Biomedical Informatics.

Awash in oceans of data

“We have come to the point in time that technology can handle this. A human’s time is better spent helping other humans by providing them with better clinical care. Everything — physician practices, health care systems, health information exchanges, insurers, as well as public health departments — are awash in oceans of data. How can we hope to make sense of this deluge of data? Humans can’t do it — but computers can.”

This is especially relevant for underserved nations, where a majority of clinical data is collected in the form of unstructured free text, he said.

The researchers sampled 7,000 free-text pathology reports from over 30 hospitals that participate in the Indiana Health Information Exchange and used open source tools, classification algorithms, and varying feature selection approaches to predict if a report was positive or negative for cancer. The results indicated that a fully automated review yielded results similar or better than those of trained human reviewers, saving both time and money.

Major infrastructure advance

“We found that artificial intelligence was as least as accurate as humans in identifying cancer cases from free-text clinical data. For example the computer ‘learned’ that the word ‘sheet’ or ‘sheets’ signified cancer as ‘sheet’ or ‘sheets of cells’ are used in pathology reports to indicate malignancy.

“This is not an advance in ideas; it’s a major infrastructure advance — we have the technology, we have the data, we have the software from which we saw accurate, rapid review of vast amounts of data without human oversight or supervision.”

The study was published in the April 2016 issue of the Journal of Biomedical Informatics. It was conducted with support from the Centers for Disease Control and Prevention.

Co-authors of the study include researchers at the IU Fairbanks School of Public Health, the IU School of Medicine and the School of Science at IUPUI.

* Grannis, a Regenstrief Institute investigator and an associate professor of family medicine at the IU School of Medicine, is the architect of the Regenstrief syndromic surveillance detector for communicable diseases and led the technical implementation of Indiana’s Public Health Emergency Surveillance System — one of the nation’s largest. Studies over the past decade have shown that this system detects outbreaks of communicable diseases seven to nine days earlier and finds four times as many cases as human reporting while providing more complete data.

Yann Lecun is Director of AI Research, Facebook and a noted deep-learning expert.