Tag Archives: learning

Organismic learning—learning to forget

This approach to mimicking the human brain differs from the memristor. (You can find several pieces about memrisors here including this August 24, 2017 post about a derivative, a neuristor).  This approach comes from scientists at Purdue University and employs a quantum material. From an Aug. 15, 2017 news item on phys.org,

A new computing technology called “organismoids” mimics some aspects of human thought by learning how to forget unimportant memories while retaining more vital ones.

“The human brain is capable of continuous lifelong learning,” said Kaushik Roy, Purdue University’s Edward G. Tiedemann Jr. Distinguished Professor of Electrical and Computer Engineering. “And it does this partially by forgetting some information that is not critical. I learn slowly, but I keep forgetting other things along the way, so there is a graceful degradation in my accuracy of detecting things that are old. What we are trying to do is mimic that behavior of the brain to a certain extent, to create computers that not only learn new information but that also learn what to forget.”

The work was performed by researchers at Purdue, Rutgers University, the Massachusetts Institute of Technology, Brookhaven National Laboratory and Argonne National Laboratory.

Central to the research is a ceramic “quantum material” called samarium nickelate, which was used to create devices called organismoids, said Shriram Ramanathan, a Purdue professor of materials engineering.

A video describing the work has been produced,

An August 14, 2017 Purdue University news release by Emil Venere, which originated the news item,  details the work,

“These devices possess certain characteristics of living beings and enable us to advance new learning algorithms that mimic some aspects of the human brain,” Roy said. “The results have far reaching implications for the fields of quantum materials as well as brain-inspired computing.”

When exposed to hydrogen gas, the material undergoes a massive resistance change, as its crystal lattice is “doped” by hydrogen atoms. The material is said to breathe, expanding when hydrogen is added and contracting when the hydrogen is removed.

“The main thing about the material is that when this breathes in hydrogen there is a spectacular quantum mechanical effect that allows the resistance to change by orders of magnitude,” Ramanathan said. “This is very unusual, and the effect is reversible because this dopant can be weakly attached to the lattice, so if you remove the hydrogen from the environment you can change the electrical resistance.”

When hydrogen is exposed to the material, it splits into a proton and an electron, and the electron attaches to the nickel, temporarily causing the material to become an insulator.

“Then, when the hydrogen comes out, this material becomes conducting again,” Ramanathan said. “What we show in this paper is the extent of conduction and insulation can be very carefully tuned.”

This changing conductance and the “decay of that conductance over time” is similar to a key animal behavior called habituation.

“Many animals, even organisms that don’t have a brain, possess this fundamental survival skill,” Roy said. “And that’s why we call this organismic behavior. If I see certain information on a regular basis, I get habituated, retaining memory of it. But if I haven’t seen such information over a long time, then it slowly starts decaying. So, the behavior of conductance going up and down in exponential fashion can be used to create a new computing model that will incrementally learn and at same time forget things in a proper way.”

The researchers have developed a “neural learning model” they have termed adaptive synaptic plasticity.

“This could be really important because it’s one of the first examples of using quantum materials directly for solving a major problem in neural learning,” Ramanathan said.

The researchers used the organismoids to implement the new model for synaptic plasticity.

“Using this effect we are able to model something that is a real problem in neuromorphic computing,” Roy said. “For example, if I have learned your facial features I can still go out and learn someone else’s features without really forgetting yours. However, this is difficult for computing models to do. When learning your features, they can forget the features of the original person, a problem called catastrophic forgetting.”

Neuromorphic computing is not intended to replace conventional general-purpose computer hardware, based on complementary metal-oxide-semiconductor transistors, or CMOS. Instead, it is expected to work in conjunction with CMOS-based computing. Whereas CMOS technology is especially adept at performing complex mathematical computations, neuromorphic computing might be able to perform roles such as facial recognition, reasoning and human-like decision making.

Roy’s team performed the research work on the plasticity model, and other collaborators concentrated on the physics of how to explain the process of doping-driven change in conductance central to the paper. The multidisciplinary team includes experts in materials, electrical engineering, physics, and algorithms.

“It’s not often that a materials science person can talk to a circuits person like professor Roy and come up with something meaningful,” Ramanathan said.

Organismoids might have applications in the emerging field of spintronics. Conventional computers use the presence and absence of an electric charge to represent ones and zeroes in a binary code needed to carry out computations. Spintronics, however, uses the “spin state” of electrons to represent ones and zeros.

It could bring circuits that resemble biological neurons and synapses in a compact design not possible with CMOS circuits. Whereas it would take many CMOS devices to mimic a neuron or synapse, it might take only a single spintronic device.

In future work, the researchers may demonstrate how to achieve habituation in an integrated circuit instead of exposing the material to hydrogen gas.

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

Habituation based synaptic plasticity and organismic learning in a quantum perovskite by Fan Zuo, Priyadarshini Panda, Michele Kotiuga, Jiarui Li, Mingu Kang, Claudio Mazzoli, Hua Zhou, Andi Barbour, Stuart Wilkins, Badri Narayanan, Mathew Cherukara, Zhen Zhang, Subramanian K. R. S. Sankaranarayanan, Riccardo Comin, Karin M. Rabe, Kaushik Roy, & Shriram Ramanathan. Nature Communications 8, Article number: 240 (2017) doi:10.1038/s41467-017-00248-6 Published online: 14 August 2017

This paper is open access.

Predicting how a memristor functions

An April 3, 2017 news item on Nanowerk announces a new memristor development (Note: A link has been removed),

Researchers from the CNRS [Centre national de la recherche scientifique; France] , Thales, and the Universities of Bordeaux, Paris-Sud, and Evry have created an artificial synapse capable of learning autonomously. They were also able to model the device, which is essential for developing more complex circuits. The research was published in Nature Communications (“Learning through ferroelectric domain dynamics in solid-state synapses”)

An April 3, 2017 CNRS press release, which originated the news item, provides a nice introduction to the memristor concept before providing a few more details about this latest work (Note: A link has been removed),

One of the goals of biomimetics is to take inspiration from the functioning of the brain [also known as neuromorphic engineering or neuromorphic computing] in order to design increasingly intelligent machines. This principle is already at work in information technology, in the form of the algorithms used for completing certain tasks, such as image recognition; this, for instance, is what Facebook uses to identify photos. However, the procedure consumes a lot of energy. Vincent Garcia (Unité mixte de physique CNRS/Thales) and his colleagues have just taken a step forward in this area by creating directly on a chip an artificial synapse that is capable of learning. They have also developed a physical model that explains this learning capacity. This discovery opens the way to creating a network of synapses and hence intelligent systems requiring less time and energy.

Our brain’s learning process is linked to our synapses, which serve as connections between our neurons. The more the synapse is stimulated, the more the connection is reinforced and learning improved. Researchers took inspiration from this mechanism to design an artificial synapse, called a memristor. This electronic nanocomponent consists of a thin ferroelectric layer sandwiched between two electrodes, and whose resistance can be tuned using voltage pulses similar to those in neurons. If the resistance is low the synaptic connection will be strong, and if the resistance is high the connection will be weak. This capacity to adapt its resistance enables the synapse to learn.

Although research focusing on these artificial synapses is central to the concerns of many laboratories, the functioning of these devices remained largely unknown. The researchers have succeeded, for the first time, in developing a physical model able to predict how they function. This understanding of the process will make it possible to create more complex systems, such as a series of artificial neurons interconnected by these memristors.

As part of the ULPEC H2020 European project, this discovery will be used for real-time shape recognition using an innovative camera1 : the pixels remain inactive, except when they see a change in the angle of vision. The data processing procedure will require less energy, and will take less time to detect the selected objects. The research involved teams from the CNRS/Thales physics joint research unit, the Laboratoire de l’intégration du matériau au système (CNRS/Université de Bordeaux/Bordeaux INP), the University of Arkansas (US), the Centre de nanosciences et nanotechnologies (CNRS/Université Paris-Sud), the Université d’Evry, and Thales.

 

Image synapse


© Sören Boyn / CNRS/Thales physics joint research unit.

Artist’s impression of the electronic synapse: the particles represent electrons circulating through oxide, by analogy with neurotransmitters in biological synapses. The flow of electrons depends on the oxide’s ferroelectric domain structure, which is controlled by electric voltage pulses.


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

Learning through ferroelectric domain dynamics in solid-state synapses by Sören Boyn, Julie Grollier, Gwendal Lecerf, Bin Xu, Nicolas Locatelli, Stéphane Fusil, Stéphanie Girod, Cécile Carrétéro, Karin Garcia, Stéphane Xavier, Jean Tomas, Laurent Bellaiche, Manuel Bibes, Agnès Barthélémy, Sylvain Saïghi, & Vincent Garcia. Nature Communications 8, Article number: 14736 (2017) doi:10.1038/ncomms14736 Published online: 03 April 2017

This paper is open access.

Thales or Thales Group is a French company, from its Wikipedia entry (Note: Links have been removed),

Thales Group (French: [talɛs]) is a French multinational company that designs and builds electrical systems and provides services for the aerospace, defence, transportation and security markets. Its headquarters are in La Défense[2] (the business district of Paris), and its stock is listed on the Euronext Paris.

The company changed its name to Thales (from the Greek philosopher Thales,[3] pronounced [talɛs] reflecting its pronunciation in French) from Thomson-CSF in December 2000 shortly after the £1.3 billion acquisition of Racal Electronics plc, a UK defence electronics group. It is partially state-owned by the French government,[4] and has operations in more than 56 countries. It has 64,000 employees and generated €14.9 billion in revenues in 2016. The Group is ranked as the 475th largest company in the world by Fortune 500 Global.[5] It is also the 10th largest defence contractor in the world[6] and 55% of its total sales are military sales.[4]

The ULPEC (Ultra-Low Power Event-Based Camera) H2020 [Horizon 2020 funded) European project can be found here,

The long term goal of ULPEC is to develop advanced vision applications with ultra-low power requirements and ultra-low latency. The output of the ULPEC project is a demonstrator connecting a neuromorphic event-based camera to a high speed ultra-low power consumption asynchronous visual data processing system (Spiking Neural Network with memristive synapses). Although ULPEC device aims to reach TRL 4, it is a highly application-oriented project: prospective use cases will b…

Finally, for anyone curious about Thales, the philosopher (from his Wikipedia entry), Note: Links have been removed,

Thales of Miletus (/ˈθeɪliːz/; Greek: Θαλῆς (ὁ Μῑλήσιος), Thalēs; c. 624 – c. 546 BC) was a pre-Socratic Greek/Phoenician philosopher, mathematician and astronomer from Miletus in Asia Minor (present-day Milet in Turkey). He was one of the Seven Sages of Greece. Many, most notably Aristotle, regard him as the first philosopher in the Greek tradition,[1][2] and he is otherwise historically recognized as the first individual in Western civilization known to have entertained and engaged in scientific philosophy.[3][4]

Removing gender-based stereotypes from algorithms

Most people don’t think of algorithms as having biases and stereotypes but Michael Zou in his Sept. 26, 2016 essay for The Conversation (h/t phys.org Sept. 26, 2016 news item) says different, Note: Links have been removed,

Machine learning is ubiquitous in our daily lives. Every time we talk to our smartphones, search for images or ask for restaurant recommendations, we are interacting with machine learning algorithms. They take as input large amounts of raw data, like the entire text of an encyclopedia, or the entire archives of a newspaper, and analyze the information to extract patterns that might not be visible to human analysts. But when these large data sets include social bias, the machines learn that too.

A machine learning algorithm is like a newborn baby that has been given millions of books to read without being taught the alphabet or knowing any words or grammar. The power of this type of information processing is impressive, but there is a problem. When it takes in the text data, a computer observes relationships between words based on various factors, including how often they are used together.

We can test how well the word relationships are identified by using analogy puzzles. Suppose I ask the system to complete the analogy “He is to King as She is to X.” If the system comes back with “Queen,” then we would say it is successful, because it returns the same answer a human would.

Our research group trained the system on Google News articles, and then asked it to complete a different analogy: “Man is to Computer Programmer as Woman is to X.” The answer came back: “Homemaker.”

Zou explains how a machine (algorithm) learns and then notes this,

Not only can the algorithm reflect society’s biases – demonstrating how much those biases are contained in the input data – but the system can potentially amplify gender stereotypes. Suppose I search for “computer programmer” and the search program uses a gender-biased database that associates that term more closely with a man than a woman.

The search results could come back flawed by the bias. Because “John” as a male name is more closely related to “computer programmer” than the female name “Mary” in the biased data set, the search program could evaluate John’s website as more relevant to the search than Mary’s – even if the two websites are identical except for the names and gender pronouns.

It’s true that the biased data set could actually reflect factual reality – perhaps there are more “Johns” who are programmers than there are “Marys” – and the algorithms simply capture these biases. This does not absolve the responsibility of machine learning in combating potentially harmful stereotypes. The biased results would not just repeat but could even boost the statistical bias that most programmers are male, by moving the few female programmers lower in the search results. It’s useful and important to have an alternative that’s not biased.

There is a way according to Zou that stereotypes can be removed,

Our debiasing system uses real people to identify examples of the types of connections that are appropriate (brother/sister, king/queen) and those that should be removed. Then, using these human-generated distinctions, we quantified the degree to which gender was a factor in those word choices – as opposed to, say, family relationships or words relating to royalty.

Next we told our machine-learning algorithm to remove the gender factor from the connections in the embedding. This removes the biased stereotypes without reducing the overall usefulness of the embedding.

When that is done, we found that the machine learning algorithm no longer exhibits blatant gender stereotypes. We are investigating applying related ideas to remove other types of biases in the embedding, such as racial or cultural stereotypes.

If you have time, I encourage you to read the essay in its entirety and this June 14, 2016 posting about research into algorithms and how they make decisions for you about credit, medical diagnoses, job opportunities and more.

There’s also an Oct. 24, 2016 article by Michael Light on Salon.com on the topic (Note: Links have been removed),

In a recent book that was longlisted for the National Book Award, Cathy O’Neil, a data scientist, blogger and former hedge-fund quant, details a number of flawed algorithms to which we have given incredible power — she calls them “Weapons of Math Destruction.” We have entrusted these WMDs to make important, potentially life-altering decisions, yet in many cases, they embed human race and class biases; in other cases, they don’t function at all.
Among other examples, O’Neil examines a “value-added” model New York City used to decide which teachers to fire, even though, she writes, the algorithm was useless, functioning essentially as a random number generator, arbitrarily ending careers. She looks at models put to use by judges to assign recidivism scores to inmates that ended up having a racist inclination. And she looks at how algorithms are contributing to American partisanship, allowing political operatives to target voters with information that plays to their existing biases and fears.

I recommend reading Light’s article in its entirety.

Memristor-based electronic synapses for neural networks

Caption: Neuron connections in biological neural networks. Credit: MIPT press office

Caption: Neuron connections in biological neural networks. Credit: MIPT press office

Russian scientists have recently published a paper about neural networks and electronic synapses based on ‘thin film’ memristors according to an April 19, 2016 news item on Nanowerk,

A team of scientists from the Moscow Institute of Physics and Technology (MIPT) have created prototypes of “electronic synapses” based on ultra-thin films of hafnium oxide (HfO2). These prototypes could potentially be used in fundamentally new computing systems.

An April 20, 2016 MIPT press release (also on EurekAlert), which originated the news item (the date inconsistency likely due to timezone differences) explains the connection between thin films and memristors,

The group of researchers from MIPT have made HfO2-based memristors measuring just 40×40 nm2. The nanostructures they built exhibit properties similar to biological synapses. Using newly developed technology, the memristors were integrated in matrices: in the future this technology may be used to design computers that function similar to biological neural networks.

Memristors (resistors with memory) are devices that are able to change their state (conductivity) depending on the charge passing through them, and they therefore have a memory of their “history”. In this study, the scientists used devices based on thin-film hafnium oxide, a material that is already used in the production of modern processors. This means that this new lab technology could, if required, easily be used in industrial processes.

“In a simpler version, memristors are promising binary non-volatile memory cells, in which information is written by switching the electric resistance – from high to low and back again. What we are trying to demonstrate are much more complex functions of memristors – that they behave similar to biological synapses,” said Yury Matveyev, the corresponding author of the paper, and senior researcher of MIPT’s Laboratory of Functional Materials and Devices for Nanoelectronics, commenting on the study.

The press release offers a description of biological synapses and their relationship to learning and memory,

A synapse is point of connection between neurons, the main function of which is to transmit a signal (a spike – a particular type of signal, see fig. 2) from one neuron to another. Each neuron may have thousands of synapses, i.e. connect with a large number of other neurons. This means that information can be processed in parallel, rather than sequentially (as in modern computers). This is the reason why “living” neural networks are so immensely effective both in terms of speed and energy consumption in solving large range of tasks, such as image / voice recognition, etc.

Over time, synapses may change their “weight”, i.e. their ability to transmit a signal. This property is believed to be the key to understanding the learning and memory functions of thebrain.

From the physical point of view, synaptic “memory” and “learning” in the brain can be interpreted as follows: the neural connection possesses a certain “conductivity”, which is determined by the previous “history” of signals that have passed through the connection. If a synapse transmits a signal from one neuron to another, we can say that it has high “conductivity”, and if it does not, we say it has low “conductivity”. However, synapses do not simply function in on/off mode; they can have any intermediate “weight” (intermediate conductivity value). Accordingly, if we want to simulate them using certain devices, these devices will also have to have analogous characteristics.

The researchers have provided an illustration of a biological synapse,

Fig.2 The type of electrical signal transmitted by neurons (a “spike”). The red lines are various other biological signals, the black line is the averaged signal. Source: MIPT press office

Fig.2 The type of electrical signal transmitted by neurons (a “spike”). The red lines are various other biological signals, the black line is the averaged signal. Source: MIPT press office

Now, the press release ties the memristor information together with the biological synapse information to describe the new work at the MIPT,

As in a biological synapse, the value of the electrical conductivity of a memristor is the result of its previous “life” – from the moment it was made.

There is a number of physical effects that can be exploited to design memristors. In this study, the authors used devices based on ultrathin-film hafnium oxide, which exhibit the effect of soft (reversible) electrical breakdown under an applied external electric field. Most often, these devices use only two different states encoding logic zero and one. However, in order to simulate biological synapses, a continuous spectrum of conductivities had to be used in the devices.

“The detailed physical mechanism behind the function of the memristors in question is still debated. However, the qualitative model is as follows: in the metal–ultrathin oxide–metal structure, charged point defects, such as vacancies of oxygen atoms, are formed and move around in the oxide layer when exposed to an electric field. It is these defects that are responsible for the reversible change in the conductivity of the oxide layer,” says the co-author of the paper and researcher of MIPT’s Laboratory of Functional Materials and Devices for Nanoelectronics, Sergey Zakharchenko.

The authors used the newly developed “analogue” memristors to model various learning mechanisms (“plasticity”) of biological synapses. In particular, this involved functions such as long-term potentiation (LTP) or long-term depression (LTD) of a connection between two neurons. It is generally accepted that these functions are the underlying mechanisms of  memory in the brain.

The authors also succeeded in demonstrating a more complex mechanism – spike-timing-dependent plasticity, i.e. the dependence of the value of the connection between neurons on the relative time taken for them to be “triggered”. It had previously been shown that this mechanism is responsible for associative learning – the ability of the brain to find connections between different events.

To demonstrate this function in their memristor devices, the authors purposefully used an electric signal which reproduced, as far as possible, the signals in living neurons, and they obtained a dependency very similar to those observed in living synapses (see fig. 3).

Fig.3. The change in conductivity of memristors depending on the temporal separation between "spikes"(rigth) and thr change in potential of the neuron connections in biological neural networks. Source: MIPT press office

Fig.3. The change in conductivity of memristors depending on the temporal separation between “spikes”(rigth) and thr change in potential of the neuron connections in biological neural networks. Source: MIPT press office

These results allowed the authors to confirm that the elements that they had developed could be considered a prototype of the “electronic synapse”, which could be used as a basis for the hardware implementation of artificial neural networks.

“We have created a baseline matrix of nanoscale memristors demonstrating the properties of biological synapses. Thanks to this research, we are now one step closer to building an artificial neural network. It may only be the very simplest of networks, but it is nevertheless a hardware prototype,” said the head of MIPT’s Laboratory of Functional Materials and Devices for Nanoelectronics, Andrey Zenkevich.

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

Crossbar Nanoscale HfO2-Based Electronic Synapses by Yury Matveyev, Roman Kirtaev, Alena Fetisova, Sergey Zakharchenko, Dmitry Negrov and Andrey Zenkevich. Nanoscale Research Letters201611:147 DOI: 10.1186/s11671-016-1360-6

Published: 15 March 2016

This is an open access paper.

Robo Brain; a new robot learning project

Having covered the RoboEarth project (a European Union funded ‘internet for robots’ first mentioned here in a Feb. 14, 2011 posting [scroll down about 1/4 of the way] and again in a March 12 2013 posting about the project’s cloud engine, Rapyuta and. most recently in a Jan. 14, 2014 posting), an Aug. 25, 2014 Cornell University news release by Bill Steele (also on EurekAlert with some editorial changes) about the US Robo Brain project immediately caught my attention,

Robo Brain – a large-scale computational system that learns from publicly available Internet resources – is currently downloading and processing about 1 billion images, 120,000 YouTube videos, and 100 million how-to documents and appliance manuals. The information is being translated and stored in a robot-friendly format that robots will be able to draw on when they need it.

The news release spells out why and how researchers have created Robo Brain,

To serve as helpers in our homes, offices and factories, robots will need to understand how the world works and how the humans around them behave. Robotics researchers have been teaching them these things one at a time: How to find your keys, pour a drink, put away dishes, and when not to interrupt two people having a conversation.

This will all come in one package with Robo Brain, a giant repository of knowledge collected from the Internet and stored in a robot-friendly format that robots will be able to draw on when they need it. [emphasis mine]

“Our laptops and cell phones have access to all the information we want. If a robot encounters a situation it hasn’t seen before it can query Robo Brain in the cloud,” explained Ashutosh Saxena, assistant professor of computer science.

Saxena and colleagues at Cornell, Stanford and Brown universities and the University of California, Berkeley, started in July to download about one billion images, 120,000 YouTube videos and 100 million how-to documents and appliance manuals, along with all the training they have already given the various robots in their own laboratories. Robo Brain will process images to pick out the objects in them, and by connecting images and video with text, it will learn to recognize objects and how they are used, along with human language and behavior.

Saxena described the project at the 2014 Robotics: Science and Systems Conference, July 12-16 [2014] in Berkeley.

If a robot sees a coffee mug, it can learn from Robo Brain not only that it’s a coffee mug, but also that liquids can be poured into or out of it, that it can be grasped by the handle, and that it must be carried upright when it is full, as opposed to when it is being carried from the dishwasher to the cupboard.

The system employs what computer scientists call “structured deep learning,” where information is stored in many levels of abstraction. An easy chair is a member of the class of chairs, and going up another level, chairs are furniture. Sitting is something you can do on a chair, but a human can also sit on a stool, a bench or the lawn.

A robot’s computer brain stores what it has learned in a form mathematicians call a Markov model, which can be represented graphically as a set of points connected by lines (formally called nodes and edges). The nodes could represent objects, actions or parts of an image, and each one is assigned a probability – how much you can vary it and still be correct. In searching for knowledge, a robot’s brain makes its own chain and looks for one in the knowledge base that matches within those probability limits.

“The Robo Brain will look like a gigantic, branching graph with abilities for multidimensional queries,” said Aditya Jami, a visiting researcher at Cornell who designed the large-scale database for the brain. It might look something like a chart of relationships between Facebook friends but more on the scale of the Milky Way.

Like a human learner, Robo Brain will have teachers, thanks to crowdsourcing. The Robo Brain website will display things the brain has learned, and visitors will be able to make additions and corrections.

The “robot-friendly format” for information in the European project (RoboEarth) meant machine language but if I understand what’s written in the news release correctly, this project incorporates a mix of machine language and natural (human) language.

This is one of the times the funding sources (US National Science Foundation, two of the armed forces, businesses and a couple of not-for-profit agencies) seem particularly interesting (from the news release),

The project is supported by the National Science Foundation, the Office of Naval Research, the Army Research Office, Google, Microsoft, Qualcomm, the Alfred P. Sloan Foundation and the National Robotics Initiative, whose goal is to advance robotics to help make the United States more competitive in the world economy.

For the curious, here’s a link to the Robo Brain and RoboEarth websites.