Neural Network

New Approaches to Risk and Decision Analysis

Wednesday, March 17, 2010 by DMUU Training Team

Risk analysis and decision-making tools are relevant to most organisations, in most industries around the world. This is demonstrated by the speaker line-up at this year's European User Conference, an event at which we believe it is important to bring together customers from a wide range of market sectors.

We are holding 'New Approaches to Risk and Decision Analysis' at the Institute of Directors in central London on 14th and 15th April 2010. As with previous years, the programme aims to provide everyone attending with practical advice to enhance the decision-making capabilities of their organisation. Customer presentations, which offer insight into a wide variety of business applications of risk and decision analysis, include:

CapGemini: Faldo's folly or Monty's Carlo – The Ryder Cup and Monte Carlo simulation
DTU Transport: New approaches to transport project assessment; reference scenario forecasting and quantitative risk analysis
Georg-August University Research: Benefits from weather derivatives in agriculture: a portfolio optimisation using RISKOptimizer
Graz University of Technology: Calculation of construction costs for building projects – application of the Monte Carlo method
Halcrow: Risk-based water distribution rehabilitation planning – impact modelling and estimation
Pricewaterhouse Coopers: PricewaterhouseCoopers and Palisade: an overview
Noven: Use of Monte Carlo simulations for risk management in pharmaceuticals
SLR Consulting: Risk sharing in waste management projects - @RISK and sensitivity analysis
Statoil: Put more science into cost risk analysis
Unilever: Succeeding in DecisionTools Suite 5 rollout – Unilever's story

We will also look at the recently-launched language versions of @RISK and DecisionTools Suite, which are now available in French, German, Spanish, Portuguese and Japanese. Software training sessions will provide delegates with practical knowledge to ensure they can optimise their use of the tools and implement business best practise and methodologies.

With over 100 delegates from around the world attending, the event is also a good opportunity to network and knowledge-share with risk professionals from around the world.

» Complete programme schedule, more information on each presentation,
and registration details

Palisade is proud to announce our first Health Risk Analysis Forum in San Diego on March 31st 2010

Wednesday, March 10, 2010 by DMUU Training Team

Why attend?

This one-day forum is a great way to find out how others in the Healthcare Industry are using our software, as well as to learn new approaches to the problems Healthcare professionals face every day. We will have six software training sessions, and six real-world case studies presented by industry experts covering risk and decision analysis from all angles specific to the Healthcare sector.

You will also see how new versions of @RISK, PrecisionTree, RISKOptimizer, TopRank, NeuralTools, StatTools, and other Palisade software tools work together to give you the most complete picture possible in your situation.

Who should attend?

Professionals in risk and financial analysis in: Care Equipment & Services, Pharmaceuticals, Biotechnology & Life Sciences, Hospital Care & Management, or related services

How much?

For a limited time, the cost for attending the Health Risk Analysis Forum is has been discounted $100.

$295 covers all sessions, continental breakfast, lunch and a cocktail networking reception. Attendees will also receive a welcome package that includes a 15% discount on their next software purchase.

Please contact Jameson Romeo-Hall at jromeo-hall@palisade.com if you are interested in attending.

Location
The Westin Gaslamp Quarter
910 Broadway Circle
San Diego, CA 92101
(619) 239-2200

Book your room at a discounted rate (subject to availability.)

The Rise of the NOMFET

Friday, February 5, 2010 by Holly Bailey

By now we've become accustomed to the marvels of neural network technology and, in fact, inured to the advances it brought in statistical analysis with its computational simulations of nerve cells. Its many everyday applications--especially in online retailing--seem kind of ho-hum, and we'd be put out if for some reason they weren't in use. Wasn't it only four or five short years ago that neural nets themselves were big news?

Last week there was more big news about neural networks: a French research team's announcement of an "organic" transistor that mimics a brain's synapse. Neural network computing is based on computational stand-ins for biological neurons, and linking these neurons with electronic synapses currently requires at least seven transistors. One new "organic" transistor can take the place of those.

The key here is nano. Tiny. Tinier than tiny. The new transistors are made of nanoparticles of gold and pentacene on a plastic substrate. The resulting connector is called a nanoparticle organic memory field-effect transistor: a NOMFET.

Not only will the NOMFETs accelerate the performance of neural network circuits, but because the human brain uses 10 to the fourth times as many synapses as neurons, the space saving NOMFETs will help make possible a generation of computers inspired by the human brain.

The rise of the NOMFET may also make possible another kind of advance, one that I find a little scary to contemplate. Because its built on plastic, the NOMFET could potentially be used to link a computer with living tissue. Get back, Frankenstein.

A Downturn for the Better

Thursday, January 7, 2010 by Holly Bailey

Honoring a time-honored tradition for the turn of the year, I've been looking back over the year just past to do a little retrospective trend-spotting. Here's one that took me by surprise: in spite of the downturn in the economy, there was also a downturn in online fraud. It's counterintuitive--historically, hard times are correlated with rising crime--but apparently true.

Late last year, DigitalTransactions, an online publication catering to businesses engaged in the "electronic exchange of value," reported that the results of a survey of principals in these businesses showed an overall decline in fraud of about 1 percent.

The survey, sponsored and carried out annually by a California risk management company, is the first in its eleven-year history to show a fraud rate this low. In 2009 North American merchants were expected to lose (a mere) $3.3 billion, in contrast to their loss of $4.0 billion in 2008.

What's behind this good-news downturn? Probably not increased honesty. There was no data on attempted fraud, and the assumption is that the increased use of automated fraud detection tools cut the merchant's losses. The level of sophistication of these tools has ratcheted up to the level where neural network classification, risk analysis, and statistical analysis of correlated data can take place in real time during the processing of a transaction. Furthermore, the combination of operational risk software with device identification of the purchaser's computer now make it difficult for a single computer to mob an online merchant with multiple bogus orders.

So the good news is not about improvements in human nature. It's about improving the defenses of this booming sector of the economy.

Digital Eyes on Alien Life

Wednesday, December 9, 2009 by Holly Bailey

University of Chicago geoscientist Patrick McGuire has big plans for Mars. Previously he worked on an imager for a Mars orbiter that could identify different types of soil and rock by detecting infrared and other wavelengths, and now he is drawing on that experience to develop a space suit with digital "eyes" and a neural network that rides on the hips of the spacesuit and can sort out living biological material from other matter.

The digital eyes will detect and plot colors, and the neural net, which is known as a Hopfield neural network, will compare these color patterns to a database of information previously gathered from that area of planet in order to make an animal-vegetable-mineral determination.

This complex AI system has already been tested at the Mars Desert Research Station in Utah, and McGuire and his colleagues were satisfied that the Hopfield algorithm could learn colors from just a few images and could recognize units that had been observed earlier.

McGuire's concept is that a human wearing this neural network could simply walk around the red planet and record every nearby object, rapidly gathering information.

Obviously, such a clothing item awaits a manned Mars mission. But in the meantime, why not have the next Rover suit up?

The Cat is Out of the Bag

Thursday, December 3, 2009 by Holly Bailey

At November's supercomputing conference in Portland, Oregon, IBM announced that its researchers working with a team from Stanford University had succeeded in developing an accurate simulation of human brain function. The simulation will be capable of emulating sensation, perception, action, interaction and cognition.

This algorithm simulating a living neural network, called BlueMatter (spelled as one word like everything else in computerese these days) is an important milestone in IBM's mission to build a cognitive computing chip because it begins to advance large-scale simulation of a cortical neural network and it synthesizes neurological data. BlueMatter is built with Blue Gene (two words for this pun in the singular) architecture, which, in combination with specialized MRI images, allowed the team to create a wiring diagram of the human brain. This map of the brain is, according to IBM's press release "crucial to untangling its vast communication network and understanding how it represents and processes information."

To be more accurate, what BlueMatter has thus far demonstrated is the potential to achieve neural network technology that operates on the scale of complexity of the human brain. The algorithm's current simulation approximates the cortical system of a cat. Hence, the title of the paper announcing IBM's accomplishments: "The Cat Is Out of the Bag." Even so, this is an operations research accomplishment that dwarfs such mundane analytical tasks as option valuation, value-at-risk, or reserve estimation.

One of the goals of the company's cognitive computing program is to create a chip that operates with the energy efficiency of the human brain (20 watts). But in order to emulate the brain activity of a cat, the research team had to bring out one of the largest supercomputers in the world, the IBM Dawn Blue Gene/P--which comprises about 150 thousand processors and contains 144 terrabytes of main memory.

This cat came out of a pretty big bag.

New Approaches to Risk & Decision Analysis at the 2010 Conference in London

Friday, November 13, 2009 by DMUU Training Team

Following on from the resounding success of the last Palisade Risk Conference in London, which attracted over 110 attendees from industry and academia, the 2010 Palisade Risk Conference will be taking place on April 14th-15th. The location for this event will again be the Institute of Directors on Pall Mall, London, and already there are a number of exciting presentations confirmed from the likes of Unilever, Pricewaterhouse Coopers and Halcrow.

The 2010 Palisade Risk Conference will be a two-day forum which will cover a wide variety of innovative approaches to risk and decision analysis. Featuring real-world case studies from industry experts, best practices in risk and decision analysis, risk analysis software training, and sneak previews of new software in the pipeline, the event is also an excellent opportunity to network with other professionals and find out how they’re using Palisade risk analysis solutions to make better decisions.

Call for Papers

If you have an unusual or interesting application of Palisade software which you would like to present, please send a short abstract to cferri@palisade.com. The closing date for abstracts to be submitted is Friday, 11th December, 2009.

A Random Walk

Friday, October 23, 2009 by Holly Bailey

It has struck me often that the target of statistical analysis writ large is randomness. Because it is the fundamental concern of all the techniques for decision making under uncertainty that I try to track--risk analysis, genetic algorithm optimization, and neural network prediction--I assumed I understood the term randomness pretty well.

But it turns out that what you think randomness is depends upon what line of work you're in. If you are a statistician, randomness occurs in a repeating process the results of which follow no discernible pattern. If you are a geneticist, randomness applies to genetic mutations that are not controlled by genes. If you are a financier, randomness is blamed for controlling stock prices, which respond instantaneously changes changes in available information.

In all these fields, randomness is closely allied with chance and probability, and the human struggle against chance is epic. We resort to sharper and sharper tools to pare down what appears to be random, and just when we think we have the magic tool--say, Monte Carlo software--or the magic idea--the random walk hypothesis, for example--the definition of randomness changes.

Neural Network Zeros in on Quarks

Monday, October 19, 2009 by Holly Bailey

Having successfully dodged high school physics I would not normally be sucked in by an article on quarks, but this one involved neural network computing, which I do understand pretty well.

It seems that a couple of weeks ago physicists at Fermilab, near Chicago, made the most precise measurement yet of a top quark. A quark is an elementary particle, the most fundamental building block of matter. Quarks come in six flavors (I'm not making this up!), four of which can be produced only by high-energy collisions. Think updated cyclotron. The top quark is one of these four, and first observed in 1995, it is the most recently discovered quark. The physicists--unsatisfied, of course, with having simply identified the particle--wanted to measure it.

It turns out that the way to measure a quark is to observe its decay and work backward from non-quark to quark. This involves heavy-duty statistical analysis of many, many observations. The scientists at Fermilab collected a large set of sample data on quark decay, and then in order to zero in on bona fide quarks, they trained a neural network to identify which particle events were not related to top quark decay. When the neural net had sorted out the quark imitators, the physicists could size up the real quarks more accurately.

The top quark is relatively large for an elementary particle. Until last month it was believed to be about the size of an atom of gold. What is the current estimate? Too daunting a calculation to quote. But if you go to the information the Fermilab has on display, you--or some of you, anyway--will begin to get the picture.

Introduction to DecisionTools Suite 5.5 Products: Software training at the NYC Conference

Monday, October 12, 2009 by DMUU Training Team

Palisade is gathering trainers from our New York and London offices to present software training seminars next week at the 2009 Palisade Conference: Risk Analysis, Applications, & Training. The conference is set to take place on 21 - 22 October at the Hyatt Regency in Jersey City, 10 minutes by PATH from Manhattan's Financial District.

Take this convenient and inexpensive opportunity to learn from Palisade’s trainers and software developers. Learn how to use the elements of the new DecisionTools Suite 5.5 as a comprehensive risk analysis, optimization, and statistical analysis toolkit. See how each of the products in the Suite — @RISK, RISKOptimizer, Evolver, PrecisionTree, TopRank, StatTools, and NeuralTools — can be used to solve practical problems in the real-world.

The conference also features case studies demonstrating how to use @RISK and DecisionTools Suite, from risk management experts in the fields of finance, healthcare and pharmaceuticals, energy, oil and gas, DFSS and Six Sigma, project management, operations management, manufacturing, and more.

See the full schedule for the Conference here.

Next Week: October 21-22 in NYC

Building on the success of last year’s record-breaking event, the conference will offer a wide range of software training, model building, and real-world case study sessions. Last year, the event drew over 150 practitioners and decision-makers from a broad spectrum of industries. The @RISK and DecisionTools software tracks were more popular than ever. This year, we’re expanding software training with sessions that let you walk through examples and try the tools directly. This will enable you to take some new tips back to the office. Please join us in October for a great opportunity to learn and connect with colleagues.

A Neural Network Is Never Too Old to Be Immature

Wednesday, October 7, 2009 by Holly Bailey

Here's a follow-on to my column about evolving neural networks. That one involved deception, this one involves distractiion. That one involved robots, this one involves babies.

About fifty years ago, logician and child psychologist Jean Piaget designed what has become a classic experiment to test the memory and learning of babies. It is a game of hiding and finding, and through it we have discovered that when infants up to 10 months old are repeatedly shown an toy being hidden in a certain place, they continue to look for the toy there, even when they have also seen it hidden in some other place. By the age of one year, however, they get it. They figure out they can look in more than one place.

This experiment has been used for decades by scientists interested in human development, and most recently has led to a finding by a Hungarian team that the ability of young infants to read social cues actually misleads them and causes them to perform worse in the hide-and-seek game. Older babies, however, were still able to read through the deception.

Now scientists at the University of Iowa have used a neural network model to prove that the problem is not the infants' mistaken reading of a social cue but simply distraction claiming the attention of the young infants and thereby disrupting their memory of the actual hiding of the toy.

How, you are wondering, did the UI team verify an internal cognitive process with neural network? Their neural net trained on the responses of the infants in many different versions of the hiding game, and the team then programmed an interruption in the flow of computation so that the computer stopped "paying attention" to the hiding event. Then it too flunked the memory test.

What is the take-home message from this? That you can fool all of the babies only part of their lives? Or that an neural network is never too old to be immature?

The Evolution of Deception

Friday, October 2, 2009 by Holly Bailey

No sooner did I mention an apparent trend toward using neural networks to analyze biological processes than another fascinating study came up on my radar: in Switzerland two engineers and a biologist are using a neural network to simulate the biological evolution of social communication.

The Swiss research employs camera-equipped robots instead of animals, and it sends a group of robots out "foraging." The robots are equipped with light sensors, and the "food" at one end of the foraging area is a lighter color than the "poison" at the other end. The bots are scored for their success at locating the "food," and they "talk about" their success by flashing a blinking blue light. Which, in turn, draws robots that are computationally attracted to the blue light.

After that, the self-prepetuating algorithms of the neural network experiment begin to look a little like genetic algorithm optimization. The robots are randomly "mated," and their neural nets are mingled. In only a few generations, the robots learn to head towards the blue lights. But here there is a hitch: the "food source" is available to only a few robots at a time.

What evolves next? By the 50th generation the robots begin to do what you and I would probably do if we had the advantage of a precious resource we didn't want to share: they deceive. They stop signaling with their blue lights when they find "food" because their signals will draw a crowd of greedy robots.

Interestingly, this story of evolving robots mirrors research done on live chickens in the mid-1990s, except that there was also sex and mating motivation--the real kind--at work in those experiments. Animal behaviorists studying vocal communication observed that roosters will call to notify other chickens about the availability of food only if those other chickens are hens.

In spite of this difference, the evolutionary logic of both the chickens and the robots is the same: why share the good stuff and lose your advantage?

A Neural Network by Any Other Name

Wednesday, September 30, 2009 by Holly Bailey

You say toe-MAY-toe, and I say toe-MAH-toe. But I still know exactly what you mean: a big, round, juicy red veggie that slices up nice for the burgers. But how I know the difference between toe-MAH-toe and toe-MAY-toe is a mystery that has eluded neuroscientists. Until now.

In the many press releases I see on scientific topics, I've noticed a trend toward using neural network technology to analyze its biological namesakes, the neural networks in the human brain. One of the latest examples of this is research from a team of scientists at Hebrew University who have used computational neural networks to analyze the cellular processes by which sensory neurons adjust to differences in speech for the same word.

The differences in the way I say tomato and you say tomato are largely a matter of timing and durations, and these sounds are received by single nerve cells. The neural net algorithms devised by Dr. Robert Gutig and Dr. Haim Sompolinsky identify these differences by classifying the way the single nerve cells respond. This innovation will not only be useful in such speech decoding applications as telephone voice dialing, but they also have promise in treating auditory problems.

Toe-MAY-toe? Toe-MAH-toe? Let's call the whole thing....Naw, the two brain scientists aren't calling anything off. Their neural network is just getting started.

Simulating the U.S. Economy: Where will we be in 100 years?

Friday, September 4, 2009 by DMUU Training Team

William Strauss is the President and founder of FutureMetrics. He brings more than thirty years of strategic planning, project management, data analysis, and modeling experience into the company’s stock of knowledge capital. Bill’s professional history includes executive positions as director, president, and senior vice president, as well as positions as senior analyst and field coordinator. He has an MBA (specializing in Finance) and a PhD (Economics).

Dr. Strauss will present a case study at the 2009 the 2009 Palisade Conference: Risk Analysis, Applications, & Training. The conference is set to take place on 21 - 22 October at the Hyatt Regency in Jersey City, 10 minutes by PATH from Manhattan's Financial District.

See the abstract for his case study below, and see the full schedule for the Conference here.

Simulating the U.S. Economy:
Where will we be in 100 years?

There is an assumption that drives all of our expectations for how our economy will be in the future. That assumption is one of endless economic growth. Clearly endless exponential growth is impossible. Yet that is what we base all of our expectations upon. We all agree that zero or negative economic growth is bad (just look around now at the effects of the Great Recession). But we also know logically that 2% or 4% annual growth every year leads to an exponential growth outcome that is unsustainable.

To see where this growth imperative will take us we first have to see how we go to where we are today. This work first models the 20th century. The model is both complex and simple. The basic schematic of the model’s relationships is easy to understand. Furthermore, the core of the model is a simple production function that combines capital, labor, and the useful work derived from energy to generate the output of the economy. Complexity is contained in the solutions to the internal workings of the model. What is unique is that there are no exogenous economic variables. Once the equations’ parameters are calibrated, setting the key outputs to "one" in 1900 results in their time paths very closely predicting the U.S. GDP and its key components from 1900 to 2006.

The experiment in this work is about the future. If the model can very closely replicate the last 100 years, what does it have to say about the next 100 years? From 1900 to 2006 there are periods in which there was parameter switching. (The optimal parameters and the years for the switching were found using a constrained optimization technique.) That suggests that in the future there will also be changes. The experiment uses @RISK’s features to generate new combinations of parameters for each of tens of thousands of runs of the simulation. Changes in the parameters represent potential exogenous policy choices.

The "doing what you did gets you what you got" scenario leads to a surprising and unsettling outcome. The experiments using @RISK do find a path that works. Obviously if it is not "business-as-usual" that leads to a stable outcome, it is some other way. The policy choices that lead to a stable outcome suggest that the future of capitalism is not going to be what we expect it to be.

Please join us in October in New York for software training in best practicies in quantitiative risk analysis and decision making under uncertainty, real world case studies from risk services consultants and experts, and networking with practicioners from many different fields including oil and gas, pharmaceuticals, academics, finances, Six Sigma, and more.

Palisade’s Custom Development Services

Wednesday, August 12, 2009 by DMUU Training Team

Palisade Corporation now offers custom development services. Our consulting team can help you to automate your risk and decision analysis models so they can be easily used by everyone in your company, or even outside of it.

We offer different options that include Excel add-ins, Windows, and Web based applications. Our consultants can help you to design, program and deploy these applications. A typical application might connect an Excel spreadsheet to your company’s database, extract data, then adjust it to probability distributions so they can be used in dynamic risk or optimization models. The structure of reports can be also customized and published as PDFs, or to the Web.

Palisade Custom Development can incorporate Monte Carlo simulation, probability distributions, distribution fitting, graphs, reports, and many other features of @RISK into any Windows-based application. In addition, we can integrate genetic algorithm optimization from RISKOptimizer or Evolver. This allows you to apply powerful, proven analytics to applications outside Excel. Applications can be run in a desktop, network, or Web environment.

You may wish to customize your @RISK or DecisionTools Suite spreadsheet models, restricting access to model components for some users or automating reports and other aspects of your analysis. Using the DecisionTools built-in Excel Developer Kit (XDK) and custom Excel VBA programming language, Palisade can help you build powerful, easy-to-use risk models for one user or for an entire work group.

We are currently working on a new website where you will find more information and project samples. Upcoming posts will discuss examples of custom Excel VBA programming.

» More about Palisade Custom Development

Dr. Javier Ordóñez
Director of Custom Development

Faces in the Risk Analysis Crowd

Thursday, August 6, 2009 by Holly Bailey

About ten years ago, I went to my first software user's conference. It was directed at people working in risk analysis, and because it was held in a university business school, I was expecting to find a room full of techno-geeks with superb chops. But it wasn't that daunting. Many of the users in that early crowd were people who were trying to learn about Monte Carlo software because their bosses had decided the outfit needed to be doing something called risk analysis, but there was one toxicologist who was frequently called upon to serve as an expert witness in the courtroom. The classes all had titles like "Introduction to. . ." and " Basic Components of Decision. . . ."

Today, some of the old familiars sport a few more gray hairs and there are many new faces, but overall, their expertise and their titles have ramped up. Their experiences working with colleagues on Six Sigma projects and decision analysis efforts and competing with other companies implementing Monte Carlo techniques and other decision support techniques have made them much more aware of the capabilities of statistical analysis techniques like genetic algorithm optimization and neural networks. And by now, of course, Monte Carlo software is standard issue in their business sectors.

Because of the tremendous savvy of people like you, my readers, I find I have to be very careful in my use of terminology, etc. I often flee to Wikipedia. Clearly it's time for me to attend another user's conference because I know for a fact the everybody in the crowd will have some truly daunting chops.

Neural Networks for Neural Security?

Friday, July 24, 2009 by Holly Bailey

Here is a futuristic follow-on to yesterday's neural network column about a computer security system that evaluates typing signatures. In the July 1 issue of Neurosurgical Focus, two computer scientists published their concerns about potential security hazards of "neural devices," computers that operate on a direct brain-computer interface. Neuroprotesthetics have been used for the past decade or so to help people with damaged hearings, sight or movement, and they have been targeted for possible military and commercial applications, especially gaming.

Because neural devices involve harnessing a human brain to a computer, there is some potential for hackers to actually mess with a person's mind. While one of the article's authors, Tadayoshi Kohno, believes that most of the current devices are safe--here, I suggest, is an issue ripe for risk analysis--but says it would be a mistake not to gird up in advance against hackers and viruses. Gird up how?

Neural networks, the machine kind. One writer envisions security software that comes in both standard and neural network versions. Kind of like hair of the dog that bit you, or it takes a neural network to know one.

Neural Networks and the Quest for Identity

Wednesday, July 22, 2009 by Holly Bailey

I'm a pretty fair typist. I'd say my current speed is roughly 60 words per minute--I was more accurate before this keyboard made it so easy to delete. What I am not good at and the instance in which I do more deleting per character is inputting passwords. Password requests make me tense up, and even for those passwords I remember perfectly well, I tend to let the characters ripple off my fingertips. Often my efforts to establish my identity are rejected, and this means I have to return to the dialog box, slow down, and pick at the keys like a chicken going for the grit.

So you can understand why the latest innovation based on a neural network is bad news for me. A Boston company, Delfigo, has just released computer security system that uses a neural network to authenticate whether or not the user is who he or she is attempting to pass as.

As you're probably aware, neural networks are good at recognizing and classifying patterns. The new identity authentication system attends to the electrical signals coming from the computer keyboard and attempts to match them to the known typing "signature" of anyone with legitimate access to a network or website. The input data for the neural network consists of "dwell" time and "flight" time, how long a finger remains on a key and how much time that finger spends in suspended in the air above the key. This does not bode good things for a person whose key striking is as inconsistent as mine. My identity will be hard to authenticate, and the confidence score generated by the neural network will be a little shaky.

My identity in question--but how can this be? Who I am is how I type?

Sure is, and a word to the wise: Protect your identity. Hunt and peck no more.

The Tank, The Volkswagen, and the Neural Network

Thursday, June 25, 2009 by Holly Bailey

Six years ago when Dale Addison was speaking to a group of engineers and trying to pitch "artificial intelligence"--meaning neural networks someone in his audience asked him if it was true that a neural network had once mistakenly classified a T-62 tank as a Volkswagen. Although the incident had occurred years before, Addison seems to admit that there was some truth to the story. At the time, he was dismayed to see how little confidence this technologically sophisticated audience had in neural nets because even by that time the technology had made huge progress since computer scientists started tinkering with it.

Addison is on the faculty of the University of Sunderland, and in spite of his audience's skepticism, he should be feeling fairly smug that many of the applications he foresaw for neural network technology, especially those that involve accurate classification, have been exploited. We have customers who use it on a daily basis for such tricky classification tasks as cancer diagnosis, emergency response systems, exploration and production of gas and oil, and operations research issues in manufacturing.

Nevertheless, Addison still sees resistance to emerging artifical intelligence techniques among engineers and business people, and he is still out there pitching neural nets, especially their use in combination with other new computational analysis methods, such as genetic algorithm optimization and neuro-fuzzy logic. Addison himself is working on some really tough classification problems now in the CASSANDRA project, where the goal is to develop an insider trading and market abuse detection system. He doesn't worry about the tank and the Volkswagen because he's confident that sooner rather than later, a neural network will learn to recognize a suspicious transaction when it sees one.

The Neural Network and the Click-Through

Friday, May 29, 2009 by Holly Bailey

As the market for neural network software has become more and more competitive, I've been intrigued to watch the proliferation of applications for this breed of statistical analysis. In a week that has produced news of neural networks put to use to diagnose epilepsy, pick stocks, protect children from internet pornography, and predict wind power came a particularly intriguing item that might not get the attention it deserves: Penn State information sciences professor Jim Jansen and his colleague Amanda Spink of the Queensland University used a neural net for a study of internet search engines and user satisfaction. Because click-throughs mean potential sales for businesses that rely on internet advertising, the study could send search engine developers scrambling to retool their engines.

Here was the question Jansen and Spink posed: what is it about the results a search engine produces that causes the person who receives them to click through on a particular result?

Here was how they went about answering that question: they obtained data on 7 million interactions from the search aggregator Dogpile and used neural networks software to classify the purpose of the search--e.g., information gathering, navigation help--used other statistical analysis methods to relate their classifications to the number of click-throughs. Here a click-through indicates that the user was satisfied enough with the result to pursue it further, and obviously a search engine that produces more click-through responses is more commercially desirable.

I'll be interested to see the reaction to the study from the search engine community, but in the meantime I'll certainly be more aware of click-throughs in my work on commercial websites.