Menu
Rivals question CA's prepackaged neural nets

Rivals question CA's prepackaged neural nets

Comments

Some customers and analysts warn not to expect too much from prepackaged neural net software coming from Computer Associates.

CA will announce next week databases of neural network "experience" to give a learning boost to networks running its InterBiz business intelligence framework.

Billed as Unicenter TNG for the CEO, InterBiz uses CA's Neugent neural net software to track and predict business -- rather than IT infrastructure -- events. CA plans to debut InterBiz at its annual conference in April.

The packaged experiences, called Intellikits, are created using the InterBiz accounts payable application, BizWorks. As an enterprise using the application trains its systems, the Intellikits supply data on business events to the BizWorks control centre, said Israel Levy, vice president for business development at InterBiz.

But neural network developer Richard Gordon, from consultancy Predictive Systems in New York, warned against expecting customised results from prepackaged models.

Neural nets "are high-powered algorithms and require detailed knowledge" to train and initialise, he said.

What neural nets offer that pure statistics don't is the ability to elicit results when working with incomplete data that is specific to an organisation, he said.

But Neugents encompass more than neural nets, said Brandon Musler, a Neugent business manager at CA. Neugents' "adaptive pattern recognition" also includes statistical algorithms and other artificial intelligence technologies, he said.

Custom neural network developer Nestor specialises in fraud detection in the financial industry. Typically, a Nestor spokesman said, neural net developers would review three months of transactions, taking all of those that the company had labelled fraudulent and a subset of good transactions to build a data set.

The developers would examine transaction data elements such as frequency and time to discern patterns that correlate with fraudulent transactions. Using that data, they would refine the company's rules and build a neural net model that would work in conjunction with the rule set.

At CA's annual conference in July, Charles Wang, CA's president and CEO, said: "We may write a rule that says, 'Only the president may write a cheque for $1 million,' " he said. But the rules alone wouldn't stop somebody "from writing 10 cheques, each for $100,000."

In writing those rules, an enterprise can fine-tune its implementation of Neugents, Levy said. "BizWorks uses [that] rules-based system to correlate business events for the user," he said.

Companies can use CA's application development tool Jasmine II to create more customised implementations or have CA do it for them, Levy said.

Gordon advised companies that try the technology to "see how it operates in their environment, track the results to see how well they correlate to what they know about their business and their industry".

The first kits are targeted for release mid next year and will target cheque bouncing, account delinquency and fraud detection. No pricing is yet available.


Follow Us

Join the newsletter!

Error: Please check your email address.
Show Comments