A neural network is a new type of artificial intelligence that uses a different type of computer hardware or specifically designed software for use on conventional computers. It can recognize a pattern or predict an outcome on les than complete or accurate data. Neural networks may be the key to voice recognition by computers.

Artificial intelligence (AI) is a phrase denoting a computer or computer progra that exhibits one or more characteristics humans consider to be intelligent. Fo example, the ability to learn from experience is considered one characteristic of intelligence. One type of AI many accountants are familiar with is expert systems. Expert systems typically use rules obtained from experts to analyze problems in their field of expertise.

Although expert systems are widely accepted and used, they have several limitations. An expert system can only analyze situations in a very specific area. For example, an expert system designed to assist with audit planning cannot be used to evaluate internal controls. This limitation also often extend to nonroutine situations. For example, an expert system used to analyze the financial health of an unconventional firm within that industry. Another limitation is the time and cost of expert system development. The rules used in many expert systems are extracted by "picking the brains" of experts. This technique usually involves having experts analyze cases and determining what rules they used. This process may take months to complete.

Another problem is presented if experts disagree. Finally, if rules for analyzing problems change or if the structure of the problem changes (for example, new tax regulations) rules used by the expert system also must be changed.

Neural networks (NNs) are a type of AI that can overcome many limitations found in expert systems. NNs are based on the structure of the human brain and are composed of a large number of interconnected processors. Although NNs typically are designed for the analysis of a specific type of situation (e.g., bankruptcy prediction), they are able to extend their analyses to nonroutine situations. For example, a NN designed to predict the bankruptcy of financial institutions should be able to perform adequately even for unconventional financial institutions.

Neural Networks in Accounting

NNs can be used in a variety of situations; however, they are most useful when complex pattern recognition is necessary. Complex pattern recognition tasks performed by NNs include: 1) forecasting earnings, 2) analyzing the going concern assumption, 3) detecting fraud, 4) recognizing hand-written characters (useful for pen-based computer systems), and 5) recognizing natural language (i.e., speech). KPMG Peat Marwick has already developed a NN for bankruptcy prediction. NNs also can be used to find optimal solutions for complex problems

For the CPA, the ability to forecast a company's earnings may be useful in planning an audit or assisting management in developing an operating strategy. If actual earnings are significantly different from forecasted earnings, an auditor may decide to expand substantive testing to provide a reason for the discrepancy. Similarly, the ability to analyze the going concern situation of a entity may be very useful to a CPA. While SAS 59, The Auditor's Consideration o an Entity's Ability to Continue as a Going Concern, disclaims auditor responsibility for predicting financial distress, auditors are continually bein named as codefendant in lawsuits for not detecting and reporting the financial distress of a client firm. The use of a NN may provide an additional tool to evaluate whether an entity has a going concern problem. Furthermore, the use of a NN to analyze the going concern situation constitutes documentation of the auditor's consideration of SAS 59.

Neural Networks in Auditing

NNs also may be useful in detecting fraud by learning patterns that typically accompany fraud. Real and hypothetical examples of fraud may be used to train NNs. Any technique that reduces the probability of major, undetected fraud should assist boards of directors in exercising their governance responsibilities and decrease auditor exposure to litigation.

NNs have a major advantage over traditional techniques (statistical analysis or auditor judgment) in going-concern analysis and fraud detection. Statistical techniques and accountants are limited in their ability to discern complex patterns with a large number of data items (e.g., current ratio, percentage of accounts receivable write-offs, etc.). Consequently, a small number of data items typically is used when examining the going-concern assumption or searchin for fraud. The result is that a large amount of relevant data may be excluded from the decision. NNs do not have this limitation. They can effectively proces a large number of data items. Thus, no relevant data need be excluded from either a going concern or fraud analysis. The increased effectiveness of NNs in these situations compared to traditional techniques may help protect investors and creditors and save CPAs millions of dollars per year in litigation costs.

In addition to their usefulness in predicting financial distress and uncovering fraud, NNs can increase productivity by making data entry easier. For example, NNs can be used to recognize hand-writing and speech. This ability would be ver useful for those who lack typing skills. Commercial NN pen-based systems and natural language interfaces are currently available. CPAs, however, may choose to develop a NN to serve as either a hand-written character or speech recognition device and integrate the NN with their existing software (for example, word processor, spreadsheet, etc.).

Neural Networks in Consulting

CPAs make use of NNs in their management advisory services (MAS) practice or in solving problems in their organization. NNs can be used to find optimal solutions to very complex problems. For example, a NN can be used to determine the optimal resource allocation and production schedule for a manufacturer who produces 100 different products where each product requires a unique production process. Several firms are currently using NNs to their advantage. For example, a Kodak plant in Texas used a NN to reduce costs by three million dollars per year while maintaining production yield and quality. Several direct-mail firms use NNs to reduce total mailing by 5-15% while maintaining the same number of responses. Nikko Securities and Frontier Financial have increased profits by using NNs to improve trading strategies. While there are many potential uses fo NNs in accounting practice, many CPAs are unaware of how they operate and how they can be obtained. Presented below are the basics of how NNs function, and a sidebar presents a table of some NN products and vendors.

How They Work

NNs are computers modeled after the human brain. The brain is composed of a vas network of interconnected brain cells called neurons. Similarly, NNs are constructed of hundreds or thousands of processing elements typically arranged in layers. Figure 1 presents a diagram of a single processing element, while Figure 2 provides an illustration of how processing elements are interconnected to form a NN. Processing elements are usually arranged in layers and connected to other processing elements contained in the layers that precede or follow it (sometimes the processing elements within a layer are connected to each other a well). The normal pattern is to have three layers as indicated in Figure 2. Thi includes input and output layers as well as a hidden processing layer. More layers may be added as needed.

NNs do not derive their knowledge from experts, but from data. The training of NN for a specific application is a repetitive process. Generally speaking, ther are two approaches used to train a NN-- unsupervised learning and supervised learning. Because the unsupervised learning approach has limited applications i accounting, the following discussion will focus on the supervised learning approach to developing a bankruptcy prediction NN.

The first step in developing a bankruptcy prediction NN entails entering data (e.g., financial ratios, etc.) about known bankrupt and known nonbankrupt firms In addition, the actual state of the firms (bankrupt or nonbankrupt) is also entered into the NN. The NN processes the data and identifies a) patterns contained in the data common to bankrupt firms, and b) patterns contained in th data common to nonbankrupt firms. This process of identifying patterns containe in data is how the NN is trained. The data used to train the NN should consist of all data relevant to the decision the network will be making. The training data also should contain as many relevant examples of bankrupt and nonbankrupt firms as possible so the network will have no difficulty in recognizing relevan data patterns once training is complete. No data items useful in discriminating between bankrupt and nonbankrupt firms should be omitted.

Once training is complete, the network can be used to solve problems for which it was created. When the NN is given data about a new firm (for example an audi client), it compares the client firm's data pattern to the data patterns with which it was trained (i.e., the bankrupt and nonbankrupt firms). If the pattern contained in the client firm's data is similar to that of a bankrupt firm, the client firm is classified as a potential candidate for bankruptcy. If the pattern contained in the client firm's data is similar to a nonbankrupt firm's data pattern, the client firm is classified as a nonbankrupt firm.

The basic components of a NN include--

* Processing Elements. The electronic counterpart of a neuron (brain cell) in the human brain, referred to as a neurode.

* Connections. The electronic counterpart of the connections between neurons in a human brain. Connections provide signal transmission pathways between processing elements.

* Weights. These determine the strength of the relationship (i.e., connection between two processing elements and may either foster or inhibit a processing element's output. A positive weight encourages a processing element to output a signal, while a negative weight discourages a processing element's output. A processing element has a weight associated with every processing element from which it receives input. In most systems the assignment of weights is done entirely by computer as it processes data. Modification of the weights allows the NN to "learn."

* Transfer Function. A mathematical equation that determines the output of a processing element in response to the input data and the weights.

* Learning Law. A mathematical equation that modifies the weights based on input values and (sometimes) feedback values (defined below). The learning law enables the NN to produce more reliable outputs as additional data is entered into the network.

* Feedback. Data regarding the correct outcome for a particular item entered into the NN. Feedback is used only during the training of a NN. Providing the actual (correct) classification of a firm as either bankrupt or nonbankrupt after entering financial ratios for the firm would constitute feedback for a bankruptcy prediction NN.

Types of Neural Networks

Several types of NNs have been developed, each for a specific type of problem. Thus, the kinds of problems a particular network can solve depends on the type of NN used. While many types of NNs exist, those useful to CPAs can be classified into the following categories:

Prediction Networks. These NNs are used to predict the value of an item given values of other items. For example, this type of NN may be given current firm earnings, firm growth data, industry growth data, and general economic growth data to predict future firm earnings. This type of network also may be used to forecast capital markets or individual stock prices. A large number of prediction networks exist and most are based on a NN learning method called backpropagation.

Classification Networks. These NNs are used to classify an item as a member of specific group. For example, a network of this type may be given financial ratios to classify a firm as bankrupt or nonbankrupt. Another possible use is discovering the presence of management fraud. Examples of classification networks include categorical learning networks, counterpropagation networks, learning vector quantization networks, probabilistic networks, and self-organizing-map into categorization networks.

Data Filtering Networks. These NNs are used when input data tends to be noisy (imperfect). Because spoken language is a noisy signal, data filtering networks can be used as natural language interfaces between computers and end-users. A recirculation network is an example of a data filtering network.

Optimization Networks. These networks are used to find optimal (or near optimal solutions to very complex problems. The discussion of a NN used to determine th best resource allocation and production schedule for a client firm is one application of an optimization network.

Neural Networks Compared to Conventional Computing

NNs differ from conventional computers in several ways. First, NNs have a different architecture than conventional computers. Conventional computers are built around one (or more) complex processor(s) TABULAR DATA OMITTED that execute sequences of instructions, one instruction at a time. Conventional computers also process input data one item at a time. NNs, on the other hand, are not sequential and do not execute instructions, but react to input data. NN are composed of a large number of simple processing elements that respond to inputs in parallel (all input data is processed simultaneously). Unlike conventional computers, NNs are capable of operating even if some (up to 15%) o their processing elements are damaged.

Although the architectures and operations of NNs and conventional computers differ, NNs are capable of performing tasks normally performed by conventional computers (e.g., arithmetic). Conventional computers can also simulate NNs when used with special NN software. However, because conventional computers process inputs sequentially, NN software designed for use on a conventional computer will process data much slower than an actual NN computer. NN software is generally less expensive than hardware-based NNs.

NNs are much more tolerant of noise (imperfections) in the input data than conventional computers. While conventional computers require precise input data to produce accurate output, NNs often operate effectively with input noise levels as high as 30 or 40%. Therefore, NNs can provide useful output from less than ideal input.

NNs are much more efficient at solving pattern recognition problems (e.g., categorizing firms as bankrupt or nonbankrupt) than conventional computers. In this respect, a NN approximates human problem solving capabilities. However, conventional computers are more efficient than NNs in solving problems involvin arithmetic and logic.

Programming a neural computer differs from programming a conventional computer. A NN is "programmed" by providing the network with data relevant to the situation it will analyze. Step-by-step instructions are not needed for the NN to determine what the correct output should be--it learns what the correct output should be from the data.

Neural Networks Compared to Expert Systems

Expert systems are based on the concept expertise can be represented through formal rule-based (if-then rules) and frame-based structures. Although human novices apply rules to facts to reach decisions--most experts do not. Experts typically compare a problem to past experiences and, through a process of pattern-matching, analyze the situation and reach a decision. Unfortunately, traditional expert systems are not very good at pattern recognition and often exhibit less than expert performance. NNs on the other hand, are quite good at pattern recognition. They are also very good at classification of items and may be superior to expert systems in problems involving pattern recognition, continuous speech recognition, forecasting and modelling, diagnosis, and so forth.

NNs also overcome many of the limitations of expert systems. Limitations of expert systems include: 1) the necessity to extract knowledge from experts, 2) the inability to learn, and 3) unpredictable behavior outside the expert system's area of expertise. NNs do not require knowledge acquisition from outside sources: they obtain their knowledge from the data itself and can adapt to changes in the data (learn). The NN learning process, however, does not eliminate the need for experts. Experts are still needed to indicate which data is relevant for a particular problem.

Expert systems do have one clear advantage compared to NNs: the ability to explain their reasoning. Expert systems can explain the logic behind a particular decision, why particular questions were asked, and/or why an alternative was eliminated. NNs lack this capacity. Integration of expert systems and NNs may be a solution to this problem.

Harlan L. Etheridge, PhD, CPA, is an Assistant Professor of Accountancy and Taxation at the University of Houston. Richard C. Brooks, PhD, is an Assistant Professor of Accounting at West Virginia University.