<p><b>The problem of uncertainty. </b>The third issue which has posed problems for AI programs is the factor of &quot;uncertainty&quot;. Computers work with a &quot;Yes or No&quot; logic. A characteristic belongs to a pattern, or it does not. A pattern can be selected, or rejected on this basis. Unfortunately many characteristics have vague relationships to patterns. They are only sometimes present. &quot;Fuzzy logic&quot; attempts to handle vagueness by giving grades to a characteristic, such as short, medium height, tall and very tall. While this helps to define a characteristic in greater detail, it fails to handle identification of a person who sometimes wears spectacles. A computer can match &quot;wears glasses&quot;, or &quot;does not wear glasses&quot;. It cannot handle both. Unfortunately most patterns have such variable qualities. This essay attempts to show how such uncertainty can still help pattern recognition.
<p><b>Instant identification of context.</b> The fourth issue, which has frustrated AI research is the inadequacy of available tools to gauge the awesome size of the search space. When an AI program attempts machine translation of a word in context, it must store contextual data and recall this through a search process. It is like searching for a needle on the beach. The mind instantly identifies context. Every seen object or event fetches its own contextual background. When the word &quot;pool&quot; is used with &quot;swim&quot;, it suggests one meaning and quite another when used with &quot;cartel&quot;. As we read, specific meanings, which exactly suit the context, are instantly recalled. The mind holds a lifetime of memories and associative thoughts. Yet it instantly identifies a single contextual meaning from such a gargantuan search space. Computers seek an item in memory through a serial match. One characteristic of the perceived object is compared with the characteristic of an item in memory. If this matches, the second characteristic is compared and so on, in a systematic search.
<p><b>An intractable search problem. </b>The search space is enormous. In AI, a systematic search brings related problems as to where to begin a search, and the direction of the search. &quot;Heuristics&quot; is a term used for determining a search direction. If one is searching for a needle on the beach, heuristics would suggest a search to the North to locate it. But such solutions work only in small search spaces. In spite of many attempted shortcuts, all such search algorithms eventually face the problem of a &quot;combinatorial explosion&quot;. The back and forth search paths become intractably prolonged and cumbersome. While it takes milliseconds for the mind to locate a memory in context, the AI search and match algorithm would take years, if it was to recall a single memory from a lifetime of memories. This essay suggests an algorithm which can make instant identification practical for the mind in the context of a large search space.
<p><b>A slower processing mechanism.</b> The fifth puzzle is that the human nervous system is known to process data far slower than a computer. (1) While messages in integrated circuits travel at the speed of light, nerve impulses travel just a few yards per second. While computers process information in millions of cycles per second, the mind runs at between 50 and 10,000 cycles per second. When one considers the enormous size of the memory bank of the mind, how does a slower processing system achieve such incredible speed in locating one memory from trillions of memory traces? This process of instant identification is usually called intuition, a hitherto unexplained and mysterious capability of the mind. Parallel processing by the billions of nerve cells in the nervous system does explain some of the complexity of the mind. Even then, no known search algorithm can achieve such precision with such speed. This essay suggests a search algorithm which could be used by the mind to practically achieve the speed of intuition, even within the limitations of the slower processing speeds of the mind.
<p><b>No chain of reasons. </b>The sixth issue is the mystery surrounding the reasoning processes of the mind. AI programs attempt to give &quot;backward chaining&quot;. When a solution is offered for a problem, step by step reasoning is provided for the final conclusions. A chain of reasons links the premise to the conclusion. Yet, the average person detects a mistake in the syntax of a sentence, without necessarily knowing anything about nouns, verbs, prepositions, deep structure, or other intricacies of grammar. When a person pays attention to a sentence, errors are detected, without always knowing why they are errors. Thus the reasoning processes used by AI do not appear to be the methods used by the mind. This essay suggests that the mind may be constructed around a pattern recognition model, which does not apply reasoning chains to draw its conclusions.
<p><b>Where does memory reside ?</b> The seventh issue that has baffled scientific research is the scarcity of data concerning the location of human memory. (2) Classic experiments carried out in the early part of this century on the memories of rats concluded that no particular location of the brain stored memories and that memories were somehow stored in a distributed fashion across the entire network. Current theory supports this hypothesis that memory is a network phenomenon. Research from the seventies in &quot;neural networks&quot; suggested that a network could be induced to carry a memory through their tendency to balance the relationships between various nodes. By providing &quot;weightage&quot; to nodes, it was possible for units of memory to be stored. Such an explanation implied that the nodes were devices which received inputs, carried out certain computation and sent out nerve signals. Opposing this theory, this essay suggests a recognition rather than a computational role for nerve cells. In the process, the paper suggests a location for human memory.
<A NAME="2"><p align="center"><H1>A New Algorithm</H1>
<p><b>Recognition and intelligence. </b>Consider the process of reading. The words are just black and white patterns on paper. Recognition of the patterns conveys the purpose of the author to the reader. A single message on paper can move an army. The act of recognition of the patterns on the paper provides a powerful, but invisible link. If we did not comprehend the recognition process, the arrival of a march order would appear to have a puzzling response. The nervous system appears a mysterious network, with billions of inter-linked communicating nodes. The process of becoming conscious, or of paying attention appear as baffling activities of the system, without any rational explanation. This essay shows how instant recognition of patterns by neural processes can reasonably trigger intelligent activity in real time. Recognition appears to be the key to intelligence.
<p><b>The Intuitive Algorithm (IA). </b>While the geography and functions of the human nervous system are well known and well documented, the mind remains a mysterious entity. The key insight to the answers suggested in this essay come from a diagnostic expert system which uses a new pattern recognition algorithm. It logically achieves virtually instant recognition in a large search space - the suspected quality of intuition. A similar logic can enable intuition to achieve the equivalent of instantly finding a needle on the beach. It removes the mystery surrounding intuition. It can be viewed as a practical process which can identify a single item from an astronomically large database. It grants the mind the ability of timely recognition in context. The insight opens to view the awesome range and power of an intelligently interactive mind. The concept begins with the expert system. It uses a singular algorithm. Let us call it the Intuitive Algorithm (IA).
<p><b>The conventional expert system.</b> When presented with a list of indicated symptoms, a diagnostic expert system identifies a disease. Its database contains hundreds of diseases and their symptoms, including many commonly shared symptoms. If a disease is a pattern, the objective is to identify a single pattern in a collection of interweaving patterns. As explained before, traditional expert systems achieve this with an open ended search, based on indicated symptoms. The database is searched for a disease that exhibits the first symptom. The first located disease having the first symptom is tested for the second symptom. If the test fails, a new disease with the first symptom is located and the second symptom is again tested. Each new symptom brings new diseases into evaluation. The search ends when all the presented symptoms match the indicators of a single disease.
<p><b>The IA process. </b>IA uses a different approach in a logical search of a database. Each disease is stored with one of three (&quot;<u>Y</u>es&quot; (Y), &quot;Ne<u>u</u>tral&quot; (U), or &quot;<u>N</u>o&quot; (N) ) relationships to each symptom question. Y means a positive link - the symptom is always present in the disease. U means the symptom is sometimes present. And N means the symptom is absent for the disease. After each answer to a presented symptom question, the Y/U/N relationships of all diseases are tested in a single step, just the way all cells in a spreadsheet are instantly recalculated. The Y/U/N relationships are entered specifically for their negative impact. An &quot;Yes&quot; answer eliminates all &quot;N&quot; diseases. If the problem is unilateral, all bilateral eye diseases are eliminated. A &quot;No&quot; answer eliminates all &quot;Y&quot; diseases. If visual acuity is not affected, all eye diseases which impact on visual acuity are eliminated. IA also purges questions which have &quot;Y' relationships only to eliminated diseases. The questioning process begins with the question which has the maximum number of &quot;Y&quot; relationships. It ends when the presented symptoms eliminate all but a single disease. Specific questions can then confirm the diagnosis. If all diseases are eliminated, the conclusion is that the presented symptoms do not match any disease in the database. For IA, it is then an unknown disease. Such a problem solving approach gives IA some exceptional capabilities.
<p><b>IA circumvents &quot;stupid questions&quot;. </b>Normal search algorithms serially seek to match a symptom with a single disease. IA narrows the search faster by evaluating the entire database concerning the current answer. IA is holistic. Doctors know that the lack of a particular symptom clearly indicates the absence of a particular disease. So, a subsequent query which suggests the possibility of that disease is a &quot;stupid question&quot;. If a patient reports a lack of pain, a subsequent question posing the possibility of a disease which always presents a powerful pain symptom is, naturally, considered stupid. Such a question annoys the user. With their &quot;back and forth, open ended&quot; serial searches, a traditional expert system is blind to the global impact of a previous answer on subsequent questions. Additional steps are required to correct this defect. IA avoids &quot;stupid questions&quot; by purging all &quot;Y&quot; questions which relate only to diseases eliminated by the process.
<p><b>IA logically manages &quot;uncertainty&quot;. </b>When a disease exhibits a symptom only occasionally, (a &quot;U&quot; condition), it is retained within the database regardless of whether the answer to the symptom question is &quot;Yes&quot; or &quot;No&quot;. The disease is not eliminated. It remains available for &quot;further consideration&quot;. IA continues the elimination process. Each answer eliminates &quot;Y&quot; or &quot;N&quot; diseases as per the entered relationships, taking IA ever closer to the answer. IA achieves the subtle objective of making a decision on an uncertain piece of information. While the disease with the uncertain condition is &quot;retained&quot;, every answer continues the elimination process. On the other hand, an uncertain condition is &quot;garbage&quot; for a traditional expert system, which cannot &quot;match&quot; a disease which has a &quot;maybe&quot; relationship to a symptom. Since IA does not seek an exact match, it logically handles &quot;uncertainty&quot;. For correctly entered relationships, the IA logic is flawless in diagnosis. Traditional expert systems are slowed down through the exponential growth of their back and forth search steps. They ask a tediously long series of questions, including stupid ones. They fail to handle uncertainty. IA is generations ahead of current expert systems. Doctors certify that IA is fast and never asks stupid questions.
<p><b>Inductive logic. </b>But, IA follows the logic that a person does not have a particular disease if he does not have a particular symptom. This is not a conventional logical derivation. In any diagnostic process, we can use deductive, or inductive reasoning. In deductive reasoning, a generally accepted principle is used to draw a specific conclusion. All men are mortal. Socrates is a man. Therefore Socrates is mortal. When a person uses a number of established facts to draw a general conclusion, he uses inductive reasoning. For instance, the observation of swans over the centuries has led to the conclusion that all swans are white. This is the kind of logic which is normally used in the sciences. An inductive argument, however, is never final. It is always open to the possibility of being falsified. The discovery of one black swan would falsify &quot;the white swan theory&quot;. Inductive reasoning is always subject to revision if new facts are discovered. The sciences progress through this process of induction and falsification.
<p><b>Exclusion is also a logical process.</b> Inductive reasoning has traditionally been based on the principle of inclusion. The white swan theory is a result of experience over time. If we saw a white bird, we would move one step forward in identifying it as a swan. But logic is equally sound in exclusion. If the bird was black, we could conclude that it is not a swan. Subsequent discovery of a black swan would make this induction wrong. But, if the reasoning that all swans are white was true, then the induction that a black bird is not a swan would be equally true. The white swan theory can logically lead to both conclusions. In a similar manner, if a symptom is always present for a particular disease, inductive logic also implies that an absence of the symptom excludes that disease from further consideration. This is not a conventional conclusion, but is accurate and unassailable.
<p><b>IA avoids an exact match and uses elimination. </b>A conventional search algorithm seeks an exact match between indicated symptoms and the symptoms in memory for a known disease. The objective of IA is not to find an exact match, but to eliminate those diseases which fail to meet the search criteria. Both &quot;Yes&quot; and &quot;No&quot; answers are specifically encoded to eliminate unrelated diseases. Consider a patient with a disease, who approaches a computer diagnostic session. Let us say the computer has a list of 200 diseases, which can be identified by 1000 symptom specific questions stored in the system. (Many diseases will share common symptoms). In practice, on an average, each disease may answer &quot;Yes&quot; to 20 of the 1000 questions.
<p><b>More clues in elimination. </b>But, upto 200 &quot;Yes&quot; answers may justify the elimination of the disease, since most symptoms will promptly point to specific groups of diseases, excluding others. The conventional expert system looks only for &quot;Yes&quot; answers. It will match the answers for the disease of the patient to just 20 of the 1000 questions. For this patient, 980 answers will not take the search forwards. But for IA, every &quot;Yes&quot; answer can eliminate up to 20 percent of the diseases. Elimination of a disease also removes its related questions. The elimination process will yield speedy results even for &quot;No&quot; answers. IA will identify the disease long before the 20 relevant questions for the disease are exhausted by swiftly purging any remaining alternatives. In pattern recognition, an elimination procedure is unbelievably faster than one which seeks an exact match.
<A NAME="3"><p align="center"><H1>Instant Recognition</H1>
<p><b>A logic for instant recognition. </b>The speed of the elimination process is even more striking for IA in a special situation. When IA identifies a special condition, its recognition process is virtually instantaneous. Its memory stores the relationships of all diseases to symptoms. Suppose only one disease has a &quot;Y&quot; relationship and all others, an &quot;N&quot; relationship to an exceptional symptom. The symptom is unique to the disease. Then, an &quot;Yes&quot; answer to this symptom eliminates all &quot;N&quot; diseases, leading immediately to recognition. The symptom indicates the disease. It is recognised in a single step of massive elimination. The process is logical. It evaluates every disease in its database against a single clue from one symptom. A doctor may walk into a surgery and instantly attend to a patient suffering from a heart attack. He may not even ask a question. With minimum visual clues, he instantly identifies a single disease from his &quot;known database&quot; of thousands of diseases. He instantly recognises a single pattern in a maze of interweaving patterns. IA may be imitating the logic of this recognition process.
<p><b>Unique features can identify a pattern. </b>The IA logic does not seek an exact match, but concentrates on the elimination of alternate possibilities. Elimination is most effective when there are unique features. It is a practical strategy for recognition in nature. All the recognised objects in our environment are unique. Despite millions of shared characteristics, they also have individual qualities. Even where patterns shift constantly, some characteristics remain stable. Consider a face in a newspaper cartoon. It contains the barest minimum of information - a few lines which define the edges of facial features. But a public figure is identified by just the curve of a nose. The context of being in the newspaper eliminates all ordinary people. The turn of the nose eliminates all politicians with straight noses. Unique features and elimination can determine the outcome. Massive amounts of data are not evaluated. A few clues. Recognition is virtually instant. Elimination based on uniqueness can achieve logical and acceptable recognition.
<p><b>IA imitates parallel processing. </b>With the discovery of the spreadsheet, it became possible for computers with single processors to imitate one characteristic of parallel processing. Even if a spread sheet has thousands of cells, a single entry in one cell is instantly reflected in all the related cells. Thousands of serial calculations appear to the user as a single parallel calculation. Logically, the spreadsheet can have billions of cells and a sufficiently powerful processor can still deliver this result. The spreadsheet is holistic, since every cell reflects the current re-calculated position. IA is similar. By evaluating the results of a single answer on all the diseases in its database, it is holistic and imitates parallel processing. Logically, IA too can produce instant recognition in any size of search space. Any unique symptom can enable IA to instantly identify one among several thousand diseases. If IA is to attempt a problem on the scale of the human nervous system, the only limitation will be the practical problem of data entry.