Learning matrix for automatic character recognition, maker: Institute of Communications Engineering and Communications Transmission, 1960, two units each approx. 80 × 112 × 30 cm. KIT, Institute of Information Processing Technology.
The learning matrix was developed by the pioneer Karl Steinbuch (1917–2005) at the Institute of Computer Science by 1960. The device shown here is a demonstration model built at that time. The illuminated display could indicate characters, and the reading unit hung in front of it by hand registered them. By pressing a button, a link could be established between the displayed pattern and a character from the set of characters stored in the learning matrix. This linking process thus first required a decision by a human. Once linked, the character recognition set up in this manner could be reproduced by the system without further intervention. This was the innovative achievement of the learning matrix. By using the word “learning” in its name — a term generally associated specifically with human ability — the learning matrix not only garnered significant attention but also sparked concerns. Its development triggered discussions about the concept of Artificial Intelligence at Karlsruhe Polytechnic in the early 1960s, which term was in use already then. This partly controversial debate attracted not only engineers and scientists but also philosophers, who came to be known as the Karlsruhe School of Technical Philosophy. The establishment of the Karlsruhe Department of Informatics in 1972 was largely based on the activities at the Institute of Communications Engineering and Communications Transmission under Steinbuch’s directorship since 1958. Today, this institute exists at KIT under the name Institute of Information Processing Technology. Another key development strand leading to the KIT Department of Informatics began at the Institute of Applied Mathematics, which collaborated early on with the Nuclear Research Center and also founded the University’s computing center. kn
The success of the early applications of the then-new electronic calculating machines sparked associations with human abilities and were often referred to as the “electronic brain.” This notion provided a global impetus for research on biological nerve functions in combination with communications technology. How does human sensory perception work? It was already known that nerves transmit signals through electrical impulses and that the central nervous system, particularly the brain, contains an extensive number of links. This understanding prompted attempts to implement mechanisms in machines that could simulate biological functions. Reflexes, especially conditioned reflexes as demonstrated by Pavlov, had long been recognized as simple bodily functions. Pavlov showed that the reflex of “gastric acid secretion by food stimulus” could be triggered in dogs by the sound of a bell, even in the absence of food, if the sound had previously been associated with food multiple times. This suggested that dogs could “learn” such associations, implying that the higher nervous system in an animal organism could store such learned connections. Inspired by this model, Karl Steinbuch proposed an adaptive categorizer, which he called a “learning matrix.” This matrix consists of a grid of connection points where associations between rows and columns could be stored. The model shown here was built to illustrate this concept in its simplest form. The core of this model is a relay matrix with 10 rows and 20 columns. A second panel device is used to generate input or display signals, referred to here as “events.” These signals are actually entirely equivalent but they are arranged here in 5 rows with 4 columns each, so that this can represent the 5 capital and lower-case vowels of the alphabet using 20 points of light. A scanner with 20 corresponding photocells detects the light or dark status of each point, generating the column signals — i.e., either 1 or 0. During the so-called “learning phase,” this scanner acts like an eye, feeding these signals into the learning matrix. When one of the 10 rows is selected as the “meaning” and a “learning step” is then performed, which means a memory signal is entered into the latching relays, the event pattern is linked to the selected meaning. This process can be repeated for all 10 event patterns. In the so-called recall phase, it is assumed that the learning phase has been completed, consequently that all event patterns are stored correspondingly in the relay states. When an event pattern is presented again, the relays generate an output signal for each row, with the row that best matches the input pattern producing a maximum value. An extremum detection mechanism can identify this maximum value and output the corresponding meaning associated with the event input. By appropriately configuring the extreme value detection, it is also possible to display patterns that contain minor errors corresponding to each event pattern. During the recall phase, events that are identical or most similar to those encountered in the learning phase are displayed. It is also feasible to display multiple similar meanings by modifying the row output accordingly. Conversely, by setting a high threshold, it is possible to prevent any row in the matrix from producing a sufficient sum signal, resulting in no meaning being displayed. This typically occurs when no learning step has yet been performed for a given event pattern. This model essentially realizes an adjustable assignment between events and meanings, with the extreme value determination displaying this assignment. However, the learning matrix is directed rather than learning autonomously. Additionally, the coding of the events is a crucial aspect of the function. It quickly became clear that automatic character recognition is far more complex than a simple pixel-to-pattern assignment. While the term “learning matrix” is catchy and targeted, it was soon considered controversial, as it overstated the machine’s capabilities, especially when it draws an association with biological organisms and their learning behaviors. Unfortunately, the expected results concerning the learning abilities of automated devices did not materialize. Nevertheless, Wolfgang Hilberg later demonstrated that the learning matrix was a misunderstood precursor to network structures, which today form the basis of artificial intelligence in the form of neural networks. Karl Steinbuch can therefore be regarded as one of the pioneers in this field. Winfried Görke