ai-one’s Learning Algorithm: Biologically Inspired Neural Network
– Introduction to HSDS vs ANN in Text Applications
Unlike any of the traditional neural nets, the neural network based on ai-one, the HoloSemantic Data Space neural network (invented by Manfred Hoffleisch) or in short “HSDS”, are massively connected, asymmetrical graphs which are stimulated by binary spikes. HSDS do not have any neural structures pre-defined by the user. Their building blocks resemble biological neural networks: a neuron has dendrites, on which the synapses from other neurons are placed, and an axon which ends in synapses at other neurons.
The connections between the neurons emerge in an unsupervised manner while the learning input is translated into the neural graph structure. The resulting graph can be queried by means of specific stimulations of neurons. In traditional neural systems it is necessary to set up the appropriate network structure at the beginning according to what is to be learned. Moreover, the supervised learning employed by neural nets such as the perceptron requires that a teacher be present who answers specific questions. Even neural nets that employ unsupervised learning (like those of Hopfield and Kohonen) require a neighborhood function adapted to the learning issue. In contrast, HSDS require neither a teacher nor a predefined structure or neighborhood function (note that although a teacher is not required, in most applications programmatic teaching is used to insure the HSDS has learned the content needed to meet performance requirements). In the following we characterize HSDS according to their most prominent features.
Exploitation of context
In ai-one applications like BrainDocs, HSDS is used for the learning of associative networks and feature extraction. The learning input consists of documents from the application domains, which are broken down into segments rather than entered whole: all sentences may be submitted as is or segmented into sub-sentences according to grammatical markers. By way of experimenting, we have discovered that a segment should ideally consist of 7 to 8 words. This is in line with findings from cognitive psychology. Breaking down text documents into sub-sentences is the closest possible approximation to the ideal segment size. The contexts given by the sub-sentence segments help the system learn. The transitivity of term co-occurrences from the various input contexts (i.e. segments) are a crucial contribution to creating appropriate associations. This can be compared with the higher-order co-occurrences explored in the context of latent semantic indexing.
Continuously evolving structure
The neural structure of a HSDS is dynamic and changes constantly in line with neural operations. In the neural context, change means that new neurons are produced or destroyed and connections reinforced or inhibited. Connections that are not used in the processing of input into the net for some time will get gradually weaker. This effect can also be applied to querying, which then results in the weakening of connections that are rarely traversed for answering a query.
The connections between the neurons need not be equally strong on both sides and it is not necessary that a connection should exist between all the neurons (cp. Hopfield’s correlation matrix).
The HSDS is stimulated by spikes, i.e. binary signals which either fire or do not. Thresholds do not play a role in HSDS. The stimulus directed at a neuron is coded by the sequence of spikes that arrive at the dendrite.
Whenever a new input document is processed, new (groups of) neurons are created which in turn stimulate the network by sending out a spike. Some of the neurons reached by the stimulus react and develop new connections, whereas others, which are less strongly connected, do not. The latter nevertheless contribute to the overall connectivity because they make it possible to reach neurons which could not otherwise be reached. Given the high degree of connectivity, a spike can pass through a neuron several times since it can be reached via several paths. The frequency and the chronological sequence in which this happens determine the information that is read from the net
There is no need to define a topology before starting the learning process because the neural structure of the HSDS develops on its own. This is why it is possible to retrieve a wide range of information by means of different stimulation patterns. For example, direct associations or association chains between words can be found, the words most strongly associated with a particular word can be identified, etc.