PSI - Issue 5

Alexander Serov et al. / Procedia Structural Integrity 5 (2017) 1160–1167 Alexander Serov / Structural Integrity Procedia 00 (2017) 000 – 000

1163

4

representations that accumulate with experience. At the time of inference, stored knowledge atoms are dynamically assembled into context-sensitive schemata. Assembly of schemata (activation of atoms) and inference (completing missing parts of the representation) are both achieved by finding maximally self-consistent states of the system that are also consistent with the input. Each schema encodes the statistical relations among a few representational features. During inference, the probabilistic information in many active schemata are dynamically folded together to find the most probable state of the environment. Smolensky hypothesized that there is a procedure for accumulating knowledge atoms through exposure to the environment so that the system will perform the completion task optimally. Realization of this procedure is a key for implementation of Machine Learning methods for the development of Dynamic Neural Networks. Harmony Theory became the basis for creation of Deep Learning architectures, such as Helmholtz Machine, Hinton et al. (1995), Bornschein et al. (2016), and Restricted Boltzmann Machine, Salakhutdinov et al. (2007). In paper Serov (2016) we mentioned that it is the methods of unsupervised Machine Learning that must be the basis for the construction of cognitive technical systems. First stage of self-development of consciousness cannot be realized in supervised mode. Because cognitive system at this stage has no any representations of outer world. That is why Harmony Theory gives a most suitable way for the development of methods of machine learning applicable for DANN architectures. On a very first stage of evolution neural network must be learned with unsupervised methods. And when cognitive system already have created the set of basic representations, DANN may be learned further with supervised Machine Learning techniques. For the needs of Structural Health Monitoring we proposed DANN architecture which is based on the model of Cognitive Sensor. In our previous work, Serov (2016), we proposed the model which may be used as a basis for the development of a new cognitive architecture: model of Cognitive Sensor (CS). We suppose that Cognitive Sensor is an element of a cognitive system which evolves according principles of self-organized systems. Each CS can percept outer world and can construct representations of this world in the range of perception which is provided by design of sensor. Architecture of Cognitive Sensor includes three subsystems: sensory subsystem, pre processing subsystem and logical subsystem. Sensory subsystem is responsible for measurements of scalar value x which characterizes some property of external world. Perception of CS is one-dimensional, and we assume that x can take values from some bounded set of values. Pre processing subsystem realizes three main functions. First, it makes the discretization of the input stream of values coming from sensory subsystem. Second, it makes the ordering of input values sampled in the stream according to their position within the stream. And third, it makes pre processing of sampled and ordered set of values. Logical subsystem can store and process the sequence of preprocessed values. A feature of logical subsystem is that it is capable of processing only a discrete set of values. Overall operating logic of CS is based on the ability to compare the values that x takes at different positions within the stream of data which must be analyzed. Mathematical model of Cognitive Sensor may be formulated as a non-deterministic finite state automaton: Kripke model. In this case we represent Kripke model as 4-tuple of the following type: K = ( S , I , R, F ), where S is the finite set of states of automaton; I is the set of initial states: I S  ; R is the set of transitions between the states: R S S   , where s S s S     * , so that   s s R  * , ; F is the function which makes labeling of states. Here term “state” has the meaning of the representation of world whic h was previously recorded in the experience of perception of CS as something which can be distinguished by CS from other representations. Each state s and each transition between states ( s , s* ) are characterized in CS model by statistics of data stream processing, i.e. by the number of times that a given state or transition was observed during experience of perception. Represented model has the following features. First, this model is a non-stationary model. Contents of the sets S , R and their statistics depend upon the time. This feature represents the ability of Cognitive Sensor to perform learning in data streams. Second, both these sets depend upon the type of discretization of input data stream. In accordance with principles of Artificial Subjective Reality (ASR), Serov (2016) , the memory of CS doesn’t include any representations of perceived world at the initial state of automaton: the set I is empty. 3. Model of Cognitive Sensor