Loading...

A Novel Spiking Neural Network Structure for Active Learning Method Fuzzy Algorithm, (Spike-IDS)

Firouzi, Mohsen | 2011

1007 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 41843 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Bagheri Shouraki, Saeed
  7. Abstract:
  8. Human brain is one of the most wonderful complex machine which is designed for ever. A huge complex network, composed of neurons as tiny biological and chemical processors which are distributed and work together as a super parallel system to do control and vital activities of human body. Today the main secrecies of operation mechanism in individual neurons as fundamental elements of brain are reasonably understood, but network interactions of this wonderful processors and full understanding of information coding in brain seems elusive and remains as a big challenge in many interdisciplinary fields of science, from biology to cognitive science and engineering.
    Thus human brain learning simulation and hardware implementation is one of the most interesting research areas in order to make artificial brain and exploit human brain abilities. Active Learning Method in brief ALM is one of the accomplished researches in this field. ALM is an adaptive-recursive fuzzy learning algorithm which is inspired by some behavioral features of human brain and active learning ability of human. In this algorithm, a Multi-Input Multi-Output system is interpreted as a fuzzy combination of some simpler Single-Input Single-Output systems. Each SISO system has been expressed as xi-y grid plane which is called IDS unit, consists of projected data points corresponding to interval domains of another input variables. So a complex system has been broken down into simpler concepts to acquire information in more comprehensible form, the way that consumes less energy to acquire information and obtain knowledge. This viewpoint of ALM to human learning process is compatible with reductionism concept in philosophy of mind in which a complex system is supposed as sum of its parts. In other words this fact can be represented more specifically in causal reductionism, which implies that the causes acting on the whole are simply the sum of the effects of the individual causalities of the parts.
    ALM has a basic operator which is called Ink Drop Spread in brief IDS, which is inspired by non-exact operation paradigm in brain, whether in hardware level or inference layer. This operator enables fine grained tunable knowledge extraction from information which is captured by IDS units as sensory layer of ALM. IDS units as partial knowledge are fed by IDS operator to extract two important features in partial knowledge space: Narrow trajectory which describes input-output characteristic of IDS units and Spread value which shows importance degree and effectiveness of partial knowledge in overall system. Finally these features are consolidated by inference layer of algorithm to make final modeling surface. In this work we propose a novel spiking neural network structure (Spike-IDS) for ALM which is inspired by biological neural networks. In proposed neural structure, extracted knowledge by IDS units can be captured and stored in the form of Hebbian type Spike-Time Dependent Synaptic Plasticity as is the case in the brain. Spike-IDS comprises multiple delayed synaptic interconnections of SRM neurons in which synaptic weights are adjusted by a kind of Spike Time Dependent Hebbian Reinforcement algorithm. From another point of view, proposed structure is a hybrid system in which spiking neural networks capture and prepare partial knowledge for inference layer of system. Inference unit is a fuzzy rule base that unifies partial knowledge in order to make overall modeling and extract human expertness. This work is a good reminder to need for unification of brain studies in different point of view. Validation results show that Spike-IDS has good ability in modeling and classification problems and can extract Narrow and Spread in IDS units as well as IDS operator. Also to assess performance of Spike-IDS in classification problems, three different type of classification problems are used. The results show Spike-IDS good ability in classification problems.
    It is worth to mention that because of matrix like architecture of IDS, it is memory hungry and makes ALM slow. So an arithmetical form of IDS called Replacing IDS is presented in this work with same functionality and better performance. Also novel pipeline digital architecture of RIDS is proposed with high throughput and high speed learning process and simple hardware structure with more flexibility and scalability in comparison with another architectures. Finally an application benchmark is used for validation process
  9. Keywords:
  10. Active Learning ; Ink Drop Spread (IDS)Operator ; Piplined Replacing Ink Drop Spread (RIDS) ; Hebbian Time Dependent Plasticity (TDP) ; Brain Learning Simulation

 Digital Object List

 Bookmark

No TOC