Skip to content


You're viewing the new version of our site. Please leave us feedback.

Learn more

BMC Neuroscience

Open Access

Artificial grammar recognition using spiking neural networks

  • Philip Cavaco1Email author,
  • Baran Çürüklü1 and
  • Karl Magnus Petersson2
BMC Neuroscience200910(Suppl 1):P352

Published: 13 July 2009


A biologically inspired neocortical model consisting of spiking neurons is designed to perform artificial grammar processing. Building on work in [1], the model is designed to categorize symbol strings as belonging to a Reber grammar [2]. Columnar organization of the cortex is used as the general inspiration of the network [3, 4].

The model consists of an input layer and a recognition layer. Input layer has six DC generator nodes designed to represent the presence of a specific symbol of input. The symbols are all those that the Reber grammar can produce (Figure 1). When a symbol is presented to the input layer, the corresponding node generates a direct current that is fed into the recognition layer. The recognition layer consists of 20 nodes connected in a way designed to sustain activity at around 50 Hz, called the recognition level, when the correct sequence of symbols occurs. Each node is a minicolumn model consisting of 100 neurons, of which 80 are excitatory and 20 inhibitory. Excitatory connections inside the node sustain activity, indicating when input is received. Indication of longer sequences is achieved by combining inputs from nodes representing symbols which can appear after another in the grammar, thus the recognition layer consists of the following 20 nodes: {#, M, T, R, V, X, MT, MV, RM, VT, VX, XT, XV, XR, XM, VXT, VXV, VXR, VXM, OUT}. Three levels are used in which the first layer receives input directly from DC generators, effectively recognizing single character input. The next layer is activated by two-symbol substrings in the grammar. A third layer is activated by subsequent three-symbol substrings. When the activity of the OUT node exceeds 50 Hz, it is said to recognize the string. Each symbol is presented for 500 ms; the next symbol is presented immediately afterwards. The nodes are designed to sustain the activity for 1000 ms after the node received an input. After that time, local inhibition overcomes excitatory activity silences the node.
Figure 1

Finite State Machine (FSM) defining Reber grammar. Minicolumn model activity.


The results from two strings are presented here. One belonging to the grammar '#MTVT#' and one string not belonging to the grammar '#MTRT#'. The OUT node passes the recognition level of 50 Hz, recognizing the grammar string (Figure 2). The other string, is recognized as not being in the grammar (Figure 3). Strings with between 4 and 15 symbols, representing all paths through the grammar FSM produce promising results. Further work will explore different network paradigms and their ability to recognize strings from the grammar.
Figure 2

Grammar string is recognized by the network.

Figure 3

A random string does not excite the OUT node above the recognition level.

Authors’ Affiliations

School of Innovation Design and Engineering, Mälerdalen University
F.C. Donders Centre for Cognitive Neuroimaging, Radboud University of Nigmegen


  1. Petersson KM: Artificial grammar learning and neural networks. Proc Cogn Sci Soc. 2005, 1726-1731.Google Scholar
  2. Reber AS: Implicit learning of artificial grammar. J Verb Learn & Verb Behav. 1967, 6: 855-863. 10.1016/S0022-5371(67)80149-X.View ArticleGoogle Scholar
  3. Çürüklü B: A Canonical Model of the Primary Visual Cortex. 2005, Mälardalen University PressGoogle Scholar
  4. Mountcastle VB: The columnar organization of the neocortex. Brain. 1997, 701-722. 10.1093/brain/120.4.701.Google Scholar
  5. Gewaltig MO, Diesmann M: NEST. Scholarpedia. 2007, 2: 1430.View ArticleGoogle Scholar


© Cavaco et al; licensee BioMed Central Ltd. 2009

This article is published under license to BioMed Central Ltd.