Learning Cross-Modal Spatial Transformations through Spike Timing-Dependent Plasticity

Abstract

A common problem in tasks involving the integration of spatial information from multiple senses, or in sensorimotor coordination, is that different modalities represent space in different frames of reference. Coordinate transformations between different reference frames are therefore required. One way to achieve this relies on the encoding of spatial information with population codes. The set of network responses to stimuli in different locations (tuning curves) constitutes a set of basis functions that can be combined linearly through weighted synaptic connections to approximate nonlinear transformations of the input variables. The question then arises: how is the appropriate synaptic connectivity obtained? Here we show that a network of spiking neurons can learn the coordinate transformation from one frame of reference to another, with connectivity that develops continuously in an unsupervised manner, based only on the correlations available in the environment and with a biologically realistic plasticity mechanism (spike timing-dependent plasticity).

Training a spiking neural network to control a 4-DoF robotic arm based on Spike Timing-Dependent Plasticity

“…Common differences between the various versions of STDP is the amount of change in weights, the dependence or not of the weight update on the current synaptic weight, and the time windows that are examined before and after the spike of the postsynaptic neuron. In this work, we use the symmetric version of STDP [17], which has been found in [6] to be robust to the temporal structure of the input patterns. In this version of STDP, the decision of whether a synapse should be potentiated or depressed does not depend on the temporal order of the events (arrival of the presynaptic spike at postsynaptic neuron before/after the firing of the postsynaptic neuron), but instead on their absolute time difference | t post − t pre |.…”

“…An important feature of the proposed network is that it consists of individual spiking neurons that exhibit realistic behaviour and uses a biologically plausible learning mechanism for modifying the synaptic weights, namely Spike TimingDependent Plasticity (STDP). Spiking neural networks are considered to be biologically realistic and many researchers have lately used them for coordinate transformations ( [5], [6]), object segmentation [7], visual pattern recognition [8], etc.…”

The virtual haptic display: A device for exploring 2-D virtual shapes in the tactile modality

“…The carriage holds the tactile output unit which consists of six standard piezoelectric Braille display elements. Each Braille element (METEC GmbH, Stuttgart, Germany) has eight independently movable plastic actuators arranged in sequently be used in the development of artificial intelligence machines (e.g., Davison & Frégnac, 2006) …”

Optimality in mono- and multisensory map formation

“…Although the precise nature of this teaching signal has not been clarified experimentally, selective neuronal disinhibition, or gating, seems to play a key role (Gutfreund et al 2002;Winkowski and Knudsen 2006). Theoretical studies have confirmed that excitatory and inhibitory teaching input can account for proper map alignment and thus development of multimodal space (Friedel and van Hemmen 2008;Davison and Frégnac 2006). It is, however, only by inhibitory teaching input that an already aligned map can be re-aligned later on (Friedel and van Hemmen 2008).…”