IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 13, NO. 3, MAY 2002 619 Learning Sensory Maps With Real-World Stimuli in Real Time Using a Biophysically Realistic Learning Rule Manuel A. Sánchez-Montañés, Peter König, and Paul F. M. J. Verschure Abstract—We present a real-time model of learning in the auditory cortex that is trained using real-world stimuli. The system consists of a peripheral and a central cortical network of spiking neurons. The synapses formed by peripheral neurons on the central ones are subject to synaptic plasticity. We implemented a biophysically realistic learning rule that depends on the precise temporal relation of pre- and postsynaptic action potentials. We demonstrate that this biologically realistic real-time neuronal system forms stable receptive fields that accurately reflect the spectral content of the input signals and that the size of these representations can be biased by global signals acting on the local learning mechanism. In addition, we show that this learning mechanism shows fast acquisition and is robust in the presence of large imbalances in the probability of occurrence of individual stimuli and noise. Index Terms—Auditory system, learning, natural stimuli, real time, real world, spiking neurons. I. INTRODUCTION O VER the past years neuroscientists have gained insight in the neural mechanisms responsible for the ability of learning and adaptation in biological systems [1], [11]. The sub- strate of learning in these systems is thought to be provided by the mechanisms which regulate the change of synaptic effica- cies of the connections among neurons [31], [43]. In his seminal work Hebb proposed that neurons which are consistently coac- tivated strengthen their coupling [20] and form associative net- works. Since then many experiments have addressed different mechanisms which regulate changes in synaptic efficacies de- pendent on specific properties of pre- and postsynaptic activity [9], [11]. Based on these experiments, a number of Hebbian learning rules have been proposed with different desirable prop- erties [8], [10], [14], [36], [39]. These learning rules have been considered physiologically re- alistic when they only rely on signals which are available to Manuscript received September 20, 2000; revised July 23, 2001 and January 3, 2002. This work was supported by SPP Neuroinformatics and SNF (Grant 31-51059.97, awarded to P. König). The work of M. A. Sánchez-Montañés was supported by a FPU grant from MEC and Grant BFI2000-0157 from MCyT (Spain). This work was done in part at the Telluride Workshop on Neuromorphic Engineering (1999). M. A. Sánchez-Montañés is with the Institute of Neuroinformatics, ETH/University, 8057 Zürich, Switzerland, and also with the Grupo de Neurocomputación Biológica, ETS de Informática, Universidad Autónoma de Madrid, 28049 Madrid, Spain. P. König and P. F. M. J. Verschure are with the Institute of Neuroinformatics, ETH/University, 8057 Zürich, Switzerland. Publisher Item Identifier S 1045-9227(02)02405-0. the synapse local in time and space. However, recent physio- logical results on neurons in mammalian cortex give a richer picture. These studies demonstrate, first, that an action poten- tial triggered at the axon hillock propagates not only antero- gradely along the axon, but also retrogradely through the den- drites [40], [12]. Second, on its way into the dendrite the action potential may be attenuated or blocked by inhibitory input from other neurons [38], [45]. Third, it has been demonstrated that these backpropagating action potentials directly affect mecha- nisms regulating synaptic plasticity [30] which depends on post- synaptic calcium dynamics [27]. In addition, the dramatic effect of even single inhibitory inputs on the calcium dynamics in the dendritic tree, in particular in its apical compartments, suggests that regulation of synaptic plasticity can be strongly influenced by inhibitory inputs [28]. Thus, the backpropagating action po- tential can make information on the output of the neuron avail- able locally at each of its afferent synapses, and inhibitory in- puts onto a neuron can in turn regulate the effectiveness of this signal. The above described mechanism makes a change in synaptic efficacy dependent on the temporal relation between pre- and postsynaptic activity. In particular, it will be strongly affected by the temporal relation between the inhibition and excitation a neuron receives and its own activity. Neurons which fire with the shortest latency to a stimulus will receive inhibition after they have generated backpropagating action potentials. In this case active synapses can be potentiated [28]. In contrast, neu- rons which fire late to a stimulus would receive inhibition be- fore they have generated a spike. Their backpropagating action potentials are modulated by this inhibition preventing potentia- tion of their active synapses. This dynamic seems to be reflected in the physiology of the visual system where the optimality of the tuning of a neuron seems to be directly reflected in its re- sponse latency to a stimulus [24]. Given the above mechanism this would imply that the optimally tuned neurons prevent fur- ther learning by other neurons in the map. Synaptic plasticity, however, is not only dependent on the dynamics of the local network but also on modulatory signals [2] arising from subcortical structures. For instance, it has been shown that cholinergic and gabaergic neurons in the basal fore- brain, which project to many areas including the cerebral cortex, can strongly regulate mechanisms of synaptic plasticity [50]. These results were obtained in classical conditioning experi- ments where tones were paired with aversive stimuli such as a footshock [51]. Subsequently it was shown that the aversive stimulus could be replaced by direct stimulation of the basal 1045-9227/02$17.00 © 2002 IEEE