direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Es gibt keine deutsche Übersetzung dieser Webseite.

Search in All Publications

Suche nach Publikationen

All Publications by GRK Members

Neuromorphic classifier microcircuits
Zitatschlüssel Schmuker2010
Autor Schmuker, M., Häusler, C., and Nawrot, M.
Jahr 2010
ISSN 1662-5188
DOI 10.3389/conf.fncom.2010.51.00048
Journal Front. Comput. Neurosci. Conference Abstract: Bernstein Conference on Computational Neuroscience.
Jahrgang 4
Nummer 00048
Zusammenfassung Insects can identify odorant stimuli in a fast and reliable manner. For example, honeybees can be trained to perform astonishing odorant discrimination experiments (Giurfa 2007)⁠. man of the neuronal circuits involved in this task are described in the experimental literature. We virtually dissect these circuits, implement them in spiking neuronal network models analyze their contribution to the performance of a neuronal implementation of a probabilistic classifier (Soltani et al. 2010)⁠. Our ultimate aim is to implement fast and powerful neuromorphic classification devices based on these circuits. In the olfactory system, primary receptor neurons project to the antennal lobe, where they synapse onto their downstream partners in compartments called glomeruli. The antennal lobe network is characterized by strong lateral inhibitory interactions between glomeruli, which makes an impact on information processing (Wilson et al. 2006)⁠. We illustrate how lateral inhibition in a network model with spiking neurons enhances separability of stimulus patterns by increasing contrast between input dimensions. However, at this stage, the system is limited to linearly separated partitions of the input space, which is a major shortcoming since real-world tasks often require non-linear separation. From the glomeruli, projection neurons send their axons to kenyon cells in the mushroom body, where multimodal integration takes place, and stimulus associations are formed (Heisenberg 1998)⁠. Neuronally, this stage is characterized by a massive fan-out of neurons; In the honeybee, about 900 projection neurons synapse onto 160,000 kenyon cells. Moreover, these connections are made within small microcircuits (Ganeshina et al. 2001)⁠. We demonstrate how these microcircuits can create non-linear transformations of the input patterns. Together with the fan-out organization of the mushroom body, non-linearly separable data is transformed into a higher-dimensional space, in which linear separation is possible, similar to the working principle of support vector machines. At each stage of the model, we use a two-dimensional toy data set to illustrate the processing principle and the classification problem. In addition, we test the performance of the neuronal classifier on real-world data sets, such as an odorant data set (Schmuker et al. 2007)⁠, and various benchmark data sets from the machine learning literature. All modeling code is written in PyNN, which allows direct execution on neuromorphic hardware (Davison et al. 2008)
Link zur Originalpublikation Download Bibtex Eintrag

Nach oben

Import Publication

Upload BibTeX

Alle Einträge nach BibTex exportieren

Zusatzinformationen / Extras


Schnellnavigation zur Seite über Nummerneingabe