TY  - GEN
A1  - Friedmann, Simon
N2  - This thesis presents a novel, highly flexible approach to plasticity and learning in
brain-inspired computing systems. A classical digital processor was combined with
local analog processing to achieve flexibility and efficiency. In particular, this allows
for the implementation of modulated spike-timing dependent plasticity. The approach
was formalized into an abstract hybrid hardware model. This model was used to
simulate a reward-based learning task to estimate the effect of hardware constraints.
To investigate the feasibility of the proposed architecture, a synthesizeable plasticity
processor was designed and tested using the CoreMark general purpose benchmark
(best score: 1.89 per MHz). The processor was also produced as part of a 65 nm proto-
type chip, requiring 0.14 mm2 of die-area, and reaching a maximum clock frequency of
769 MHz. In a preparatory step a non-programmable plasticity implementation was
developed, that is now part of the operational BrainScaleS wafer-scale system. This
design was later extended with the plasticity processor to implement the proposed
hybrid architecture. Simulations show a speed improvement of 42 % over the non-
programmable variant. By preparation for production, the area requirement for the
digital part is estimated to be 6.2 % of total area.

AV  - public
UR  - https://archiv.ub.uni-heidelberg.de/volltextserver/15359/
Y1  - 2013///
TI  - A New Approach to Learning in Neuromorphic Hardware
KW  - Neuromorphic engineering
KW  -  Belohnungslernen
KW  -  STDP
ID  - heidok15359
ER  -