Learning to approximate functions

using niobium doped strontium titanate memristors

Thomas Tiotto, Jelmer Borst & Niels Taatgen
CogniGron (Groningen Cognitive Systems and Materials Centre), University of Groningen, The Netherlands

Introduction The learning rules The experiments

Memristors are a novel fundamental two-terminal circuit Existing supervised and unsupervised learning algorithms We simulate 30 seconds of neuronal activity in response to
element whose resistance value (PES², BCM and Oja) are adapted in order to modulate the a multidimensional input signal.
depends both on the past state of memristor resistance (only results for PES are shown).
the device and on the input current. Three-quarters of the time is dedicated to learning and the
At each timestep, discrete voltage pulses are applied to last eight seconds to testing the discovered weights.
This change in resistance resembles the potentiation and one of the memristors in each pair in order to increase its
depression of synapses in the brain. conductance. The pre-synaptic neuronal ensemble is fed either a band-
limited white noise signal ( ) or a set of uniformly phase-
There is strong research interest in integrating memristors This iterative process minimises the error between the pre- shifted sine waves ( ).
into neuromorphic machine learning models. and post-synaptic neuronal ensembles’ representations of
the original signal. The post-synaptic ensemble is trying to either represent
the same signal as the pre-synaptic ensemble or the square.
The Question 1. e 2.
Results

pr
o st
Can we build a model that uses a particular niobium p
E>0

3 dimensions
doped strontium titanate (SrTiO₃) memristor¹ to support

100 neurons

²
learning multidimensional functions and their

y=
transformations?
post
The model

20 dimensions
200 neurons
t pre
The model is built using the Nengo Brain Maker package.

y=
Each artificial synapse is composed of a “positive” and a 1. The error is positive 3.
“negative” simulated SrTiO₃ memristor.
2. Find contributing
The weight of the connection is given by the difference in pre-synaptic neuron

50 dimensions
1000 neurons
the normalised conductances of the two paired memristors. 3. Pulse its inhibitory M₋↑

y=
The initial state of the memristive devices is unknown as is memristors
the result of each update, through the addition of 15% 4. The connection weights
noise. decrease
5. The error is minimised
M₊ Conclusions
t
M₊ The training yields a set of memristor conductances that,
4. 5. e used as weights, enable the post-synaptic ensemble to
represent functions of the pre-synaptic signal.
pr
G ₋
M₋ G₊ - st
po These weights are found using only discrete updates and
w₁₁ ... local knowledge.
The results hold for both periodic and random high-
dimensional input signals, if the neuronal ensembles are
... w₃₃ large enough to have sufficient representational power.
The learning is robust to the presence of hardware noise
Pre Weights Post t E→0 t and device-to-device variation.

References Acknowledgments More information

We thank Anouk Goossens and prof. Tamalika Banerjee (Zernike Institute for Advanced Contact me: t.f.tiotto@rug.nl
¹Goossens et al., 2018, Journal of Applied Physics 124 Materials, University of Groningen) for providing the memristor experimental data used as a My personal page: bit.ly/2W2dbUQ
²MacNeil, Eliasmith, 2011, PLoS One 6(9) basis for the simulated model. Download the code: bit.ly/3egJeXm