Analog synaptic circuit design competition

We are going to learn to design analog weight-update synaptic circuits for implementing spike-based learning in neural processing systems. One of our goals will be to design synaptic update circuits independent of the absolute value of the weight voltage. We will then investigate the implementation of different types of learning rules that have different synaptic weight update strategies (including weight-dependent ones). I will present preliminary attempts with current-mode analog VLSI designs, discussing over their advantages and drawbacks, and then open the discussion for new exciting ideas. The "best" synapse circuit will win a prize (to be determined) at the end of the workshop. Looking forward to brainstorming over novel ideas!


The Main objective of this work group was to brain storm over different analog circuit architectures in order to translate linear changes in the weight value on the input to linear changes in the efficacy of the synaptic circuit. Many creative ideas arrived. To mention only few of them over threshold source-follower, adaptive receptor circuit and pulse extender were presented.


The idea of logarithmically pulse extender was based on controlling the weight update mechanism in order to compensate exponential dependency of the stored weight value represented as the gate to source voltage of the transistor and the current flowing through it – synaptic efficacy. Unfortunately no final conclusion was met and the idea applied in analog hardware proved to be difficult.

The other idea arrived from Sim Bamford. Using the source follower configuration the lower transistor was biased to high constant value assuring that it is operating in over threshold regime.

source follower

By sweeping linearly the voltage on the input of the upper transistor it is made sure that the drain to source voltage of the lower transistor does not exist the Vds=Vgs-Vth range.

IV curve
Since this condition is met for any input voltage the lower transistor in the source-follower configuration is biased to work in triode region.

However at this point the output stage where the design will be applied has to be mentioned. The synapse itself consists in the differential pair integrator producing exponential spikes on its outputs. The circuit we were to design was supposed to control the current flowing through the bottom current source.

DPI circuit

Therefore connecting the source follower to differential pair in DPI is required. This connection makes the upper transistor of the SF to leave the linear region. It is still not clear if it will affect the rest of transistors in the DPI forcing them to leave the sub threshold operation region and therefore disrupt the DPI´s functionality.

The other idea arrived from Chiara Bartolozzi and it is based on the adaptive receptor circuit.

adaptive receptor circuit
By applying the feed-back loop we assure that the linear changes on the input in current domain are inducing logarithmic changes in voltage representing the stored weight value. When applied to the output transistor they are transformed to linear changes in current domain. This solution requires elaborate parameters tuning of the DPI circuit connected to the output transistor. Nevertheless it promises for acquiring the desired function.

Last modified 4 years ago Last modified on 05/09/15 10:24:31

Attachments (4)

Download all attachments as: .zip