1
2015 International Joint Conference on Neural Networks (IJCNN) July 12 – July 17, 2015, Killarney, Ireland 2015 International Joint Conference on Neural Networks Arithmetic Computing via Rate Coding in Neural Circuits with Spike: triggered Adaptive Synapses Sushrut Thorat, Bipin Rajendran Indian Institute of Technology : Bombay Addition of spike infos Arithmetic operations were previously implemented using neuronal preBsynaptic transfer functions. We implement them by controlling the neural network connectivity B> Network topology, and Synaptic weight adaptation. Our circuits encode and process information in the spike rates that lie between 40−140 Hz. The synapses in our circuit obey simple, local and spikeBtime dependent adaptation rules. Spike rate response to input DC AEIF RS Neuron Linearised Neuron Linearisation: Spike Info (sout) = Spike Rate (fout) : 14.55 Linear Operations: SNN for scaling and offsetting Non:Linear Operations: s1 a s2 b = exp(a log s1 + b log s2) Spike rate response Spike rate response calculation Iin= w12(βs1 + γ) s2 α(Iin+ Ib) s2 w12αβs1 + α(Ib+ γw12) The relaNonship is approximate as Iin is not a DC current. Weight adaptation Exponential spike rate response Logarithmic Spike rate response Multiplication Scheme • Add logs of two spike rates. • Match the range of log addiNon, and domain of exp, and implement exp. • 0.5 X 0.1862 X 10.73 = 0.99 SNN for Exponentiation Exponentiation scheme • Expected SuperBlinear response. •wexp should increase with increase in s1. • Weight AdaptaNon Rule: Weight Adaptation Rule: Multiplication SNN Square of Spike Rate Application to Signal Estimation (Echolocation): Reciprocal of Spike Rate Noise Resilience: Multiplication response to Noise with CV 0.1 Multiplication of Signal and Echo Max. Spike response curve Other Operations •s1 a s2 b = exp(a log s1 + b log s2) • We can theoretically implement any power law. The SNNs for Exp and Log have to be tuned appropriately. • Using the linear operations, we can implement polynomial operations on spike trains. Conclusion: • The building blocks we have designed in this paper can perform the fundamental operations – addition, subtraction, multiplication and division, as well as other nonBlinear transformations such as exponentiation and logarithm for time dependent signals in realBtime, in presence of noise. • Though our circuits use the AEIF neuron model, the outlined design methodology can be readily used to design SNN circuits that use other spiking neuron models. • We have illustrated the power of these circuits to perform complex computations based on the frequency of spike trains in realBtime, and thus they can be used in a wide variety of hardware and software implementations for navigation, control and computing. Spike rate response to input DC

ijcnn2015 poster thorat · 2020. 4. 3. · 2015%International%Joint%Conference%on%Neural%Networks%(IJCNN) July%12%–%July%17,%2015,%Killarney,%Ireland% 2015%International%Joint%Conference%on%Neural%

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: ijcnn2015 poster thorat · 2020. 4. 3. · 2015%International%Joint%Conference%on%Neural%Networks%(IJCNN) July%12%–%July%17,%2015,%Killarney,%Ireland% 2015%International%Joint%Conference%on%Neural%

2015%International%Joint%Conference%on%Neural%Networks%(IJCNN)July%12%–%July%17,%2015,%Killarney,%Ireland%

2015%International%Joint%Conference%on%Neural%Networks

Arithmetic)Computing)via)Rate)Coding)in)Neural)Circuits)with)Spike:triggered)Adaptive)Synapses

Sushrut)Thorat,)Bipin)Rajendran)Indian)Institute)of)Technology):)Bombay

Addition)of)spike)infos

Arithmetic%operations%were%previously%implemented%using%neuronal%preBsynaptic%transfer%functions.%We%implement%them%by%controlling%the%neural%network%connectivity%B>%Network)topology,%and%Synaptic)weight)adaptation.%Our%circuits%encode%and%process%information%in%the%spike%rates%that%lie%between%40−140%Hz.%The%synapses%in%our%circuit%obey%simple,%local%and%spikeBtime%dependent%adaptation%rules.%%

Spike)rate)response)to)input)DC)AEIF)RS)Neuron Linearised)Neuron

))Linearisation:) Spike)Info)(sout))=)Spike)Rate)(fout)):)14.55))))))

))Linear)Operations:)

SNN)for)scaling)and)offsetting

))Non:Linear)Operations:) s1as2b)=)exp(a)log)s1)+)b)log)s2)))

Spike)rate)responseSpike)rate)response)calculation

⟨Iin⟩ = w12(βs1 + γ)

s2 ≈ α(⟨Iin⟩ + Ib′ )

s2 ≈ w12αβs1 + α(Ib′ + γw12)

The%relaNonship%is%approximate% as%Iin%is%not%a%DC%current.%

Weight)adaptation Exponential)spike)rate)response

Logarithmic)Spike)rate)response Multiplication)Scheme

• Add%logs%of%two%spike%rates.%

• Match%the%range%of%log%addiNon,%and%domain%of%exp,%and%implement%exp.%

• 0.5%X%0.1862%X%10.73%=%0.99

SNN)for)Exponentiation Exponentiation)scheme

• Expected%SuperBlinear%response.%

• wexp%should%increase%with%increase%in%s1.%

• Weight%AdaptaNon%Rule:%

Weight%Adaptation%Rule:

Multiplication)SNN Square)of)Spike)Rate

))Application)to)Signal)Estimation)(Echolocation):)Reciprocal)of)Spike)Rate ))Noise)Resilience:)Multiplication)response)to)Noise)

with)CV)0.1Multiplication)of)Signal)and)Echo Max.)Spike)response)curve

Other)Operations

• s1as2b%=%exp(a%log%s1%+%b%log%s2)%%• We%can%theoretically%implement%any%power%law.%The%SNNs%for%Exp%and%Log%have%to%be%tuned%appropriately.%

• Using%the%linear%operations,%we%can%implement%polynomial%operations%on%spike%trains.

Conclusion:%%• The%building%blocks%we%have%designed%in%this%paper%can%perform%the%fundamental%operations%–%addition,%subtraction,%multiplication%and%division,%as%well%as%other%nonBlinear%transformations%such%as%exponentiation%and%logarithm%for%time%dependent%signals%in%realBtime,%in%presence%of%noise.%

• Though%our%circuits%use%the%AEIF%neuron%model,%the%outlined%design%methodology%can%be%readily%used%to%design%SNN%circuits%that%use%other%spiking%neuron%models.%• We%have%illustrated%the%power%of%these%circuits%to%perform%complex%computations%based%on%the%frequency%of%spike%trains%in%realBtime,%and%thus%they%can%be%used%in%a%wide%variety%of%hardware%and%software%implementations%for%navigation,%control%and%computing.%%

Spike)rate)response)to)input)DC)