Energy-efficient neuromorphic computing with magnetic tunnel junctions

Transcription

1 Energy-efficient neuromorphic computing with magnetic tunnel junctions CNRS/Thales, France Jacob Torrejon Mathieu Riou Flavio Abreu Araujo Paolo Bortolotti Vincent Cros Julie Grollier AIST, Japan Sumito Tsunegi Kay Yakushiji Akio Fukushima Hitoshi Kubota Shinji Yuasa NIST, USA Guru Khalsa Mark Stiles C2N, France Damien Querlioz 1

2 WHAT: Brain-inspired computing for cognitive tasks Neural networks. Already in heavy commercial application. Quasi-static. inputs outputs Dynamic processing. Make use of dynamical properties of devices. Spikes, rates, oscillations, timing, Simulate brain function Voltage Time 2

3 Sophisticated CMOS-based neuromorphic chip development Loihi (Intel) TrueNorth (IBM) BrainScales (Human Brain Project) Brain-like features natural in CMOS local energy source Incoming signal does not power outgoing signal. Connectivity Digital spikes shared communication channels Energy efficiency and complexity is still far from the brain! 3

4 HOW: Augment CMOS with efficient neural devices (MTJs) Features for which CMOS may be inefficient (energy and/or device area) Non-volatility Plasticity (local learning) Stochasticity Oscillators Nano-oscillators P k B T AP Source MRAM Bit Bit Line Line Drain MTJ Superparamagnetic tunnel junctions R AP R AP Resistance 4 current R P R P Time (s) 4

5 Single oscillator reservoir computing Time multiplexed reservoir computing Sine/Square Identification (Intrinsic memory) Spoken digit recognition (Non-linearity) Sine/Square Identification (Delayed feedback memory) 5

6 Feed-forward networks one direction of information flow information flow inputs outputs hidden layers Non-linear nodes (neurons) rearranges spaces to allow classification Train off-line adjust synaptic weights to optimize fit to test data 6

7 Recurrent networks have intrinsic time scales information flow Input time series Output time series hidden layers Output time dependent input becomes time series Training protocol not simple 7

8 Reservoir computing a simply trainable recurrent network reservoir Input time series Output time series Non-linear, Fixed synaptic weights Linear, Trained synaptic weights 8

9 Reservoir computing ring geometry reservoir Input time series to each device Output time series Non-linear, Fixed synaptic weights Linear, Synaptic weights trained off-line 9

10 Reservoir computing time multiplexed single device Input time series outputs t n-1 times mask Time multiplexed input to reservoir t n time t n+1 t n+2 t n+3 reservoir Device at different times gives virtual Devices. output time series Output weights, time dependent, accumulated. 1

11 Spin-torque nano-oscillators: non-linear amplitude dynamics and memory Input (mv) V osc (mv) Time (µs) ) Voltage amplitude (mv) V in = ± 25 mv VV dc current (ma) Torrejon et al., Nature 547, 428 (217). Input: current Output: amplitude of the oscillator s voltage Input Arbitrary Waveform Generator I DC + FeB MgO CoFeB H Diode V osc (t) VV(tt) Input (mv) V osc (mv) VV 1 VV 2 VV 5 VV 7 VV 3 VV 4 VV 6 VV(tt) Time (µs) 11

12 Single oscillator reservoir computing Time multiplexed reservoir computing Sine/Square Identification (Intrinsic memory) Spoken digit recognition (Non-linearity) Sine/Square Identification (Delayed feedback memory) 12

13 Task: recognizing sines from squares at each point in time with a single oscillator Voltage (V).2. Input sine square square sine square sine square Time (µs) Paquot et al, Scientific Reports 2:287 (212) Same inputs 13

14 For extrinsic memory, each node should couple to a few other nodes. Voltage (V) Time (µs) Preprocessing Different input to each (virtual) node τ Voltage amplitude VV N τ VV 1 VV 2 VV 3 VV N VV 1 Time 14

15 Experimental trajectories of oscillators amplitude Preprocessed Input Voltage (mv) Voltage (mv) sine square Time (µs) Oscillator s emitted voltage VV(tt) Time (µs) τ 15

16 Experimental trajectories of oscillators amplitude Preprocessed Input Voltage (mv) 2-2 sine square Time (µs) Oscillator s emitted voltage VV(tt) Voltage (mv) Time (µs) τ 16

17 Different trajectories: data separation is achieved Same input for sine and square Different outputs VV(tt) Voltage (mv) 2-2 Sine Square Time (µs) Voltage (mv) Sine Square Time (µs) 17

18 Trajectories need to be grouped to be classified in sines and squares Voltage (mv) Time (µs) VV(tt) Square 1 Target Sine Time (µs) 18 18

19 Classification: constructing the output for sine, 1 for square Voltage (mv) Time (µs) VV(tt) Square 1 Target VV oooooooooooo = ww ii VV ii Sine computer Time (µs) 19 19

20 Experimental result : RMS = 1% perfect classification of sines and squares waveform with 8 randomly arranged sines and squares Target Reconstructed output Time / τ H = 38 Oe, I DC = 6.4 ma 8 τ per period, 24 nodes, θ = 1 ns 64 first τ for training, 64 next τ for classification 2

21 Single oscillator reservoir computing Time multiplexed reservoir computing Sine/Square Identification (Intrinsic memory) Spoken digit recognition (Non-linearity) Sine/Square Identification (Delayed feedback memory) 21

22 Amplitude (a.u.) Spoken digit recognition (NIST TI-46 corpus) Input: audio file 5 Time (a.u.) "1" Acoustic features Spectrogram or Cochlear Preprocessed input Pre-processing Oscillator Recorded trace Computer Output "1" 8 Spectrogram Success rate (%) With oscillator Utterances for training 22

23 Input: audio file Amplitude (a.u.) Spoken digit recognition (NIST TI-46 corpus) 5 Time (a.u.) "1" Acoustic features Spectrogram or Cochlear Preprocessed input Pre-processing Oscillator No oscillator Recorded trace Computer Output "1" 8 Spectrogram Success rate (%) With oscillator + 7% Without oscillator Utterances for training 23

24 Input: audio file Amplitude (a.u.) Spoken digit recognition (NIST TI-46 corpus) 5 Time (a.u.) "1" Acoustic features Spectrogram or Cochlear Preprocessed input Pre-processing Oscillator No oscillator Recorded trace Computer Output "1" Spectrogram Cochlear Success rate (%) With oscillator + 7% Without oscillator Success rate (%) With oscillator Without oscillator 99.6% Utterances for training Utterances for training 24

25 Spoken digit recognition (NIST TI-46 corpus) Jacob Torrejon-Diaz, Mathieu Riou, Flavio Abreu-Araujo, et al, Nature 547, (217) State of the art: 96 to 99.8 % Spectrogram Cochlear Success rate (%) With oscillator + 7% Without oscillator Success rate (%) With oscillator Without oscillator 99.6% Utterances for training Utterances for training 25

26 Recognition results are sensitive to the noise, an optimal bias area is found 26

27 Single oscillator reservoir computing Time multiplexed reservoir computing Sine/Square Identification (Intrinsic memory) Spoken digit recognition (Non-linearity) Sine/Square Identification (Delayed feedback memory) 27

28 Replace intrinsic memory with delayed feedback New part of reservoir τ Intrinsic memory Feedback memory 28

29 To focus on feedback, nodes only couple to past states of themselves Input u tt (b) (a) Input u(k) τ Random fixed connections iin iin ww iiii τ (c) (b) Preprocessed input J(tt) θ τ (d) (c) τ Recurrent neural network X(tt) θ Oscillation amplitude time traces x(tt) Trained connections W, b Output y(tt) (e) (d) (e) (f) (g) (f) Higher dimension Mapping X(k) Offline linear combination WX k + b Output y(k) θ 29

30 Feedback allows separation of similar inputs without delayed feedback loop Error rate is 1.78% with delayed feedback loop Error rate is.16% 3

31 Single oscillator reservoir computing Augmenting CMOS based neuromorphic circuits with energy-efficient spintronic devices. MTJs already accessible in stat-of-the-art BEOL CMOS Time multiplexed reservoir computing. Memory (intrinsic or delayed feedback) allows context dependent discrimination. Single oscillator achieves state of the art at spoken digit recognition non-linearity. Where to? Small low power oscillators. Efficient coupling. Appropriate algorithms 31