Robustness of Surface EMG classifiers with Fixed-Point Decomposition on Reconfigurable Architecture
1 like95 views
The document discusses the robustness of surface EMG classifiers using fixed-point decomposition on reconfigurable architecture for muscle contraction prostheses. It details challenges in real-time processing and power consumption, proposing an optimized FPGA implementation of algorithms like non-negative matrix factorization and recurrent neural networks. Results indicate that the hardware implementation is significantly faster than software, with LSTM providing improved accuracy for gesture recognition.
Robustness of Surface EMG classifiers with Fixed-Point Decomposition on Reconfigurable Architecture
- 1. Pierandrea Cancian <pierandrea.cancian@mail.polimi.it> Luca Cerina <luca.cerina@polimi.it> Giuseppe Franco <g.franco4@studenti.unipi.it> Marco D. Santambrogio <marco.santambrogio@polimi.it> RAW @ Vancouver 22/5/2018 Robustness of Surface EMG classifiers with Fixed-Point Decomposition on Reconfigurable Architecture
- 2. Context 2
- 4. Muscle Contraction Voltage signal EMG Prostheses 3
- 5. Voltage signal Muscle Contraction Feature Extraction EMG Prostheses 3
- 6. Voltage signal Muscle Contraction Feature Extraction Classification EMG Prostheses 3
- 7. Voltage signal Muscle Contraction Feature Extraction Gesture Actuation Classification EMG Prostheses 3
- 8. EMG Prostheses -- To understand muscle activation we turn our attention at how the brain controls our hands. Many DOF 4
- 9. 4 EMG Prostheses -- To understand muscle activation we turn our attention at how the brain controls our hands. Many DOF Complex control
- 10. 4 EMG Prostheses -- To understand muscle activation we turn our attention at how the brain controls our hands. Many DOF Complex control Stereotyped Building Blocks Cheung, Vincent CK, et al. "Muscle synergy patterns as physiological markers of motor cortical damage." (2012)
- 11. 5 Challenges -- Meeting real time constraints in Embedded devices -- Power consumption for all day activity and small batteries
- 12. 6 Our Proposal -- Optimized FPGA implementation of the NMF -- Exploration of fixed point architecture implementation -- Machine learning algorithm for the classification (in SW)
- 13. 6 Outline Feature Extraction (HW) Feature Classification (SW) Non-Negative Matrix Factorization Support Vector Machine Recurrent Neural Network
- 14. 6 Outline Feature Extraction (HW) Feature Classification (SW) Non-Negative Matrix Factorization Support Vector Machine Recurrent Neural Network
- 15. 7 Non-Negative Matrix Factorization -- Non Negative Matrix Factorization allows to extract the Muscle Synergies from input signal Input Samples Muscle Synergies Activation Coefficients
- 16. 8 Non-Negative Matrix Factorization W.shape -> (m, r) H.shape -> (r, n) V.shape -> (m, n) W and H are obtained iteratively subject to the following: Several implementations exist High cost per iteration Low cost per iteration
- 17. 9 Multiplicative Update H matrix W matrix H stop W stop && Final stop
- 18. 10 Proposed Solution Algorithm The inner loop is unrolled and the multiplication are done in parallel. (inner dimension = 8)
- 19. 10 Proposed Solution Algorithm The inner loop is unrolled and the multiplication are done in parallel. (inner dimension = 8) The inner loop is partially unrolled and some of the multiplications are executed in parallel. (inner dimension = 256)
- 20. 10 Proposed Solution Algorithm The norm of the matrices is executed in parallel
- 21. 10 Proposed Solution Algorithm The norm of the matrices is executed in parallel Data is consumed as soon as it becomes available
- 22. 12 Fixed Point Solution Floating point Sign 8-bit Exponent 23-bit Mantissa Fixed point Sign Integer Decimal -- Fixed point arithmetic requires less resources and allows faster computation Arbitrary
- 23. 12 Fixed Point Solution Floating point Sign 8-bit Exponent 23-bit Mantissa Fixed point Sign Integer Decimal Not used 4-bit 28-bit -- Fixed point arithmetic requires less resources and allows faster computation
- 24. 6 Outline Feature Extraction (HW) Non-Negative Matrix Factorization Support Vector Machine Feature Classification (SW) Recurrent Neural Network
- 25. 13 Classification 2 -- Easy -- ‘’just the gesture’’ -- 3 repetition per gesture -- Hard -- ‘’reach + gesture’’ -- 12 repetitions per gesture Benchmark: 2 different datasets E.Pirondini, M.Coscia, J.Minguillon, J.d.R.Milla, D.Ville and S. Micera, “Eeg topographies provide subject-specific correlates of motor control”,2017. R. N. Khushaba and S. Kodagoda, “Electromyogram (emg) feature reduction using mutual components analysis for multifunction prosthetic fingers control”, 2012.
- 26. 13 Classification 2 -- Easy -- ‘’just the gesture’’ -- 3 repetition per gesture -- Hard -- ‘’reach + gesture’’ -- 12 repetitions per gesture Benchmark: 2 different datasets E.Pirondini, M.Coscia, J.Minguillon, J.d.R.Milla, D.Ville and S. Micera, “Eeg topographies provide subject-specific correlates of motor control”,2017. R. N. Khushaba and S. Kodagoda, “Electromyogram (emg) feature reduction using mutual components analysis for multifunction prosthetic fingers control”, 2012. Signals have dynamic nature!
- 27. 14 Classification Recurrent Neural Networks (RNN)Support Vector Machines (SVM) 2 Classifiers are considered
- 28. 15 Classification -- The use of Recurrent Neural Networks intrinsically exploit time dependencies in learning -- Long Short Term Memory (LSTM) are particular RNNs that use complex memory cells in place of neurons 2 Forget gate Input gate Output gate
- 29. 15 Key Performance Indicators 2 -- Speed for responsive movement and smooth motion -- Resources to minimize power consumption -- Accuracy to guarantee the correct gesture at any time Created by Eucalyp from the Noun Project
- 30. 16 Results: Speed-Up ARM <= 650Mhz Cortex A9 FPGA <= pynq-z1 HW Fixed point design About 20x times faster than SW Slightly faster than floating point
- 31. 17 Results: Resources Fixed Point Design Less DSPs More FFs and LUTs
- 32. 18 Results: Fixed point influence Fixed Point Negligible accuracy degradation
- 33. 19 Results: accuracy LSTM network Up to 19.4% more accurate than vanilla SVM
- 34. 20 Sum Up Conclusions -- Our HW implementation runs 20x times faster than the comparable CPU -- LSTM provide an accuracy improvement of ~12% in every class -- Assessed the advantages of the fixed point architecture Future Work -- Implement the LSTM network in FPGA -- Test the device on real patients -- Assess performance and accuracy with less bits
- 35. 21 Pierandrea Cancian <pierandrea.cancian@mail.polimi.it> Luca Cerina <luca.cerina@mail.polimi.it> Giuseppe Franco <g.franco4@studenti.unipi.it> Marco D. Santambrogio <marco.santambrogio@polimi.it> necst.it/ slideshare.net/necstlab/ facebook.com/groups/ReconfigurableArchitecturesWorkshop You can find us at: