第116回音楽情報科学研究会
MMアルゴリズムの説明を追加しました.
English title: Nonnegative Matrix Factorization Based on Complex Laplace Distribution
Authors: H. Tanji, T. Murakami, H. Kamata
Institution: Meiji University
Presented in IPSJ Music and Computer 116th Domestic Workshop, Aug. 2017.
Detail of MM algorithm for Laplace-NMF is added to the presented slide.
Several recent papers have explored self-supervised learning methods for vision transformers (ViT). Key approaches include:
1. Masked prediction tasks that predict masked patches of the input image.
2. Contrastive learning using techniques like MoCo to learn representations by contrasting augmented views of the same image.
3. Self-distillation methods like DINO that distill a teacher ViT into a student ViT using different views of the same image.
4. Hybrid approaches that combine masked prediction with self-distillation, such as iBOT.
Effective Optimization Algorithms for Blind and Supervised Music Source Separation with Nonnegative Matrix Factorization
長倉研究奨励賞第三次審査,20分間の研究概要説明
内容は自身の学位論文の一部に相当
Several recent papers have explored self-supervised learning methods for vision transformers (ViT). Key approaches include:
1. Masked prediction tasks that predict masked patches of the input image.
2. Contrastive learning using techniques like MoCo to learn representations by contrasting augmented views of the same image.
3. Self-distillation methods like DINO that distill a teacher ViT into a student ViT using different views of the same image.
4. Hybrid approaches that combine masked prediction with self-distillation, such as iBOT.
Effective Optimization Algorithms for Blind and Supervised Music Source Separation with Nonnegative Matrix Factorization
長倉研究奨励賞第三次審査,20分間の研究概要説明
内容は自身の学位論文の一部に相当
Learning the Statistical Model of the NMF Using the Deep Multiplicative Updat...Hiroki_Tanji
?
This document presents the Deep Multiplicative Update Algorithm (DeMUA) for Nonnegative Matrix Factorization (NMF). DeMUA uses a neural network to represent the statistical model and update rules for NMF. It is applied to audio denoising and supervised signal separation tasks. Experimental results show DeMUA can learn complex distributions and achieve better performance than conventional statistical models of NMF.
A Generalization of Laplace Nonnegative Matrix Factorizationand Its Multichan...Hiroki_Tanji
?
This paper proposes a generalization of nonnegative matrix factorization (NMF) and multichannel NMF based on the complex Bessel distribution. This distribution generalizes three statistical models, including the Gaussian, exponential-function Laplace, and Bessel-function Laplace distributions. Optimization algorithms are derived for the proposed Bessel-NMF and Bessel-multichannel NMF. Simulations on music signal separation show the proposed method achieves better source-to-distortion ratio improvements than competing methods, demonstrating the effectiveness of modeling super-Gaussian observations.
Laplace Nonnegative Matrix Factorization with Application to Semi-supervised ...Hiroki_Tanji
?
1. Laplace Nonnegative Matrix Factorization (NMF) is a method that approximates observed nonnegative spectrograms using basis vectors and activation coefficients, minimizing a cost function based on the Laplace distribution.
2. The paper formulates NMF using complex Laplace distributions and derives update rules that guarantee convergence. This provides a novel statistical interpretation of NMF based on Itakura-Saito divergence.
3. Simulation results show that Laplace-NMF has favorable abilities for fitting complex-Laplace distributed data and performs better than other methods for semi-supervised audio denoising when noise follows a Laplace distribution.
2018/8/27 電気学会システム研究会
English title: A Majorization-Minimization-Based Kalman Filter with Hyperbolic Secant Measurement Noise
Authors: H. Tanji, T. Murakami, H. Kamata
Institution: Meiji University
Presented in Technical Meeting on "Systems", IEE Japan
第32回 信号処理シンポジウム
English title: Bayesian Binary Latent Feature Model Using Laplace Likelihood
Authors: H. Tanji, T. Murakami, H. Kamata
Institution: Meiji University
Presented in IEICE technical group on signal processing (SIP) 32nd Symposium, Nov. 2017.
Collapsed variational Bayes zero (CVB0) inference works well even in
Indian buffet process (IBP) -like Bayesian model.
Nonparametric Bayesian models for AR and ARX identification (CSPA 2016)Hiroki_Tanji
?
Presented in 12th IEEE Colloquium on Signal Processing and its Applications (CSPA 2016).
CSPA 2016 Best Paper Award.
The paper is available from IEEE Xplore.
3. 研究背景
非負値行列因子分解
Nonnegative matrix factorization; NMF [Lee(1999)]
? 非負行列(すべての要素が負ではない行列)から
頻出パターンを抜き出すための枠組み
? 振幅 or パワースペクトログラムに適用すれば,
個々の音源のスペクトルとアクティベーションに
分解できる [Smaragdis(2003)]
スペクトログラム Y
Time
Frequency
スペクトル W アクティベーション H
Frequency
Time
分解
2 / 17
4. NMFの定式化
一般的なモデル
Y W
H
?Y=
観測データ [ymn] 基底 [wmk] 重み [hkn] 推定値 [?ymn]
振幅 or パワースペクトルの加法性 いずれかを仮定
Y ?Y1
?Y2 + · · ·+
観測信号の振幅 音源 1 の振幅 音源 2 の振幅
観測信号のパワー 音源 1 のパワー 音源 2 のパワー
3 / 17
5. NMFの定式化
一般的なモデル
Y W
H
?Y=
観測データ [ymn] 基底 [wmk] 重み [hkn] 推定値 [?ymn]
Y と ?Y の乖離度を考慮
評価関数の最小化問題
F(W , H) =
m,n
f(ymn; ?ymn)
Y の統計的な生成過程を考慮
尤度関数の最大化問題
p(Y ; ?Y ) =
m,n
p(ymn; ?ymn)
評価関数 or 尤度関数をどう設定するか
4 / 17
23. 参考文献 I
[Lee(1999)] D.D. Lee and H.S. Seung.
“Learning the parts of objects with nonnegative matrix factorization”.
Nature, 401, pp.788–791, Oct. 1999.
[Smaragdis(2003)] P. Smaragdis and J.C. Brown.
“Non-negative matrix factorization for polyphonic music transcription”.
In Proc. 2003 IEEE International Workshop on Applications of Signal Processing to
Audio and Acoustics (WASPAA), pp. 177–180, Oct. 2003.
[Fevotte(2008)] C. Fevotte, N. Bertin, and J. L. Durrieu.
“Nonnegative matrix factorization with the Itakura-Saito divergence: with
application to music analysis”.
Neural Computation, 21(3), pp.793–830, Sep. 2008.
[Liutkus(2015)] A. Liutkus, D. Fitzgerald, and R. Badeau.
“Cauchy nonnegative matrix factorization”.
In Proc. 2015 IEEE International Workshop on Applications of Signal Processing to
Audio and Acoustics (WASPAA), pp. 1–5, Oct. 2015.
[Yoshii(2016)] K. Yoshii, K. Itoyama, and M. Goto.
“Student’s T nonnegative matrix factorization and positive semidefinite tensor
factorization for single-channel audio source separation”.
In Proc. 2016 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), pp. 51–55, Mar. 2016.
24. 参考文献 II
[Martin(2002)] R. Martin.
“Speech enhancement using MMSE short time spectral estimation with gamma
distributed speech priors”.
In Proc. 2002 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), volume 1, pp. I–253–I–256, May 2002.
[Lee(2008)] B. Lee, T. Kaler, and R.W. Schafer.
“Maximum-likelihood sound source localization with a multivariate complex
Laplacian distribution”.
In Proc. 11th International Workshop on Acoustic Echo and Noise Control (IWAENC),
Sep. 2008.
[Vincent(2006)] E. Vincent, R. Gribonval, and C. Fevotte.
“Performance measurement in blind audio source separation”.
IEEE Trans. Audio, Speech, and Language Processing, 14(4), pp.1462–1469, Jul.
2006.