11-755 Machine Learning for Signal Processing Independent Component Independent Component Analysis y Class 20. 8 Nov 2012 Instructor: Bhiksha Raj 8 Nov 2012 11755/18797 1
A brief review of basic probability Uncorrelated: Two random variables X and Y are uncorrelated iff: uncorrelated iff: The average value of the product of the variables equals the product of their individual averages Setup: Each draw produces one instance of X and one instance of Y instance of Y I.e one instance of (X,Y) E[XY] = E[X]E[Y] E[XY] E[X]E[Y] The average value of X is the same regardless of the value of Y 8 Nov 2012 11755/18797 2
Uncorrelatedness Which of the above represent uncorrelated RVs? 8 Nov 2012 11755/18797 3
A brief review of basic probability Independence: Two random variables X and Y are independent iff: Their joint probability equals the product of their individual probabilities P(X Y) = P(X)P(Y) P(X,Y) P(X)P(Y) The average value of X is the same regardless of the value of Y E[X|Y] = E[X] 8 Nov 2012 11755/18797 4
A brief review of basic probability Independence: Two random variables X and Y are independent iff: The average value of any function X is the same The average value of any function X is the same regardless of the value of Y E[f(X)g(Y)] = E[f(X)] E[g(Y)] for all f(), g() E[f(X) (Y)] E[f(X)] E[ (Y)] f ll f() () 8 Nov 2012 11755/18797 5
Independence Which of the above represent independent RVs? Which represent uncorrelated RVs? Whi h t l t d RV ? 8 Nov 2012 11755/18797 6
A brief review of basic probability p(x) ( ) f(x) The expected value of an odd function of an RV is 0 if The RV is 0 mean The RV is 0 mean The PDF is of the RV is symmetric around 0 E[f(X)] = 0 if f(X) is odd symmetric E[f(X)] 0 if f(X) i dd t i 8 Nov 2012 11755/18797 7
A brief review of basic info. theory T(all), M(ed), S(hort)… ( ) ( )[ log ( )] H X P X P X X Entropy: The minimum average number of bits to Entropy: The minimum average number of bits to transmit to convey a symbol X X T, M, S… M F F M.. Y Y ( , ) ( , )[ log ( , )] H X Y P X Y P X Y , X Y Joint entropy: The minimum average number of bits to convey sets (pairs here) of symbols 8 Nov 2012 11755/18797 8
A brief review of basic info. theory X X T, M, S… M F F M.. Y ( | ) ( ) ( | )[ log ( | )] ( , )[ log ( | )] H X Y P Y P X Y P X Y P X Y P X Y , Y X X Y Conditional Entropy: The minimum average number of bits to transmit to convey a symbol X, y y , after symbol Y has already been conveyed Averaged over all values of X and Y Averaged over all values of X and Y 8 Nov 2012 11755/18797 9
A brief review of basic info. theory ( | ) ( ) ( | )[ log ( | )] ( ) ( )[ log ( )] ( ) H X Y P Y P X Y P X Y P Y P X P X H X Y X Y X Conditional entropy of X = H(X) if X is Conditional entropy of X = H(X) if X is independent of Y ( ( , ) ) ( ( , )[ )[ log l ( ( , )] )] ( ( , )[ )[ log l ( ( ) ) ( ( )] )] H H X X Y Y P P X X Y Y P P X X Y Y P P X X Y Y P P X X P P Y Y , , X Y X Y ( , ) log ( ) ( , ) log ( ) ( ) ( ) P X Y P X P X Y P Y H X H Y X , Y X , Y Joint entropy of X and Y is the sum of the entropies of X and Y if they are independent p y p 8 Nov 2012 11755/18797 10
Onward.. 8 Nov 2012 11755/18797 11
Projection: multiple notes j M = W = P = W (W T W) ‐ 1 W T ( ) Projected Spectrogram = P * M 8 Nov 2012 11755/18797 12
We’re actually computing a score M = H = ? W = M ~ WH H = pinv (W)M 8 Nov 2012 11755/18797 13
How about the other way? M = H = ? ? ? ? U = U = W = W = M ~ WH W = M pinv (V) U = WH 8 Nov 2012 11755/18797 14
So what are we doing here? H = ? W = ? M ~ WH is an approximation Given W , estimate H to minimize error 2 2 arg min || || arg min ( ) H M W H M W H F ij ij H H i j Must ideally find transcription of given notes 8 Nov 2012 11755/18797 15
Going the other way.. H W =? ? M ~ WH is an approximation Given H , estimate W to minimize error 2 2 arg min || || arg min ( ) W M W H M W H F ij ij W H i j Must ideally find the notes corresponding to the d ll f d h d h transcription 8 Nov 2012 11755/18797 16
When both parameters are unknown H = ? W =? approx(M) = ? approx(M) ? Must estimate both H and W to best approximate M Ideally, must learn both the notes and their transcription! 8 Nov 2012 11755/18797 17
A least squares solution 2 , arg min || || W H M W H , F W H Unconstrained For any W,H that minimizes the error, W’=WA, H’=A -1 H also minimizes the error for any invertible A also minimizes the error for any invertible A H H For our problem, lets consider the “truth”.. For our problem, lets consider the truth .. When one note occurs, the other does not T h j = 0 for all i != j h i i j The rows of H are uncorrelated 8 Nov 2012 11755/18797 18
A least squares solution H Assume: HH T = I Normalizing all rows of H to length 1 g g pinv (H) = H T Projecting M onto H Projecting M onto H W = M pinv (H) = MH T WH = M H T H WH M H H 2 , arg min || || W H M W H , F W H 2 T H arg min || || H M M H Constraint: Rank(H) = 4 F H 8 Nov 2012 11755/18797 19
Finding the notes 2 T H arg min || || H M M H F H Note H T H != I Only HH T = I Could also be rewritten as T T arg min ( ) H trace M I H H M H H T T arg min ( ) H trace M M I H H H T T arg min ( )( ) H M I H H trace Correlatio n H T T T T arg max ( ) H trace Correlatio n M H H H 8 Nov 2012 11755/18797 20
Finding the notes Constraint: every row of H has length 1 T T T arg max ( ) H trace Correlatio n M H H trace H H H Differentiating and equating to 0 H T ( ( M ) M ) Correlatio Correlatio n n H H H Simply requiring the rows of H to be orthonormal p y q g gives us that H is the set of Eigenvectors of the data in M T 8 Nov 2012 11755/18797 21
Equivalences T T T arg max ( ) H trace Correlatio n M H H trace H H H is identical to 2 2 T , arg min || || || || W H M W H h h h , F i i ij i j W H i i j Minimize least squares error with the constraint that the rows of H are length 1 and orthogonal to one another 8 Nov 2012 11755/18797 22
So how does that work? There are 12 notes in the segment, hence we try to estimate 12 notes to estimate 12 notes.. 8 Nov 2012 11755/18797 23
So how does that work? The first three “notes” and their contributions The spectrograms of the notes are statistically uncorrelated The spectrograms of the notes are statistically uncorrelated 8 Nov 2012 11755/18797 24
Finding the notes Can find W instead of H 2 2 T T arg min i || || || || W W M M W W W W M M F W Solving the above with the constraints that the Solving the above, with the constraints that the columns of W are orthonormal gives you the eigen vectors of the data in M eigen vectors of the data in M T T arg max ( ) W W W M W W trace Correlatio n trace W ( ) Correlatio n M W W 8 Nov 2012 11755/18797 25
So how does that work? There are 12 notes in the segment, hence we try to estimate 12 notes.. 8 Nov 2012 11755/18797 26
Our notes are not orthogonal Overlapping frequencies O l i f i Note occur concurrently Harmonica continues to resonate to previous note More generally, simple orthogonality will not give us the desired solution 8 Nov 2012 11755/18797 28
What else can we look for? Assume: The “transcription” of one note does not p depend on what else is playing Or, in a multi ‐ instrument piece, instruments are playing independently of one another Not strictly true, but still.. 8 Nov 2012 11755/18797 29
Formulating it with Independence 2 , arg min || || ( . . . . ) W H M W H rows of H are independen t , F W H Impose statistical independence constraints on Impose statistical independence constraints on decomposition 8 Nov 2012 11755/18797 30
Recommend
More recommend