Deep Belief Networks are Compact Universal Approximators

Transcription

1 Deep Belief Networks are Compact Universal Approximators Franck Olivier Ndjakou Njeunje Applied Mathematics and Scientific Computation May 16, / 29

2 Outline 1 Introduction 2 Preliminaries Universal approximation theorem Definition Deep Belief Network 3 Sutekever and Hinton s Method Transfer of Probability 4 Le Roux and Bengio s Method Gray Codes Simultaneous Transfer of Probability 5 Conclusion 2 / 29

3 Introduction Machine learning (by Arthur Samuel ) Neural Networks: Set of transformations 3 / 29

4 Introduction Machine learning (by Arthur Samuel ) Neural Networks: Set of transformations 3 / 29

5 Introduction Machine learning (by Arthur Samuel ) Neural Networks: Set of transformations σ(x) = exp( x) (1) 3 / 29

6 Universal approximation theorem [Cybe89] Let ϕ( ) be a non constant, bounded, and monotonically-increasing continuous function. Let I m denote the m-dimensional unit hypercube [0, 1] m. The space of continuous functions on I m is denoted by C(I m ). Then, given any function f C(I m ) and ε > 0, there exists an integer N, real constants v i, b i R and real vectors w i R m, where i = 1,, N such that we may define: F (x) = N i=1 ( ) v i ϕ wi T x + b i (2) as an approximate realization of the function f where f is independent of ϕ; that is, F (x) f (x) < ε (3) for all x I m. In other words, functions of the form F (x) are dense in C(I m ). 4 / 29

7 Definition Generative probabilistic model (GPM) Application: recognition, classification, and generation 5 / 29

8 Definition Generative probabilistic model (GPM) Application: recognition, classification, and generation Restricted Boltzmann machine (RBM) [LeBe08] 2-layer GPM 5 / 29

9 Definition Generative probabilistic model (GPM) Application: recognition, classification, and generation Restricted Boltzmann machine (RBM) [LeBe08] 2-layer GPM Deep belief network (DBN) [HiOT06] Multilayer GPM First two layer form an RBM 6 / 29

10 Deep Belief Network Let h i represent the vector of hidden variable at layer i. The model is parametrized as follows: P(h 0, h 1, h 2,..., h l ) = P(h 0 h 1 )P(h 1 h 2 )... P(h l 2 h l 1 )P(h l, h l 1 ). (4) The hidden layer h i is a binary random vector with elements h i j and n i P(h i h i+1 ) = P(h i j h i+1 ). (5) The element h i j is a stochatic neuron or whose binary activation is 1: j=1 ( n i+1 P(h i j = 1 h i+1 ) = σ bj i + k=1 W i jk hi+1 k ). (6) 7 / 29

11 Universal approximator (DBN) Let p be an arbitrary distribution over binary vectors of n bits. A deep belief network that has p as its marginal distribution over h 0 is said to be a universal approximator. This means that for any binary vector x of n bits, there exist weights and biases such that given ε > 0: P(h 0 = x) p (x) < ε (7) 8 / 29

12 Sutekever and Hinton s Method 2008 [SuHi08] Define an arbitrary sequence (a i ) 1 i 2 n of binary vectors in {0, 1} n. The goal is to find appropriate weights and biases such that the marginal distribution over the set of outputs to our DBN is the same as the probability distribution over the vectors (a i ) 1 i 2 n. In the next few slides I will consider the following example for n = 4: a 1 = 1011, p (a 1 ) = 0.1 (8) a 2 = 1000, p (a 2 ) = 0.05 (9) a 3 = 1001, p (a 3 ) = 0.01 (10) a 4 = 1111, p (a 4 ) = 0.02 (11). (12) 9 / 29

13 Sutekever and Hinton s Method 2008 [SuHi08] Consider two consecutive layers h and v of size n, with W ij the weight linking unit v i to unit h j, b i the bias of unit v i and w a positive scalar. For every positive scalar ε (0 < ε < 1), there is a weight vector W i,: and a real b i such that P(v i = h i h) = 1 ε. Indeed, setting: W ii = 2w W ij = 0 for i j b i = w yields a total input to unit v i of: I (v i, h) = 2wh i w (13) Therefore, if w = σ 1 (1 ε), we have P(v i = h i h) = 1 ε. 10 / 29

14 Sutekever and Hinton s Method 2008 [SuHi08] W ii = 2w W ij = 0 for i j b i = w With I (v i, h) = 2wh i w and w = σ 1 (1 ε): P(v i = 1 h i = 1) = σ(w) = 1 ε. (14) P(v i = 0 h i = 0) = 1 P(v i = 1 h i = 0) (15) = 1 σ( w) (16) = σ(w) = 1 ε. (17) 11 / 29

15 Transfer of Probability This is what Sutskever and Hinton call transfer of probability. 12 / 29

16 Transfer of Probability 13 / 29

17 Transfer of Probability Number of parameters: We need 3(n + 1) 2 2 n parameters or 3 2 n layers. 13 / 29

18 Le Roux and Bengio s Method 2010 [LeBe10]: Gray Codes Gray codes [Gray53] are sequences (a i ) 1 i 2 n k {a k } = {0, 1} n such that: k s.t. 2 k 2 n, a k a k 1 H = 1 where H is the Hamming distance Example for n = 4: a 1 = 0000 a 5 = a 2 = 0001 a 6 = a 3 = 0011 a 7 = a 4 = / 29

19 Theorem 1 Let a t be an arbitrary binary vector in {0, 1} n with its last bit equal to 0 and p a scalar. For every positive scalar ε (0 < ε < 1), there is a weight vector W n,: and a real b n such that: if the binary vector h is not equal to a t, the last bit remains unchanged with probability greater than or equal to 1 ε, that is P(v n = h n h a t ) > (1 ε). if the binary vector h is equal to a t, its last bit is switched from 0 to 1 with probability σ(p). 15 / 29

20 Parameters for Theorem 1 With the following the weights and biases the result in Theorem 1 is achievable: W nj = w, 1 j k W nj = w, k + 1 j n 1 W nn = nw b n = kw + p Number of parameters: We need n 2 2 n parameters or 2 n layers. 16 / 29

21 Simultaneous Transfer of Probability 1 vector at a time. N vectors at a time. We need n2 n parameters or 2n n layers. 17 / 29

22 Arrangement of Gray Codes Let n = 2 t. There exist n sequences of vectors of n bits S i, 0 i n 1 composed of vectors S i,k, 1 k 2n satisfying the following conditions: n 1 {S 0,..., S n 1 } is a partition of the set of all vectors of n bits. Example for n = 4: S 0 S 1 S 2 S / 29

23 Arrangement of Gray Codes Let n = 2 t. There exist n sequences of vectors of n bits S i, 0 i n 1 composed of vectors S i,k, 1 k 2n satisfying the following conditions: n 1 {S 0,..., S n 1 } is a partition of the set of all vectors of n bits. 2 Every sub-sequence S i satisfies the second property of Gray codes: The Hamming distance between S i,k and S i,k+1 is 1. Example for n = 4: S 0 S 1 S 2 S / 29

24 Arrangement of Gray Codes Let n = 2 t. There exist n sequences of vectors of n bits S i, 0 i n 1 composed of vectors S i,k, 1 k 2n satisfying the following conditions: n 1 {S 0,..., S n 1 } is a partition of the set of all vectors of n bits. 2 Every sub-sequence S i satisfies the second property of Gray codes: The Hamming distance between S i,k and S i,k+1 is 1. 3 For any two sub-sequences S i and S j the bit switched between consecutive vectors (S i,k and S i,k+1 or S j,k and S j,k+1 ) is different unless the Hamming distance between S i,k and S j,k is 1. Example for n = 4: S 0 S 1 S 2 S / 29

25 End Game Can we retain the universal approximation property of DBN by transferring probability to n vectors at a time? For any binary vector x of length n, can we still find weights and biases such that P(h 0 = x) = p (x)? N vectors at a time. 21 / 29

26 Lemma Let p be an arbitrary distribution over vectors of n bits, where n is again a power of two. A DBN with 2n n + 1 layers such that: 1 for each i, 0 i n 1, the top RBM between layers h 2n n and h 2n n 1 assigns probability k p (S i,k ) to S i,1. 22 / 29

27 Lemma Let p be an arbitrary distribution over vectors of n bits, where n is again a power of two. A DBN with 2n n + 1 layers such that: 1 for each i, 0 i n 1, the top RBM between layers h 2n n and h 2n n 1 assigns probability k p (S i,k ) to S i,1. 2 for each i, 0 i n 1 and each k, 1 k 2n n 1, we have P(h 2n n (k+1) = S i,k+1 h 2n n (k) = S i,k ) = P(h 2n n (k+1) = S i,k h 2n n (k) = S i,k ) = 2n n t=k+1 p (S i,t ) (18) 2n n t=k p (S i,t ) p (S i,k ) (19) 2n n t=k p (S i,t ) 22 / 29

28 Lemma Let p be an arbitrary distribution over vectors of n bits, where n is again a power of two. A DBN with 2n n + 1 layers such that: 1 for each i, 0 i n 1, the top RBM between layers h 2n n and h 2n n 1 assigns probability k p (S i,k ) to S i,1. 2 for each i, 0 i n 1 and each k, 1 k 2n n 1, we have P(h 2n n (k+1) = S i,k+1 h 2n n (k) = S i,k ) = P(h 2n n (k+1) = S i,k h 2n n (k) = S i,k ) = 2n n t=k+1 p (S i,t ) (18) 2n n t=k p (S i,t ) p (S i,k ) (19) 2n n t=k p (S i,t ) 3 for each k, 1 k 2n n 1, we have P(h 2n n (k+1) = a h 2n n (k) = a) = 1 if a / i S i,k (20) has p as its marginal distribution over h / 29

29 Proof of the Lemma Let x be an arbitrary binary vector of n bits; there is a pair (i, k) such that x = S i,k. We need to show that: P(h 0 = S i,k ) = p (S i,k ). (21) 23 / 29

30 Proof of the Lemma Let x be an arbitrary binary vector of n bits; there is a pair (i, k) such that x = S i,k. We need to show that: P(h 0 = S i,k ) = p (S i,k ). (21) Example for n = 4: if x = S 22 P(h 0 = S 22 ) = p (S 22 ). (22) 23 / 29

31 Proof of the Lemma The marginal probability of h 0 = S i,k is therefore equal to: P(h 0 = S i,k ) = P(h 2n n 1 = S i,1 ) (23) k 1 ) P (h 2n n (t+1) = S i,t+1 h 2n n t = S i,t (24) t=1 ) P (h 2n n (k+1) = S i,k h 2n n k = S i,k (25) 2 n n 1 t=k+1 ) P (h 2n n (t+1) = S i,k h 2n n t = S i,k (26) 24 / 29

32 Proof of the Lemma By replacing each of those probabilities by the ones given in the Lemma we get: P(h 0 = S i,k ) = 2 n n u=1 k 1 t=1 p (S i,u ) (27) 2n n u=t+1 p (S i,u ) 2n n u=t p (S i,u ) p (S i,k ) 2n n u=k p (S i,u ) (28) (29) 1 2n n 1 k (30) = p (S i,k ) (31) The last result comes from the cancellation of consecutive terms in the product. This concludes the proof. 25 / 29

33 Theorem 4 If n = 2 t, a DBN composed of 2n n + 1 layers of size n is a universal approximator of distributions over vectors of size n. 26 / 29

34 Theorem 4 If n = 2 t, a DBN composed of 2n n + 1 layers of size n is a universal approximator of distributions over vectors of size n. Proof of Theorem 4: Using Lemma, we now show that it is possible to construct such a DBN. First, Le Roux and Bengio (2008) showed that an RBM with n hidden units can model any distribution which assigns a non-zero probability to at most n vectors. Property 1 of the Lemma can therefore be achieved. 26 / 29

35 Proof of Theorem 4 All the subsequent layers are as follows. At each layer, the first t bits of h k+1 are copied to the first t bits of h k with probability arbitrarily close to 1. This is possible as proven earlier. 27 / 29

36 Proof of Theorem 4 All the subsequent layers are as follows. At each layer, the first t bits of h k+1 are copied to the first t bits of h k with probability arbitrarily close to 1. This is possible as proven earlier. At each layer, n/2 of the remaining n t bits are potentially changed to move from one vector in a Gray code sequence to the next with the correct probability (as defined in the Lemma). 27 / 29

37 Proof of Theorem 4 All the subsequent layers are as follows. At each layer, the first t bits of h k+1 are copied to the first t bits of h k with probability arbitrarily close to 1. This is possible as proven earlier. At each layer, n/2 of the remaining n t bits are potentially changed to move from one vector in a Gray code sequence to the next with the correct probability (as defined in the Lemma). The remaining n/2 t bits are copied from h k+1 to h k with probability arbitrarily close to 1. Such layers are arbitrarily close to fulfilling the requirements of the second property of the Lemma. This concludes the proof. 27 / 29

38 Conclusion Deep belief networks are compact universal approximators: Sutskever and Hinton method (2008) Transfer of probability We need 3(n + 1) 2 2 n parameters or 3 2 n layers. LeRoux and Bengio improvements (2009) Gray codes Simultaneous transfer of probability We need n2 n parameters or 2n n layers (given n is a power of 2). 28 / 29

39 References: Le Roux, N., Bengio, Y. (2010). Deep belief networks are compact universal approximators. Neural computation, 22(8), Le Roux, N. and Bengio, Y. (2008). Representational power of restricted boltzmann machines and deep belief networks. Neural Computation, 20(6), Sutskever, I. and Hinton, G. E. (2008). Deep, narrow sigmoid belief networks are universal approximators. Neural Computation, 20(11), Hinton, G. E., Osindero, S., and Teh, Y. (2006). A fast learning algorithm for deep belief nets. Neural Computation, 18, Gray, F. (1953). Pulse code communication. U.S. Patent 2,632,058. Cybenko, G. (1989). Approximation by superpositions of a sigmoidal function. Mathematics of control, signals and systems, 2(4), / 29