Error bounds for approximations with deep ReLU networks.
Citations
467 citations
405 citations
354 citations
345 citations
Cites background or methods or result from "Error bounds for approximations wit..."
...If we take r = d+1 2 + 2 as in our previous discussion, we see that for d ≥ 6, the deep net constructed in Theorem 1 of [29] for achieving an accuracy ∈ (0, 1) for approximating f ∈ C([0, 1]) has at least 2 −d/r free parameters and at least C0d 4 (log(1/ ) + d) layers....
[...]
...To compare this result with ours when d is large, we need to derive an explicit lower bound for the number of parameters in the above net from the analysis in [29] which is based on Taylor polynomials of f and trapezoid functions defined by σ....
[...]
...When the activation function is ReLU, explicit rates of approximation by fully connected neural networks were obtained recently in [13] for shallow nets, in [24] for nets with 3 hidden layers, and in [29,2,22] for nets with more layers....
[...]
...Thus, to achieve an accuracy ∈ (0, 1) for approximating f by a ReLU deep net, one takes N = ( 2d+1dr r! )1/r and δ = 2d+1dr(d+r) as in [29] and know that the depth of the net is at least C0d 2 (log(1/ ) + (d + 1) log 2 + r log d + log(d + r)), and the total number of parameters for the net is more than the number of coefficients D f(m/N) α! which is...
[...]
...In particular, it was shown in Theorem 1 of [29] that for f ∈ C([0, 1]), the approximation accuracy ∈ (0, 1) can be achieved by a ReLU deep net with at most c(log(1/ ) + 1) layers and at most c −d/r(log(1/ ) + 1) weights and computation units with a constant c = c(d, r)....
[...]
307 citations
References
[...]
46,982 citations
14,635 citations
"Error bounds for approximations wit..." refers background in this paper
...Recently, multiple successful applications of deep neural networks to pattern recognition problems (Schmidhuber [2015], LeCun et al. [2015]) have revived active interest in theoretical properties of such networks, in particular their expressive power....
[...]
3,939 citations
"Error bounds for approximations wit..." refers background in this paper
...Bianchini and Scarselli 2014 give bounds for Betti numbers characterizing topological properties of functions represented by networks....
[...]
1,757 citations
"Error bounds for approximations wit..." refers background or methods in this paper
...For any d, n and ∈ (0, 1), there is a ReLU network architecture that 1. is capable of expressing any function from Fd,n with error ; 2. has the depth at most c(ln(1/ ) + 1) and at most c −d/n(ln(1/ ) + 1) weights and computation units, with some constant c = c(d, n)....
[...]
...Let us obtain a condition ensuring that such f ∈ Fd,n....
[...]
...…conclude c), observe that computation (4) consists of three instances of f̃sq,δ and finitely many linear and ReLU operations, so, using Proposition 2, we can implement ×̃ by a ReLU network such that its depth and the number of computation units and weights are O(ln(1/δ)), i.e. are O(ln(1/ ) + lnM)....
[...]
...Namely, let fm be the piece-wise linear interpolation of f with 2m + 1 uniformly distributed breakpoints k 2m , k = 0, . . . , 2m: fm ( k 2m ) = ( k 2m )2 , k = 0, . . . , 2m (see Fig....
[...]
...Namely, given f ∈ F1,1 and > 0, set T = d1 e and let f̃ be the piece-wise interpolation of f with T + 1 uniformly spaced breakpoints ( t T )Tt=0 (i.e., f̃( t T ) = f( t T ), t = 0, . . . , T )....
[...]
1,209 citations
"Error bounds for approximations wit..." refers background in this paper
...…network, a deep one can be viewed as a long sequence of non-commutative transformations, which is a natural setting for high expressiveness (cf. the well-known Solovay-Kitaev theorem on fast approximation of arbitrary quantum operations by sequences of non-commutative gates, see Kitaev et al.…...
[...]