Sample-efficient strategies for learning in the presence of noise

In this paper, we prove various results about PAC learning in the presence of malicious noise. Our main interest is the sample size behavior of learning algorithms. We prove the first nontrivial sample complexity lower bound in this model by showing that order of ε/Δ<supscrpt>2</supscrpt> + <italic>d</italic>/Δ (up to logarithmic factors) examples are necessary for PAC learning any target class of {0,1}-valued functions of VC dimension <italic>d</italic>, where ε is the desired accuracy and &eegr; = ε/(1 + ε) - Δ the malicious noise rate (it is well known that any nontrivial target class cannot be PAC learned with accuracy ε and malicious noise rate &eegr; ≥ ε/(1 + ε), this irrespective to sample complexity). We also show that this result cannot be significantly improved in general by presenting efficient learning algorithms for the class of all subsets of <italic>d</italic> elements and the class of unions of at most <italic>d</italic> intervals on the real line. This is especialy interesting as we can also show that the popular minimum disagreement strategy needs samples of size <italic>d</italic> ε/Δ<supscrpt>2</supscrpt>, hence is not optimal with respect to sample size. We then discuss the use of randomized hypotheses. For these the bound ε/(1 + ε) on the noise rate is no longer true and is replaced by 2ε/(1 + 2ε). In fact, we present a generic algorithm using randomized hypotheses that can tolerate noise rates slightly larger than ε/(1 + ε) while using samples of size <italic>d</italic>/ε as in the noise-free case. Again one observes a quadratic powerlaw (in this case <italic>d</italic>ε/Δ<supscrpt>2</supscrpt>, Δ = 2ε/(1 + 2ε) - &eegr;) as Δ goes to zero. We show upper and lower bounds of this order.

[1]  R. R. Bahadur Some Approximations to the Binomial Distribution Function , 1960 .

[2]  R. R. Bahadur,et al.  On Deviations of the Sample Mean , 1960 .

[3]  Michel Loève,et al.  Probability Theory I , 1977 .

[4]  S. M. Samuels,et al.  Monotone Convergence of Binomial Probabilities and a Generalization of Ramanujan's Equation , 1968 .

[5]  J. Littlewood On the probability in the tail of a binomial distribution , 1969, Advances in Applied Probability.

[6]  Norbert Sauer,et al.  On the Density of Families of Sets , 1972, J. Comb. Theory, Ser. A.

[7]  S. Shelah A combinatorial problem; stability and order for models and theories in infinitary languages. , 1972 .

[8]  Vladimir Vapnik,et al.  Estimation of Dependences Based on Empirical Data: Springer Series in Statistics (Springer Series in Statistics) , 1982 .

[9]  Leslie G. Valiant,et al.  A theory of the learnable , 1984, STOC '84.

[10]  Ming Li,et al.  Learning in the presence of malicious errors , 1993, STOC '88.

[11]  P. Laird Learning from Good and Bad Data , 1988 .

[12]  Leslie G. Valiant,et al.  A general lower bound on the number of examples needed for learning , 1988, COLT '88.

[13]  David Haussler,et al.  Learnability and the Vapnik-Chervonenkis dimension , 1989, JACM.

[14]  R. Schapire Toward Eecient Agnostic Learning , 1992 .

[15]  Linda Sellie,et al.  Toward efficient agnostic learning , 1992, COLT '92.

[16]  Robert E. Schapire,et al.  Efficient Distribution-Free Learning of Probabilistic , 1994 .

[17]  R. Schapire,et al.  Toward Efficient Agnostic Learning , 1994 .

[18]  Hans Ulrich Simon,et al.  General bounds on the number of examples needed for learning probabilistic concepts , 1993, COLT '93.

[19]  Claudio Gentile,et al.  Improved lower bounds for learning from noisy examples: an information-theoretic approach , 1998, COLT' 98.