Error tolerance (PAC learning) explained
In PAC learning, error tolerance refers to the ability of an algorithm to learn when the examples received have been corrupted in some way. In fact, this is a very common and important issue since in many applications it is not possible to access noise-free data. Noise can interfere with the learning process at different levels: the algorithm may receive data that have been occasionally mislabeled, or the inputs may have some false information, or the classification of the examples may have been maliciously adulterated.
Notation and the Valiant learning model
In the following, let
be our
-dimensional input space. Let
be a class of functions that we wish to use in order to learn a
-valued target function
defined over
. Let
be the distribution of the inputs over
. The goal of a learning algorithm
is to choose the best function
such that it minimizes
error(h)=Px}(h(x) ≠ f(x))
. Let us suppose we have a function
that can measure the complexity of
. Let
be an oracle that, whenever called, returns an example
and its correct label
.
When no noise corrupts the data, we can define learning in the Valiant setting:[1] [2]
Definition:We say that
is efficiently learnable using
in the
Valiant setting if there exists a learning algorithm
that has access to
and a polynomial
such that for any
and
it outputs, in a number of calls to the oracle bounded by
, a function
that satisfies with probability at least
the condition
.
In the following we will define learnability of
when data have suffered some modification.
[3] [4] [5] Classification noise
In the classification noise model[6] a noise rate
is introduced. Then, instead of
that returns always the correct label of example
, algorithm
can only call a faulty oracle
that will flip the label of
with probability
. As in the Valiant case, the goal of a learning algorithm
is to choose the best function
such that it minimizes
error(h)=Px}(h(x) ≠ f(x))
. In applications it is difficult to have access to the real value of
, but we assume we have access to its upperbound
.
[7] Note that if we allow the noise rate to be
, then learning becomes impossible in any amount of computation time, because every label conveys no information about the target function.
Definition:We say that
is efficiently learnable using
in the
classification noise model if there exists a learning algorithm
that has access to
and a polynomial
such that for any
,
and
it outputs, in a number of calls to the oracle bounded by
, a function
that satisfies with probability at least
the condition
.
Statistical query learning
Statistical Query Learning[8] is a kind of active learning problem in which the learning algorithm
can decide if to request information about the likelihood
that a function
correctly labels example
, and receives an answer accurate within a tolerance
. Formally, whenever the learning algorithm
calls the oracle
, it receives as feedback probability
, such that
Qf(x)-\alpha\leqPf(x)\leqQf(x)+\alpha
.
Definition:We say that
is efficiently learnable using
in the
statistical query learning model if there exists a learning algorithm
that has access to
and polynomials
,
, and
such that for any
the following hold:
can evaluate
in time
;
is bounded by
outputs a model
such that
, in a number of calls to the oracle bounded by
.
Note that the confidence parameter
does not appear in the definition of learning. This is because the main purpose of
is to allow the learning algorithm a small probability of failure due to an unrepresentative sample. Since now
always guarantees to meet the approximation criterion
Qf(x)-\alpha\leqPf(x)\leqQf(x)+\alpha
, the failure probability is no longer needed.
The statistical query model is strictly weaker than the PAC model: any efficiently SQ-learnable class is efficiently PAC learnable in the presence of classification noise, but there exist efficient PAC-learnable problems such as parity that are not efficiently SQ-learnable.
Malicious classification
In the malicious classification model[9] an adversary generates errors to foil the learning algorithm. This setting describes situations of error burst, which may occur when for a limited time transmission equipment malfunctions repeatedly. Formally, algorithm
calls an oracle
that returns a correctly labeled example
drawn, as usual, from distribution
over the input space with probability
, but it returns with probability
an example drawn from a distribution that is not related to
. Moreover, this maliciously chosen example may strategically selected by an adversary who has knowledge of
,
,
, or the current progress of the learning algorithm.
Definition:Given a bound
for
, we say that
is efficiently learnable using
in the malicious classification model, if there exist a learning algorithm
that has access to
and a polynomial
such that for any
,
it outputs, in a number of calls to the oracle bounded by
, a function
that satisfies with probability at least
the condition
.
Errors in the inputs: nonuniform random attribute noise
In the nonuniform random attribute noise[10] [11] model the algorithm is learning a Boolean function, a malicious oracle
may flip each
-th bit of example
independently with probability
.
This type of error can irreparably foil the algorithm, in fact the following theorem holds:
In the nonuniform random attribute noise setting, an algorithm
can output a function
such that
only if
.
Notes and References
- Valiant, L. G. (August 1985). Learning Disjunction of Conjunctions. In IJCAI (pp. 560–566).
- Valiant, Leslie G. "A theory of the learnable." Communications of the ACM 27.11 (1984): 1134–1142.
- Laird, P. D. (1988). Learning from good and bad data. Kluwer Academic Publishers.
- Kearns, Michael. "Efficient noise-tolerant learning from statistical queries ." Journal of the ACM 45.6 (1998): 983–1006.
- Brunk, Clifford A., and Michael J. Pazzani. "An investigation of noise-tolerant relational concept learning algorithms." Proceedings of the 8th International Workshop on Machine Learning. 1991.
- Kearns, M. J., & Vazirani, U. V. (1994). An introduction to computational learning theory, chapter 5. MIT press.
- Angluin, D., & Laird, P. (1988). Learning from noisy examples. Machine Learning, 2(4), 343–370.
- Kearns, M. (1998). [www.cis.upenn.edu/~mkearns/papers/sq-journal.pdf Efficient noise-tolerant learning from statistical queries]. Journal of the ACM, 45(6), 983–1006.
- Kearns, M., & Li, M. (1993). [www.cis.upenn.edu/~mkearns/papers/malicious.pdf Learning in the presence of malicious errors]. SIAM Journal on Computing, 22(4), 807–837.
- [Sally Goldman|Goldman, S. A.]
- Sloan, R. H. (1989). Computational learning theory: New models and algorithms (Doctoral dissertation, Massachusetts Institute of Technology).