A Note on Learning from Multiple-Instance Examples

We describe a simple reduction from the problem of PAC-learning from multiple-instance examples to that of PAC-learning with one-sided random classification noise. Thus, all concept classes learnable with one-sided noise, which includes all concepts learnable in the usual 2-sided random noise model plus others such as the parity function, are learnable from multiple-instance examples. We also describe a more efficient (and somewhat technically more involved) reduction to the Statistical-Query model that results in a polynomial-time algorithm for learning axis-parallel rectangles with sample complexity Õ(d²r/²) , saving roughly a factor of r over the results of Auer et al. (1997).     

Active Learning Using Arbitrary Binary Valued Queries

The original and most widely studied PAC model for learning assumes a passive learner in the sense that the learner plays no role in obtaining information about the unknown concept. That is, the samples are simply drawn independently from some probability distribution. Some work has been done on studying more powerful oracles and how they affect learnability. To find bounds on the improvement in sample complexity that can be expected from using oracles, we consider active learning in the sense that the learner has complete control over the information received. Specifically, we allow the learner to ask arbitrary yes/no questions. We consider both active learning under a fixed distribution and distribution-free active learning. In the case of active learning, the underlying probability distribution is used only to measure distance between concepts. For learnability with respect to a fixed distribution, active learning does not enlarge the set of learnable concept classes, but can improve the sample complexity. For distribution-free learning, it is shown that a concept class is actively learnable iff it is finite, so that active learning is in fact less powerful than the usual passive learning model. We also consider a form of distribution-free learning in which the learner knows the distribution being used, so that distribution-free refers only to the requirement that a bound on the number of queries can be obtained uniformly over all distributions. Even with the side information of the distribution being used, a concept class is actively learnable iff it has finite VC dimension, so that active learning with the side information still does not enlarge the set of learnable concept classes.     

Quantum lower bounds for the Goldreich-Levin problem

At the heart of the Goldreich-Levin theorem is the problem of determining an n-bit string a by making queries to two oracles, referred to as IP (inner product) and EQ (equivalence). The IP oracle, on input x, returns a bit that is biased towards a⋅x (the modulo two inner product of a with x) in the following sense. For a random x, the probability that IP(x)=a⋅x is at least . The EQ oracle, on input x, returns a bit specifying whether or not x=a. It has been shown that a quantum algorithm can solve this problem with O(1/?) IP and EQ queries, whereas any classical algorithm requires Ω(n/?²) such queries. Also, the quantum algorithm requires only O(n/?) auxiliary one- and two-qubit gates in addition to its queries. We show that the above quantum algorithm is optimal in terms of both EQ and IP queries. Specifically, Ω(1/?) EQ queries are necessary, and Ω(1/?) IP queries are necessary if the number of EQ queries is .     

Malicious Omissions and Errors in Answers to Membership Queries

We consider two issues in polynomial-time exact learning of concepts using membership and equivalence queries: (1) errors or omissions in answers to membership queries, and (2) learning finite variants of concepts drawn from a learnable class.To study (1), we introduce two new kinds of membership queries: limited membership queries and malicious membership queries. Each is allowed to give incorrect responses on a maliciously chosen set of strings in the domain. Instead of answering correctly about a string, a limited membership query may give a special I don't know answer, while a malicious membership query may give the wrong answer. A new parameter Lis used to bound the length of an encoding of the set of strings that receive such incorrect answers. Equivalence queries are answered correctly, and learning algorithms are allowed time polynomial in the usual parameters and L. Any class of concepts learnable in polynomial time using equivalence and malicious membership queries is learnable in polynomial time using equivalence and limited membership queries; the converse is an open problem. For the classes of monotone monomials and monotone k-term DNF formulas, we present polynomial-time learning algorithms using limited membership queries alone. We present polynomial-time learning algorithms for the class of monotone DNF formulas using equivalence and limited membership queries, and using equivalence and malicious membership queries.To study (2), we consider classes of concepts that are polynomially closed under finite exceptions and a natural operation to add exception tables to a class of concepts. Applying this operation, we obtain the class of monotone DNF formulas with finite exceptions. We give a polynomial-time algorithm to learn the class of monotone DNF formulas with finite exceptions using equivalence and membership queries. We also give a general transformation showing that any class of concepts that is polynomially closed under finite exceptions and is learnable in polynomial time using standard membership and equivalence queries is also polynomial-time learnable using malicious membership and equivalence queries. Corollaries include the polynomial-time learnability of the following classes using malicious membership and equivalence queries: deterministic finite acceptors, boolean decision trees, and monotone DNF formulas with finite exceptions.     

Equivalence of models for polynomial learnability

In this paper we consider several variants of Valiant's learnability model that have appeared in the literature. We give conditions under which these models are equivalent in terms of the polynomially learnable concept classes they define. These equivalences allow comparisons of most of the existing theorems in Valiant-style learnability and show that several simplifying assumptions on polynomial learning algorithms can be made without loss of generality. We also give a useful reduction of learning problems to the problem of finding consistent hypotheses, and give comparisons and equivalences between Valiant's model and the prediction learning models of Haussler, Littlestone, and Warmuth (in “29th Annual IEEE Symposium on Foundations of Computer Science,” 1988).     

How Many Missing Answers Can Be Tolerated by Query Learners?

We consider the model of exact learning using an equivalence oracle and an incomplete membership oracle. In this model a random subset of the learners membership queries is left unanswered. Our results are as follows. First, we analyze the obvious method for coping with missing answers: search exhaustively through all possible answer patterns associated with the unanswered queries. Thereafter, we present two specific concept classes that are efficiently learnable using an equivalence oracle and a (completely reliable) membership oracle, but are provably not polynomially learnable if the membership oracle becomes slightly incomplete. The first class demonstrates that the aforementioned method of exhaustively searching through all possible answer patterns cannot be substantially improved in general (despite its apparent simplicity). The second class demonstrates that the incomplete membership oracle can be rendered useless even if it leaves only a fraction 1/poly(n) of all queries unanswered. Finally, we present a learning algorithm for monotone DNF formulas that can cope with a relatively large fraction of missing answers (more than 60%), but is as efficient (in terms of run-time and number of queries) as the classical algorithm whose questions are always answered reliably.     

Learning Boolean Functions in an Infinite Attribute Space

This paper presents a theoretical model for learning Boolean functions in domains having a large, potentially infinite number of attributes. The model allows an algorithm to employ a rich vocabulary to describe the objects it encounters in the world without necessarily incurring time and space penalties so long as each individual object is relatively simple. We show that many of the basic Boolean functions learnable in standard theoretical models, such as conjunctions, disjunctions, K-CNF, and K-DNF, are still learnable in the new model, though by algorithms no longer quite so trivial as before. The new model forces algorithms for such classes to act in a manner that appears more natural for many learning scenarios.     

Noise-Tolerant Occam Algorithms and Their Applications to Learning Decision Trees

In the distribution-independent model of concept learning of Valiant, Angluin and Laird have introduced a formal model of noise process, called classification noise process, to study how to compensate for randomly introduced errors, or noise, in classifying the example data. In this article, we investigate the problem of designing efficient learning algorithms in the presence of classification noise. First, we develop a technique of building efficient robust learning algorithms, called noise-tolerant Occam algorithms, and show that using them, one can construct a polynomial-time algorithm for learning a class of Boolean functions in the presence of classification noise. Next, as an instance of such problems of learning in the presence of classification noise, we focus on the learning problem of Boolean functions represented by decision trees. We present a noise-tolerant Occam algorithm for k-DL (the class of decision lists with conjunctive clauses of size at most k at each decision introduced by Rivest) and hence conclude that k-DL is polynomially learnable in the presence of classification noise. Further, we extend the noise-tolerant Occam algorithm for k-DL to one for r-DT (the class of decision trees of rank at most r introduced by Ehrenfeucht and Haussler) and conclude that r-DT is polynomially learnable in the presence of classification noise.     

Randomly Fallible Teachers: Learning Monotone DNF with an Incomplete Membership Oracle

We introduce a new fault-tolerant model of algorithmic learning using an equivalence oracle and anincomplete membership oracle, in which the answers to a random subset of the learner's membership queries may be missing. We demonstrate that, with high probability, it is still possible to learn monotone DNF formulas in polynomial time, provided that the fraction of missing answers is bounded by some constant less than one. Even when half the membership queries are expected to yield no information, our algorithm will exactly identifym-term,n-variable monotone DNF formulas with an expectedO(mn ²) queries. The same task has been shown to require exponential time using equivalence queries alone. We extend the algorithm to handle some one-sided errors, and discuss several other possible error models. It is hoped that this work may lead to a better understanding of the power of membership queries and the effects of faulty teachers on query models of concept learning.     

Simple learning algorithms using divide and conquer

This paper investigates what happens when a learning algorithm for a classC attempts to learn target formulas from a different class. In many cases, the learning algorithm will find a bad attribute or a property of the target formula which precludes its membership in the classC. To continue the learning process, we proceed by building a decision tree according to the possible values of this attribute (divide) and recursively run the learning algorithm for each value (conquer). This paper shows how to recursively run the learning algorithm for each value using the oracles of the target.We demonstrate that the application of this idea on some known learning algorithms can both simplify the algorithm and provide additional power to learn more classes. In particular, we give a simple exact learning algorithm, using membership and equivalence queries, for the class of DNF that is almost unate, that is, unate with the addition ofO (logn) nonunate variables and a constant number of terms. We also find algorithms in different models for boolean functions that depend onk terms.     

Conjunctions of Unate DNF Formulas: Learning and Structure

A central topic in query learning is to determine which classes of Boolean formulas are efficiently learnable with membership and equivalence queries. We consider the class ^kconsisting of conjunctions ofkunate DNF formulas. This class generalizes the class ofk-clause CNF formulas and the class of unate DNF formulas, both of which are known to be learnable in polynomial time with membership and equivalence queries. We prove that ²can be properly learned with a polynomial number of polynomial-size membership and equivalence queries, but can be properly learned in polynomial time with such queries if and only if P=NP. Thus the barrier to properly learning ²with membership and equivalence queries is computational rather than informational. Few results of this type are known. In our proofs, we use recent results of Hellersteinet al.(1997,J. Assoc. Comput. Mach.43(5), 840–862), characterizing the classes that are polynomial-query learnable, together with work of Bshouty on the monotone dimension of Boolean functions. We extend some of our results to ^kand pose open questions on learning DNF formulas of small monotone dimension. We also prove structural results for ^k. We construct, for any fixedk2, a class of functionsfthat cannot be represented by any formula in ^k, but which cannot be “easily” shown to have this property. More precisely, for any functionfonnvariables in the class, the value offon any polynomial-size set of points in its domain is not a witness thatfcannot be represented by a formula in ^k. Our construction is based on BCH codes.     

On Polynomial-Time Learnability in the Limit of Strictly Deterministic Automata

This paper deals with the polynomial-time learnability of a language class in the limit from positive data, and discusses the learning problem of a subclass of deterministic finite automata (DFAs), called strictly deterministic automata (SDAs), in the framework of learning in the limit from positive data. We first discuss the difficulty of Pitt's definition in the framework of learning in the limit from positive data, by showing that any class of languages with an infinite descending chain property is not polynomial-time learnable in the limit from positive data. We then propose new definitions for polynomial-time learnability in the limit from positive data. We show in our new definitions that the class of SDAs is iteratively, consistently polynomial-time learnable in the limit from positive data. In particular, we present a learning algorithm that learns any SDA M in the limit from positive data, satisfying the properties that (i) the time for updating a conjecture is at most O(lm), (ii) the number of implicit prediction errors is at most O(ln), where l is the maximum length of all positive data provided, m is the alphabet size of M and n is the size of M, (iii) each conjecture is computed from only the previous conjecture and the current example, and (iv) at any stage the conjecture is consistent with the sample set seen so far. This is in marked contrast to the fact that the class of DFAs is neither learnable in the limit from positive data nor polynomial-time learnable in the limit.     

On the non-efficient PAC learnability of conjunctive queries

This note serves three purposes: (i) we provide a self-contained exposition of the fact that conjunctive queries are not efficiently learnable in the Probably-Approximately-Correct (PAC) model, paying clear attention to the complicating fact that this concept class lacks the polynomial-size fitting property, a property that is tacitly assumed in much of the computational learning theory literature; (ii) we establish a strong negative PAC learnability result that applies to many restricted classes of conjunctive queries (CQs), including acyclic CQs for a wide range of notions of acyclicity; (iii) we show that CQs (and UCQs) are efficiently PAC learnable with membership queries.     

Compact Navigation and Distance Oracles for Graphs with Small Treewidth

Given an unlabeled, unweighted, and undirected graph with n vertices and small (but not necessarily constant) treewidth k, we consider the problem of preprocessing the graph to build space-efficient encodings (oracles) to perform various queries efficiently. We assume the word RAM model where the size of a word is Ω(logn) bits. The first oracle, we present, is the navigation oracle which facilitates primitive navigation operations of adjacency, neighborhood, and degree queries. By way of an enumeration argument, which is of interest in its own right, we show the space requirement of the oracle is optimal to within lower order terms for all graphs with n vertices and treewidth k. The oracle supports the mentioned queries all in constant worst-case time. The second oracle, we present, is an exact distance oracle which facilitates distance queries between any pair of vertices (i.e., an all-pairs shortest-path oracle). The space requirement of the oracle is also optimal to within lower order terms. Moreover, the distance queries perform in O(k ³log³ k) time. Particularly, for the class of graphs of popular interest, graphs of bounded treewidth (where k is constant), the distances are reported in constant worst-case time.     

Dynamic learning from multiple examples for semantic object segmentation and search

We present a novel “dynamic learning” approach for an intelligent image database system to automatically improve object segmentation and labeling without user intervention, as new examples become available, for object-based indexing. The proposed approach is an extension of our earlier work on “learning by example,” which addressed labeling of similar objects in a set of database images based on a single example. The proposed dynamic learning procedure utilizes multiple example object templates to improve the accuracy of existing object segmentations and labels. Multiple example templates may be images of the same object from different viewing angles, or images of related objects. This paper also introduces a new shape similarity metric called normalized area of symmetric differences (NASD), which has desired properties for use in the proposed “dynamic learning” scheme, and is more robust against boundary noise that results from automatic image segmentation. Performance of the dynamic learning procedures has been demonstrated by experimental results.     

Uniform-distribution attribute noise learnability

We study the problem of PAC-learning Boolean functions with random attribute noise under the uniform distribution. We define a noisy distance measure for function classes and show that if this measure is small for a class and an attribute noise distribution D then is not learnable with respect to the uniform distribution in the presence of noise generated according to D. The noisy distance measure is then characterized in terms of Fourier properties of the function class. We use this characterization to show that the class of all parity functions is not learnable for any but very concentrated noise distributions D. On the other hand, we show that if is learnable with respect to uniform using a standard Fourier-based learning technique, then is learnable with time and sample complexity also determined by the noisy distance. In fact, we show that this style algorithm is nearly the best possible for learning in the presence of attribute noise. As an application of our results, we show how to extend such an algorithm for learning AC⁰ so that it handles certain types of attribute noise with relatively little impact on the running time.     

On Restricted-Focus-of-Attention Learnability of Boolean Functions

In the k-Restricted-Focus-of-Attention (k-RFA) model, only k of the n attributes of each example are revealed to the learner, although the set of visible attributes in each example is determined by the learner. While thek -RFA model is a natural extension of the PAC model, there are also significant differences. For example, it was previously known that learnability in this model is not characterized by the VC-dimension and that many PAC learning algorithms are not applicable in the k-RFA setting.In this paper we further explore the relationship between the PAC and k -RFA models, with several interesting results. First, we develop an information-theoretic characterization of k-RFA learnability upon which we build a general tool for proving hardness results. We then apply this and other new techniques for studying RFA learning to two particularly expressive function classes,k -decision-lists (k-DL) and k-TOP, the class of thresholds of parity functions in which each parity function takes at most k inputs. Among other results, we prove a hardness result for k-RFA learnability of k-DL,k n-2 . In sharp contrast, an (n-1)-RFA algorithm for learning (n-1)-DL is presented. Similarly, we prove that 1-DL is learnable if and only if at least half of the inputs are visible in each instance. In addition, we show that there is a uniform-distribution k-RFA learning algorithm for the class of k -DL. For k-TOP we show weak learnability by ak -RFA algorithm (with efficient time and sample complexity for constant k) and strong uniform-distribution k-RFA learnability of k-TOP with efficient sample complexity for constant k. Finally, by combining some of our k-DL and k-TOP results, we show that, unlike the PAC model, weak learning does not imply strong learning in the k -RFA model.     

Complexity theoretic hardness results for query learning

We investigate the complexity of learning for the well-studied model in which the learning algorithm may ask membership and equivalence queries. While complexity theoretic techniques have previously been used to prove hardness results in various learning models, these techniques typically are not strong enough to use when a learning algorithm may make membership queries. We develop a general technique for proving hardness results for learning with membership and equivalence queries (and for more general query models). We apply the technique to show that, assuming , no polynomial-time membership and (proper) equivalence query algorithms exist for exactly learning read-thrice DNF formulas, unions of halfspaces over the Boolean domain, or some other related classes. Our hardness results are representation dependent, and do not preclude the existence of representation independent algorithms.?The general technique introduces the representation problem for a class F of representations (e.g., formulas), which is naturally associated with the learning problem for F. This problem is related to the structural question of how to characterize functions representable by formulas in F, and is a generalization of standard complexity problems such as Satisfiability. While in general the representation problem is in , we present a theorem demonstrating that for "reasonable" classes F, the existence of a polynomial-time membership and equivalence query algorithm for exactly learning F implies that the representation problem for F is in fact in co-NP. The theorem is applied to prove hardness results such as the ones mentioned above, by showing that the representation problem for specific classes of formulas is NP-hard. Received: December 6, 1994