The Relationship Between Agnostic Selective Classification, Active Learning and the Disagreement Coefficient

Authors: Roei Gelbhart, Ran El-Yaniv

JMLR 2019 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical The main result of this paper is an equivalence between the following three entities: (i) the existence of a fast rejection rate for any PCS learning algorithm (such as ILESS); (ii) a poly-logarithmic bound for Hanneke s disagreement coefficient; and (iii) an exponential speedup for a new disagreement-based active learner called Active-ILESS. We prove that Active-ILESS, despite being very similar to Oracular CAL Hsu (2010), exhibits an improved label complexity, in comparison to that proved for Oracular CAL. Our second and main contribution is to show a complete equivalence between (i) selective classification with a fast R rejection rate, (ii) an AL algorithm, Active-ILESS, with an R exponential speedup, and (iii) the existence of an f with a disagreement coefficient bounded by polylog(1/r). This is illustrated in Figure 1, where the blue errors indicate the equivalence relationships we prove in this paper, and the red arrow indicates a previously known result (Hsu, 2010; Hanneke, 2007), and can also be deduced from the other arrows.
Researcher Affiliation Academia Roei Gelbhart EMAIL Department of Computer Science Technion Israel Institute of Technology Ran El-Yaniv EMAIL Department of Computer Science Technion Israel Institute of Technology
Pseudocode Yes Strategy 1 Agnostic Low-Error Selective Strategy (LESS) Input: Sample set of size m, Sm, Confidence level δ Hypothesis class F with VC dimension d Output: A selective classifier (h, g) ... Strategy 2 Improved Low-Error Selective Strategy (ILESS) Input: Sample set of size m, Sm, Confidence level δ Hypothesis class F with VC dimension d Output: A selective classifier (h, g) ... Strategy 3 Agnostic Low-Error Active Learning Strategy (Active-ILESS) Input: ϵ and/or m depending on the desired termination condition (error or labeling budget, respectively) Confidence level δ Hypothesis class F with VC dimension d An unlabeled input sequence sampled i.i.d from PX,Y: x1, x2, x3, . . . Output: A classifier ˆf ... Strategy 4 Batch Improved Low-Error Selective Strategy (Batch-ILESS) Input: Sample set of size m, Sm, Confidence level δ Hypothesis class F with VC dimension d Output: A selective classifier (h, g)
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets No A learning problem is specified via a hypothesis class F and an unknown probability distribution PX,Y. The paper deals with theoretical analysis of algorithms under abstract distributions and hypothesis classes, not specific datasets.
Dataset Splits No The paper does not mention any specific datasets used for empirical evaluation, and therefore does not provide dataset split information.
Hardware Specification No The paper is theoretical and does not describe experimental results or the hardware used to obtain them.
Software Dependencies No The paper is theoretical and does not mention specific software dependencies or versions.
Experiment Setup No The paper is theoretical and does not provide details on experimental setup, hyperparameters, or training configurations.