This special issue of Machine Learning focuses on the evaluation and selection of biases. The papers in this issue describe methods by which intelligent systems automatically evaluate and select their own biases, and tools for analyzing and testing various approaches to bias selection. In this paper, we motivate the importance of this topic. Since most readers will be familiar with supervised concept learning, we phrase our discussion within that framework. However, bias as we present it here is a part of every type of learning.
We outline a framework for treating bias selection as a process of designing appropriate search methods over the bias and meta-bias spaces. This framework has two essential features: it divides bias into representational and procedural components, and it characterizes learning as search within multiple tiers. The sources of bias within a system can thus be identified and analyzed with respect to their influence on this multi-tiered search process, and bias shift becomes search at the bias level. The framework provides an analytic tool with which to compare different systems (including those not developed within the framework), as well as an abstract formalism and architecture to guide the development of new systems.
We provide a summary of recent research in the field of machine learning bias, in the context of the framework presented in this paper.
Click to get a postscript version of this paper.