describes open challenges and opportunities for
advancing the state of the art in human interaction
with interactive machine-learning systems.
Developing a Common Language
Across Diverse Fields
As shown by the variety of case studies presented in
this article, many fields of computer science already
employ interactive machine learning to solve different problems, such as search in information retrieval,
filtering in recommender systems, and task learning
in human-robot interaction. However, different fields
often refer to interactive machine learning or parts of
the interactive machine-learning process in domain-specific terms (for example, relevance feedback, programming by demonstration, debugging machine-learned programs, socially guided machine learning).
This diversity in terminology impedes awareness of
progress in this common space and can potentially
lead to duplicate work. Seeking to develop a common
language and facilitate the development of new
interactive machine-learning systems, some
researchers have begun to examine this body of work
and abstract away domain-specific details from existing solutions to characterize common variables and
dimensions of the interactive machine-learning
process itself (for example, Amershi [2012]; Porter,
Theiler, and Hush [2013]).
For example, Amershi (2012) examined interactive
machine-learning systems across several fields
(including information retrieval, context-aware com-
puting, and adaptive and intelligent systems) and
identified specific design factors influencing human
interaction with machine-learning systems (for
example, the expected duration of model use, the
focus of a person’s attention during interaction, the
source and type of data over which the machine will
learn) and design dimensions that can be varied to
address these factors (for example, the type and visi-
bility of model feedback, the granularity and direc-
tion of user control, and the timing and memory of
model input). In another example, Porter, Theiler,
and Hush (2013) break down the interactive
machine-learning process into three dimensions: task
decomposition (defining the level of coordination
and division of labor between the end user and the
machine learner), training vocabulary (defining the
type of input end users can provide the machine
learner), and the training dialogue (defining the lev-
el and frequency of interaction between the end user
and the learner). Design spaces such as these can help
to form a common language for researchers and
developers to communicate new interactive
machine-learning solutions and share ideas. Howev-
er, there are many ways to dissect and describe the
various interaction points between people and
machine learners within the interactive machine-
learning process. Therefore, an important opportuni-
ty remains for converging on and adopting a com-
mon language across these fields to help accelerate
research and development in this space.
Distilling Principles and Guidelines for
How to Design Human Interaction with
Machine Learning
In addition to developing a common language, an
opportunity remains for generalizing from existing
solutions and distilling principles and guidelines for
how we should design future human interaction
with interactive machine learning, much like we
have for designing traditional interfaces (for example, Schneiderman et al. [2009]; Moggridge and
Smith [2007]; Dix et al. [2004]; Winograd [1996];
Norman [1988]). For example, Schneiderman’s golden rules of interface design advocate for designating
the users as the controllers of the system and offering them informative feedback after each interaction.
Some principles for designing traditional interfaces can directly translate to the design of interactive machine learning interfaces — interactive
machine-learning systems inherently provide users
with feedback about their actions and, as this article
discusses, giving users more control over machine-learning systems can often improve a user’s experience. However, interactive machine-learning systems also often inherently violate many existing
interface design principles. For example, research has
shown that traditional interfaces that support understandability (that is, systems that are predictable or
clear about how they work) and actionability (that
is, systems that make it clear how a person can
accomplish his or her goals and give the person the
freedom to do so) are generally more usable than
interfaces that do not support these principles. Many
machine-learning systems violate both principles:
they are inherently difficult for users to understand
fully and they largely limit the control given to the
end user. Thus, there is an opportunity to explore
how current design principles apply to the human-computer interaction in interactive machine learning.
Some researchers have started to suggest new principles for designing end-user interaction with general artificial intelligence systems, many of which
could translate to end-user interaction with interactive machine learning (for example, Norman [1994];
Höök [2000]; Horvitz [1999]; Jameson [2009]). For
example, Norman (1994) and Höök (2000) both
identified safety and trust as key factors to consider
when designing intelligent systems, referring to the
assurance against and prevention of unwanted adaptations or actions. Others have stated that artificially intelligent and machine-learning-based systems
should manage expectations to avoid misleading or
frustrating the user during interaction (for example,
Norman [1994]; Höök [2000]; Jameson [2009]). In
Horvitz’s formative paper on mixed-initiative inter-
