The unprecedented rise in the adoption of artificial intelligence techniques and automation in many contexts is concomitant with shortcomings of such technology with respect to robustness, interpretability, usability, and trustworthiness. Crowd computing offers a viable means to leverage human intelligence at scale for data creation, enrichment, and interpretation, demonstrating a great potential to improve the performance of AI systems and increase the adoption of AI in general. Existing research and practice has mainly focused on leveraging crowd computing for training data creation. However, this perspective is rather limiting in terms of how AI can fully benefit from crowd computing. In this vision paper, we identify opportunities in crowd computing to propel better AI technology, and argue that to make such progress, fundamental problems need to be tackled from both computation and interaction standpoints. We discuss important research questions in both these themes, with an aim to shed light on the research needed to pave a future where humans and AI can work together seamlessly, while benefiting from each other.
Artificial intelligence techniques and machine learning in particular, are drastically changing our
lives through technological revolutions across several domains such as transportation, health, finance,
education, and manufacturing. AI systems at the forefront of such innovations have garnered a
growing barrage of concerns, not only due to issues pertaining to performance – such systems have
been observed to easily fail in situations slightly different from those encountered in the training
instances [
Problems exist and manifest both in AI systems and in the interaction between end users with
such systems. On the one hand, machine learning models have been criticized for the lack of
robustness, fairness, and transparency [
The two schools of challenges pertaining to AI systems, characterised as computational and
interactional ones, are in fact highly related to each other. From the computation perspective, a better
understanding of user interactions can help identify the focal point of system development and
potentially spark new research directions. A prominent example is machine learning interpretability,
inspired by the observation that explainable results are more in demand by users than highly accurate
ones. From the interaction perspective, more robust and interpretable systems can help build trust
and increase system uptake [
Crowd computing has been conceptualised in various ways – as being related to crowdsourcing,
human computation, social computing, cloud computing and mobile computing [
Based on the existing evidence of how crowd computing can play an important role in tackling computational and interactional challenges in developing new-age AI systems, in this vision paper, we highlight research themes that need to be pursued to ensure that AI systems can create a future where we are better off than we currently are – both as individuals and as a society. 2
In this section, we discuss important challenges that need to be addressed to make advances in the next generation of AI systems from two main standpoints – (1) Human-in-the-loop AI, and (2) Human-AI interaction. The former concerns the computational role of humans for AI, i.e., AI by humans, while the latter concerns the interactional role of humans with AI systems, i.e., AI for humans. 2.1
In what follows, we analyze the fundamental computational challenges in the quest for robust, interpretable, and hence trustworthy AI systems. We argue that to tackle such fundamental challenges, research should explore a novel crowd computing paradigm, which we refer to as “crowd conceptual computing”. In such form of crowd computing, crowd workers can contribute knowledge at the conceptual level; this comes in contrast to the current paradigm where crowd intelligence is utilised on a per-datum basis, e.g., labelling and debugging individual data instances.
Robust AI by Crowds. Machine (deep) learning models have proven to be “shallow” – they often
learn spurious correlations in the data – and “brittle” – they are unable to make sense of situations
slightly different from the training data. Consequently, current AI systems often fail when required to
make predictions on data beyond the training distribution, which is of crucial need in practice. Those
issues constitute what is now referred to as the robustness or reliability issue, generally viewed as a
main obstacle for wide deployment of AI systems [
Robust AI requires models to be encapsulated with causality and better generalisation ability, which are the main advantageous characteristics of conventional symbolic AI methods focusing on knowledge representation and reasoning. Recent discussions in the AI community has therefore converged to the idea of developing neurosymbolic methods that benefit from both the robustness of symbolic methods and the flexibility of deep learning. Few discussions have, however, touched upon the questions of what knowledge is required, and where and how to obtain such knowledge. Historical research in expert systems has shown that the amount of knowledge for a specific task can be very large that can easily go beyond readily available knowledge bases and what individuals can provide. Building on top of the Web, crowd computing systems can reach an unprecedented number of people, thus offering a feasible approach to leveraging human intelligence at scale for knowledge creation. Classical per-datum based crowd computing techniques, however, are ill suited for the problem, when the outcome contributions are data instances as opposed to the knowledge in need. Take for example, unknown unknowns of machine learning, which is a major class of errors produced by unreliable AI systems. Such errors are caused by missing or underrepresented concepts in the model. Each of those concepts can be instantiated as various data instances. Crowd computing has been used to detect unknown unknowns and fix them by contributing instances for training data augmentation. Such an approach however, is limited not only in terms of efficiency but also effectiveness, due to the intrinsic shallowness and brittleness of machine learning models.
Interpretable AI by Crowds. Interpretability in AI refers to “the ability to explain or to present in
understandable terms to a human” [
Humans as the object in the definition of AI interpretability implies the following key requirements for
the design of interpretability methods: i) presentation of interpretations need to match humans’ mental
representations of concepts as humans understand the world through concepts that are associated
with observable properties; ii) interpretability methods also need to take the flexible needs of humans
as explanation consumers into account, allowing humans to gain insights about system behavior with
multi-concept queries that involve the (non-)presence of multiple concepts flexibly named by humans.
Existing interpretability methods, however, fail to meet those requirements. Existing local methods
generally generate explanations by highlighting relevant input units – e.g., words in a sentence
or pixels in an image [
A natural approach to fill the semantic gap is involving humans in the interpretation process. Similar to crowd computing for robust AI, where the goal is to characterise what a model has not learned, crowd computing for interpretable AI seeks to explain what a model has learned. The latter again requires crowd computing on the conceptual level for human interpretability and query flexibility. 2.2
Principles for human-AI interaction have been discussed in the HCI community for several years [
Congruence of machine learning models with human understanding. Complex machine learning
models are deployed in several critical domains including healthcare and autonomous vehicles
nowadays, albeit as functional blackboxes. Models which correspond to human interpretation of a
task are more desirable in certain contexts and can help attribute liability, build trust, expose biases
and in turn build better models [
Open-ended Crowd Knowledge Creation. For the purpose of both, robust or interpretable AI,
knowledge creation in real-world machine learning tasks is a complex, open-ended task. Research on
this problem needs to investigate not only the extraction of knowledge from the training data and
model, but also the creation of any knowledge crowds deem as relevant, which can easily go beyond
knowledge encoded in existing knowledge sources. Such a problem is related to multiple ongoing
research lines, such as crowd knowledge creation [
Conversational Human-AI Interaction. Conversational interfaces have been argued to have
advantages over traditional GUIs due to having a more human-like interaction [