Keynote Lectures

Agent-based Models for Language Adaptation and Change
Luc Steels, ICREA, Institute of Evolutionary Biology (UPF-CSIC) Barcelona, Spain, Spain

Accountability, Responsibility, Transparency - The ART of AI
Virginia Dignum, Delft University of Technology, Netherlands, Netherlands

Reading Agents that Hunger for Knowledge
Eduard Hovy, Carnegie Mellon University, United States, United States

What We Do: Challenges in Individual Rationality and Its Collective Consequences
Luís Antunes, Universidade de Lisboa, Portugal, Portugal

Abstract
Human natural languages and the conceptual frameworks underlying them are profoundly changing, not just for lexicon or phonology but also for grammar and semantics. Languages thus adapt their expressive power to changing needs of their language communities. They change because of fashions and new communication media (such as texting or twitter). They change also because there is no ideal solution to verbal communication so that a community will continue to navigate in a multi-criterion landscape trying to increase communicative success and decrease cognitive effort for one area while increasing complexity for another. 

I will argue that it is worthwhile to try and model language as a complex adaptive systems, both from the viewpoint of linguistic theory, which has so far failed to come up with adequate explanations for this phenomenon, and from the viewpoint of AI because this challenge pushes us to develop new formalisms for grammar, new language processing mechanisms that are more flexible than currently standard parsers and producers, new learning mechanisms that not only focus on parsing but also on producing in interaction.

The talk will be illustrated with videoclips of our experiments with humanoid robots playing language games. 
The agents are initialized with no or very limited forms of language and then develop grammars and lexicons in order to be successful in their language games. 

Abstract
As robots and other AI systems move from being a tool to being teammates, and are increasingly making decisions that directly affect society,, many questions raise across social, economic, political, technological, legal, ethical and philosophical issues. Can machines make moral decisions? Should artificial systems ever be treated as ethical entities? What are the legal and ethical consequences of human enhancement technologies, or cyber-genetic technologies? What are the consequences of extended government, corporate, and other organisational access to knowledge and predictions concerning citizen behaviour? How can moral, societal and legal values be part of the design process? How and when should governments and the general public intervene?

Answering these and related questions requires a whole new understanding of Ethics with respect to control and autonomy, in the changing socio-technical reality. Means are needed to integrate moral, societal and legal values with technological developments in Artificial Intelligence, both within the design process as well as part of the deliberation algorithms employed by these systems. In this talk I discuss leading Ethics theories and propose alternative ways to model ethical reasoning and discuss their consequences to the design of robots and softbots. Depending on the level of autonomy and social awareness of AI systems, different methods for ethical reasoning are needed. Given that ethics are dependent on the sociocultural context and are often only implicit in deliberation processes, methodologies are needed to elicit the values held by designers and stakeholders, and to make these explicit can lead to better understanding and trust on artificial autonomous systems.
The urgency of these issues is acknowledged by researchers and policy makers alike. Methodologies are needed to ensure ethical design of AI systems, including means to ensure accountability, responsibility and transparency (ART) in system design.

Abstract
True intelligent agenthood (as opposed to mere agency) is characterized by self-driven internal goal creation and prioritization. Few AI systems enjoy the freedom today to autonomously decide what to do next; even robots and planning systems start with a fairly concrete goal and stop acting when they have achieved it. In a small experimental project at CMU we have been exploring what it might mean for a Natural Language text reading engine to experience a ‘hunger for knowledge’ that drives what it chooses to read and learn about next, in an ongoing manner. There is no overall goal other than trying to increase its understanding (coverage and interpretations) of the world as described in Wikipedia. The starting point is a sketchy representation of the Infoboxes of all the people listed in Wikipedia, and the principal criterion for choosing what to read about next is the desire to minimize knowledge gaps and remove inconsistencies. In contrast to Freebase, Knowledge Graphs, and other text mining projects, internal generalization is central to our work. To implement the system we combine traditional AI frame proposition representation for the basic information (to make it readable by humans) with neural networks such as autoencoders to perform generalization and anomaly detection.

Abstract
I start from a notion of individual, situated and multi-varied rationality, and then integrate each individual in its own social context (which includes multiple concomitant relations and permeability from one to another). The role of the modeller is not that of an anonymous avatar for a misplaced notion of scientific objectivity. These activities have a subject, and deal with subjectivity explicitly. In particular, the source of the agents’ own motivations is not exogenous to the system. Goals are acquired through mimetic desire, a mechanism introduced by René Girard, that decomposes motivation in a particular geometric structure. These are the key ingredients to propose a framework for tackling complex social problems through simulation of autonomous computational agents. Within this framework I will present several examples of application to concrete problems. The lecture ends with a challenge for tackling a particularly shocking application field.