Interactive Learning of Grounded Concepts

                Jens Nevens, Paul Van Eecke, and Katrien Beuls

                             Artificial Intelligence Lab
                              Vrije Universiteit Brussel
                       Pleinlaan 2, B-1050 Brussels, Belgium
                       {jens|paul|katrien}@ai.vub.ac.be

     Autonomous agents perceive the world through streams of continuous sensori-
 motor data. Yet, in order to reason and communicate about their environment,
 agents need to be able to distil meaningful symbolic concepts from their raw
 observations. Without such a repertoire of concepts, communication would need
 to happen by directly sending sensori-motor values. Such a system easily leads
 to miscommunication when perfect calibration is not possible.
     Existing approaches to bridge between the continuous and symbolic domain
 include deep learning techniques (e.g. [1]) and version space learning [2]. Deep
 learning techniques generally achieve high levels of accuracy. However, they
 rely on very large amounts of training data, they often fail to adapt to unseen
 scenarios, and the resulting concepts lack transparency. Version space learning,
 on the other hand, can yield human-interpretable concept representations, but
 are notoriously brittle when faced with noisy training data.
     In this interactive demo, we introduce a novel approach to grounded concept
 learning. Using the language game methodology [3], we set up a tutor-learner
 scenario where the learner is an autonomous agent, grounded in the world using
 a Nao humanoid robot, and the participant is its tutor. Using blocks of various
 shapes, sizes and colours, the tutor first creates a scene. In this scene, the tutor
 chooses a topic and describes it to the learner using an informative concept, such as
‘red’ or ‘cube’. The learner robot observes the world through human-interpretable
 streams of numeric data, such as ‘area’, ‘colour’ and ‘XY-coordinates’. These
 are obtained through standard computer vision techniques. The robot tries to
 find the object described by the tutor. After each interaction, whether is was
 successful or not, the tutor provides feedback to the learner by showing the
 intended topic object. The robot maximally benefits from this feedback to newly
 create or extend the representation of the concept. For each concept, the robot
 has to find out which data streams are important and what the typical values for
 each data stream within a concept are. To make these decisions, the learner makes
 use of the notion of discrimination, i.e. separating one particular object from
 the other objects in the scene. Over the course of many such interactions, the
 learner incrementally and in real-time builds a complete repertoire of concepts
 that is functional in the world. An overview of the experimental set-up is shown
 in Figure 1. A video of the demonstration can be found at https://ehai.ai.
vub.ac.be/demos/interactive-concept-learning.
  Copyright c 2019 for this paper by its authors. Use permitted under Creative
  Commons License Attribution 4.0 International (CC BY 4.0)
2       J. Nevens, P. Van Eecke & K. Beuls




Fig. 1. Overview of the experimental setup. The Nao humanoid robot observes a scene
of blocks with different shapes, colours and sizes.


    During the demonstration, participants are able to inspect the conceptual
system of the agent and follow its evolution. The acquired concepts are completely
transparent and human-interpretable. We show that our approach does not rely
on huge amounts of training data, since forming a repertoire of concepts only
requires a few interactions. Additionally, the resulting concepts are general enough
to be applied to previously unseen objects and can be learned in an incremental
manner. The whole system is adaptive as it does not require us to specify the
number of concepts that should be learned. This completely depends on the
objects observed by the agent, hence there is no need for complete or even partial
retraining when the environment changes. These properties make the approach
well-suited to be used in robotic agents as the module that maps from continuous
sensori-motor input to grounded, symbolic concepts that can then be used for
higher-level reasoning tasks such as planning, explanation or communication.

References
1. Higgins, I., Matthey, L., Glorot, X., Pal, A., Uria, B., Blundell, C., Mohamed, S.,
   Lerchner, A.: Early visual concept learning with unsupervised deep learning. arXiv
   preprint arXiv:1606.05579 (2016)
2. Mitchell, T.M.: Generalization as search. Artificial intelligence 18(2), 203–226 (1982)
3. Steels, L.: Language games for autonomous robots. IEEE Intelligent systems 16(5),
   16–22 (2001)