Abstract—Exploratory activities seem to be intrinsically rewarding for children and crucial for their cognitive development. Can a machine be endowed with such an intrinsic motivation system? This is the question we study in this paper, presenting a number of computational systems that try to capture this drive towards novel or curious situations. After discussing related research coming from developmental psychology, neuroscience, developmental robotics, and active learning, this paper presents the mechanism of Intelligent Adaptive Curiosity, an intrinsic motivation system which pushes a robot towards situations in which it maximizes its learning progress. This drive makes the robot focus on situations which are neither too predictable nor too unpredictable, thus permitting autonomous mental development. The complexity of the robot’s activities autonomously increases and complex developmental sequences self-organize without

How does an individual use the knowledge acquired through self exploration as a manipulable model through which to understand others and benefit from their knowledge? How can developmental and social learning be combined for their mutual benefit? In this paper we review a hierarchical architecture (HAMMER) which allows a principled way for combining knowledge through exploration and knowledge from others, through the creation and use of multiple inverse and forward models. We describe how Bayesian Belief Networks can be used to learn the association between a robot’s motor commands and sensory consequences (forward models), and how the inverse association can be used for imitation. Inverse models created through self exploration, as well as those from observing others can coexist and compete in a principled unified framework, that utilises the simulation theory of mind approach to mentally rehearse and understand the actions of others. 1.

How can we build artificial agents that can autonomously explore and understand their environments? An immediate requirement for such an agent is to learn how its own sensory state corresponds to the external world properties: It needs to learn the semantics of its internal state (i.e., grounding). In principle, we as programmers can provide the agents with the required semantics, but this will compromise the autonomy of the agent. To overcome this problem, we may fall back on natural agents and see how they acquire meaning of their own sensory states, their neural firing patterns. We can learn a lot about what certain neural spikes mean by carefully controlling the input stimulus while observing how the neurons fire. However, neurons embedded in the brain do not have direct access to the outside stimuli, so such a stimulus-to-spike association may not be learnable at all. How then can the brain solve this problem? (We know it does.) We propose that motor interaction with the environment is necessary to overcome this conundrum. Further, we provide a simple yet powerful criterion, sensory invariance, for learning the meaning of sensory states. The basic idea is that a particular form of action sequence that maintains invariance of a sensory state will express the key property of the environmental stimulus that gave rise to the sensory state. Our experiments with a sensorimotor agent trained on natural images show that sensory invariance can indeed serve as a powerful objective for semantic grounding.

Artificial neural networks can model cortical local learning and signal processing, but they are not the brain, neither are many special purpose systems to which they contribute. Autonomous mental development models all or part of the brain (or the central nervous system) and how it develops and learns autonomously from infancy to adulthood. Like neural network research, such modeling aims to be biologically plausible. This paper discusses why autonomous development is necessary according to a concept called task muddiness. Then it introduces recent results for a series of research issues, including the new paradigm for autonomous development, mental architectures, developmental algorithm, a refined classification of types of machine learning, spatial complexity and time complexity. Finally, the paper presents some experimental results for applications, including: visionguided navigation, object finding, object-based attention (eye-pan), and attention-guided pre-reaching, four tasks which infants learn to perform early but very perceptually challenging for robots. Key words: cognitive development, autonomous learning, mental architecture, on-line learning, incremental learning, visual learning, working memory, long-term memory, self-organization, regression, autonomous navigation, attention selection, object recognition,

Currently, there is a lack of general-purpose, in-place learning engines that incrementally learn multiple tasks, to develop “soft ” multi-task-shared invariances in the intermediate internal representation while a developmental robot interacts with its environment. In-place learning is a biologically inspired concept, rooted in the genomic equivalence principle, meaning that each neuron is responsible for its own development while interacting with its environment. With in-place learning, there is no need for a separate learning network. Computationally, biologically inspired, in-place learning provides unusually efficient learning algorithms whose simplicity, low computational complexity, and generality are set apart from typical conventional learning algorithms. We present in this paper the multiple-layer in-place learning network (MILN) for this ambitious goal. As a key requirement for autonomous mental development, the network enables both unsupervised and supervised learning to occur concurrently, depending on whether motor supervision signals are available or not at the motor end (the last layer) during the agent’s interactions with the environment. We present principles based on which MILN automatically develops invariant neurons in different layers and why such invariant neuronal clusters

This paper presents Incremental Hierarchical Discriminant Regression (IHDR) which incrementally builds a decision tree or regression tree for very high dimensional regression or decision spaces by an online, real-time learning system. Biologically motivated, it is an approximate computational model for automatic development of associative cortex, with both bottom-up sensory inputs and top-down motor projections. At each internal node of the IHDR tree, information in the output space is used to automatically derive the local subspace spanned by the most discriminating features. Embedded in the tree is a hierarchical probability distribution model used to prune very unlikely cases during the search. The number of parameters in the coarse-to-fine approximation is dynamic and data-driven, enabling the IHDR tree to automatically fit data with unknown distribution shapes (thus, it is difficult to select the number of parameters up front). The IHDR tree dynamically assigns long-term memory to avoid the loss-of-memory problem typical with a global-fitting learning algorithm for neural networks. A major challenge for an incrementally built tree is that the number of samples varies arbitrarily during the construction process. An incrementally updated probability model, called sample size dependent negative-log-likelihood (SDNLL) metric is used to deal with large-sample size cases, small-sample size cases, and unbalanced-sample size cases, measured among different internal nodes of the IHDR tree. We report experimental results for four types of data: synthetic data to visualize the behavior of the algorithms, large face image data, continuous video stream from robot navigation, and publicly available data sets that use human defined features.

Abstract — The cerebral cortex uses a large number of topdown connections, but the roles of the top-down connections remain unclear. Through end-to-end (sensor-to-motor) multilayered networks that use three types of connections (bottom-up, lateral, and top-down), the new Topographic Class Grouping (TCG) mechanism shown in this paper explains how the topdown connections influence (1) the type of feature detectors (neurons) developed and (2) their placement in the neuronal plane. The top-down connections boost the variations in the neuronal between class directions during the training phase. The first outcome of this top-down boosted input space is the facilitation of the emergence of feature detectors that are purer, measured statistically by the average entropy of the neurons ’ development. The relatively purer neurons are more “abstract, ” i.e., characterizing class-specific (or motorspecific) input information, resulting in better classification rates. The second outcome of this top-down boosted input space is the increase of the distance between input samples that belong to different classes, resulting in a farther separation of neurons according to their class. Therefore, neurons that respond to the same class become relatively nearer. This results in TCG, measured statistically by a smaller within-class scatter of responses when the neuronal plane has a fixed size. Although these mechanisms are potentially applicable to any pattern recognition applications, we report quantitative effects of these mechanisms for 3D object recognition of center-normalized, background-controlled objects. TCG has enabled a significant reduction of the recognition errors. I.

Abstract—Scaffolding is a process of transferring learned skills to new and more complex tasks through arranged experience in open-ended development. In this paper, we propose a developmental learning architecture that enables a robot to transfer skills acquired in early learning settings to later more complex task settings. We show that a basic mechanism that enables this transfer is sequential priming combined with attention, which is also the driving mechanism for classical conditioning, secondary conditioning, and instrumental conditioning in animal learning. A major challenge of this work is that training and testing must be conducted in the same program operational mode through online, real-time interactions between the agent and the trainers. In contrast with former modeling studies, the proposed architecture does not require the programmer to know the tasks to be learned and the environment is uncontrolled. All possible perceptions and actions, including the actual number of classes, are not available until the programming is finished and the robot starts to learn in the real world. Thus, a predesigned task-specific symbolic representation is not suited for such an open-ended developmental process. Experimental results on a robot are reported in which the trainer shaped the behaviors of the agent interactively, continuously, and incrementally through verbal commands and other sensory signals so that the robot learns new and more complex sensorimotor tasks by transferring sensorimotor skills learned in earlier periods of open-ended development. Index Terms—Attention, classical conditioning, incremental learning, instrumental conditioning, mental architecture, mental development, multitask learning, online learning, scaffolding, skill transfer. I.

What is available to developmental programs in autonomous mental development, and what should be learned at the very early stages of mental development? Our observation is that sensory and motor primitives are the most basic components present at the beginning, and what developmental agents need to learn from these resources is what their internal sensory states stand for. In this paper, we investigate the question in the context of a simple biologically motivated visuomotor agent. We observe and acknowledge, as many other researchers do, that action plays a key role in providing content to the sensory state. We propose a simple, yet powerful learning criterion, that of invariance, where invariance simply means that the internal state does not change over time. We show that after reinforcement learning based on the invariance criterion, the property of action sequence based on an internal sensory state accurately reflects the property of the stimulus that triggered that internal state. That way, the meaning of the internal sensory state can be firmly grounded on the property of that particular action sequence. We expect the framing of the problem and the proposed solution presented in this paper to help shed new light on autonomous understanding in developmental agents such as humanoid robots.

This position paper proposes that the study of embodied cognitive agents, such as humanoid robots, can advance our understanding of the cognitive development of complex sensorimotor, linguistic and social learning skills. This in turn will benefit the design of cognitive robots capable of learning to handle and manipulate objects and tools autonomously, to cooperate and communicate with other robots and humans, and to adapt their abilities to changing internal, environmental, and social conditions. Four key areas of research challenges are discussed, specifically for the issues related to the understanding of: (i) how agents learn and represent compositional actions; (ii) how agents learn and represent compositional lexicons; (iii) the dynamics of social interaction and learning; and (iv) how compositional action and language representations are integrated to bootstrap the cognitive system. The review of specific issues and progress in these areas is then translated into a practical roadmap based on a series of milestones. These milestones provide a possible set of cognitive robotics goals and test-scenarios, thus acting as a research roadmap for future work on cognitive developmental robotics.