Section: New Results

Autonomous And Social Perceptual Learning

Participants : David Filliat [correspondant] , Freek Stulp, Celine Craye, Yuxin Chen, Clement Masson, Adrien Matricon.

Incremental Learning of Object-Based Visual Saliency

Searching for objects in an indoor environment can be drastically improved if a task-specific visual saliency is available. We describe a method to learn such an object-based visual saliency in an intrinsically motivated way using an environment exploration mechanism. We first define saliency in a geometrical manner and use this definition to discover salient elements given an attentive but costly observation of the environment. These elements are used to train a fast classifier that predicts salient objects given large-scale visual features. In order to get a better and faster learning, we use intrinsic motivation to drive our observation selection, based on uncertainty and novelty detection. Our approach has been tested on RGB-D images, is real-time, and outperforms several state-of-the-art methods in the case of indoor object detection. We published these results in two conferences [43] , [42]

Cross-situational noun and adjective learning in an interactive scenario

Learning word meanings during natural interaction with a human faces noise and ambiguity that can be solved by analysing regularities across different situations. We propose a model of this cross-situational learning capacity and apply it to learning nouns and adjectives from noisy and ambiguous speeches and continuous visual input. This model uses two different strategy: a statistical filtering to remove noise in the speech part and the Non Negative Matrix Factorization algorithm to discover word-meaning in the visual domain. We present experiments on learning object names and color names showing the performance of the model in real interactions with humans, dealing in particular with strong noise in the speech recognition. We published these results in a conference paper [41]

Learning representation with gated auto-encoders

We investigated algorithms that would be able to learn relevant visual or multi-modal features from data recorded while the robot performed some task. Representation learning is a currently very active research field, mainly focusing on deep-learning, which investigates how to compute more meaningful features from the raw high dimensional input data, providing a more abstract representation from which it should be easier to make decision or deduction (e.g classification, prediction, control, reinforcement learning). In the context of robotics, it is notably interesting to apply representation learning in a temporal and multi-modal approach exploiting vision and proprioception so as to be able to find feature that are relevant for building models of the robot itself and of its actions and their effect on the environment. Among the many existing approaches, we decided to explore the use of gated auto-encoders, a particular kind of neural networks including multiplicative connections, as they seem well adapted to this problem. Preliminary experimentations have been carried out with gated auto-encoders to learn transformations between two images. We observed that Gated Auto-Encoders (GAE) can successfully find compact representations of simple transformations such as translations, rotation or scaling between two small images. This is however not directly scalable to realistic images such as ones acquired by a robot's camera because of the number of parameters, memory size and compational power it would require (unless drastically downsampling the image which induces sensible loss of information). In addition, the transformation taking an image to the next one can be the combination of transformations due to the movement of several object in the field of view, composed with the global movement of the camera. This induces the existence of an exponential number of possible transformations to model, for which the basic GAE architecture is not suited. To tackle both issue, we are developing a convolutional architectures inspired form Convolutional Neural Networks (CNNs) that provide different modelisations for different parts of the image, which might be usefull to model combinations of transformations. Our Convolutional Gated Auto-Encoder is designed to perform generic feature learning in an unsupervised way (while most CNNs are trained in a supervised fasion) and we are currently testing it on realistic image sequences. We plan to extend this architecture to find relations between modalities as, for instance, proproceptive information and its evolution could be used to predict the next visual features. Similarly, proprioceptive information could be used as a supervising signal to learn visual features.

Learning models by minimizing complexity

In machine learning, it is commonly assumed that simpler models have better chances at generalizing to new, unseen data. Following this principle, we developped an algorithm relying on minimization of a given complexity measure to build a collection of models which jointly explain the observation of the training datapoints. The resulting collection is composed of as few models as possible, each using as few dimensions as possible and each as regular as possible. As of now, each model is a multivariate polynomial, with the complexity of a polynomial of degree N in d variables being N*d+1. The complexity of the collection is the sum of the complexity of all its models. The algorithm starts by associating each datapoint to a local model of complexity 1 (degree 0, no variables), then models are iteratively merged into models of higher complexity, as long as those merges don't increase the complexity of the collection and as long as the resulting models stay within a certain distance of their associated datapoints. We applied this algorithm to the problem of inverse dynamics, which we studied in simulation. For a given robot, torques needed to compensate gravity at equilibrium are entirely determined by the values of its joint angles. As it is common that robots actually perform only low-dimensional tasks, and do not explore their full state space during normal operation, we would like the complexity of our models to mirror the structure of the task. When the task was expressed in the joint space, we got satisfying results on that point, and got good predictions for unseen datapoints. When the task was expressend in end-effector position, it turned out to be impossible to learn the underlying manifolds because a given end-effector position can correspond to various joint configurations, and thus to various torques, making it impossible to predict those torques from the end-effector position alone. We are currently working on applying this model to data generated by an exploration algorithm on a robot arm manipulating objects.