Hanchen Xiong

Structural Bootstrapping—A Novel, Generative Mechanism for Faster and More Efficient Acquisition of Action-Knowledge

Humans, but also robots, learn to improve their behavior. Without existing knowledge, learning either needs to be explorative and, thus, slow or-to be more efficient-it needs to rely on supervision, which may not always be available. However, once some knowledge base exists an agent can make use of it to improve learning efficiency and speed. This happens for our children at the age of around three when they very quickly begin to assimilate new information by making guided guesses how this fits to their prior knowledge. This is a very efficient generative learning mechanism in the sense that the existing knowledge is generalized into as-yet unexplored, novel domains. So far generative learning has not been employed for robots and robot learning remains to be a slow and tedious process. The goal of the current study is to devise for the first time a general framework for a generative process that will improve learning and which can be applied at all different levels of the robots cognitive architecture. To this end, we introduce the concept of structural bootstrapping-borrowed and modified from child language acquisition-to define a probabilistic process that uses existing knowledge together with new observations to supplement our robots data-base with missing information about planning-, object-, as well as, action-relevant entities. In a kitchen scenario, we use the example of making batter by pouring and mixing two components and show that the agent can efficiently acquire new knowledge about planning operators, objects as well as required motor pattern for stirring by structural bootstrapping. Some benchmarks are shown, too, that demonstrate how structural bootstrapping improves performance.

Towards Sparsity and Selectivity: Bayesian Learning of Restricted Boltzmann Machine for Early Visual Features ⋆

This paper exploits how Bayesian learning of restricted Boltzmann machine (RBM) can discover more biologically-resembled early visual features. The study is mainly motivated by the sparsity and selectivity of visual neurons’ activations in V1 area. Most previous work of computational modeling emphasize selectivity and sparsity independently, which neglects the underlying connections between them. In this paper, a prior on parameters is defined to simultaneously enhance these two properties, and a Bayesian learning framework of RBM is introduced to infer the maximum posterior of the parameters. The proposed prior performs as the lateral inhibition between neurons. According to our empirical results, the visual features learned from the proposed Bayesian framework yield better discriminative and generalization capability than the ones learned with maximum likelihood, or other state-of-the-art training strategies.

Diversity priors for learning early visual features

This paper investigates how utilizing diversity priors can discover early visual features that resemble their biological counterparts. The study is mainly motivated by the sparsity and selectivity of activations of visual neurons in area V1. Most previous work on computational modeling emphasizes selectivity or sparsity independently. However, we argue that selectivity and sparsity are just two epiphenomena of the diversity of receptive fields, which has been rarely exploited in learning. In this paper, to verify our hypothesis, restricted Boltzmann machines (RBMs) are employed to learn early visual features by modeling the statistics of natural images. Considering RBMs as neural networks, the receptive fields of neurons are formed by the inter-weights between hidden and visible nodes. Due to the conditional independence in RBMs, there is no mechanism to coordinate the activations of individual neurons or the whole population. A diversity prior is introduced in this paper for training RBMs. We find that the diversity prior indeed can assure simultaneously sparsity and selectivity of neuron activations. The learned receptive fields yield a high degree of biological similarity in comparison to physiological data. Also, corresponding visual features display a good generative capability in image reconstruction.

Homogeneity analysis for object-action relation reasoning in kitchen scenarios

Modeling and learning object-action relations has been an active topic of robotic study since it can enable an agent to discover manipulation knowledge from empirical data, based on which, for instance, the effects of different actions on an unseen object can be inferred in a data-driven way. This paper introduces a novel object-action relational model, in which objects are represented in a multi-layer, action-oriented space, and actions are represented in an object-oriented space. Model learning is based on homogeneity analysis, with extra dependency learning and decomposition of unique object scores into different action layers. The model is evaluated on a dataset of objects and actions in a kitchen scenario, and the experimental results illustrate that the proposed model yields semantically reasonable interpretation of object-action relations. The learned object-action relation model is also tested in various practical tasks (e.g. action effect prediction, object selection), and it displays high accuracy and robustness to noise and missing data.

Scalable, accurate image annotation with joint SVMs and output kernels

This paper studies how joint training of multiple support vector machines (SVMs) can improve the effectiveness and efficiency of automatic image annotation. We cast image annotation as an output-related multi-task learning framework, with the prediction of each tag׳s presence as one individual task. Evidently, these tasks are related via dependencies between tags. The proposed joint learning framework, which we call joint SVM, is superior to other related models in its impressive and flexible mechanisms in exploiting the dependencies between tags: first, a linear output kernel can be implicitly learned when we train a joint SVM; or, a pre-designed kernel can be explicitly applied by users when prior knowledge is available. Also, a practical merit of joint SVM is that it shares the same computational complexity as one single conventional SVM, although multiple tasks are solved simultaneously. Although derived from the perspective of multi-task learning, the proposed joint SVM is highly related to structured-output learning techniques, e.g. max-margin regression (Szedmak and Shawe-taylor [1]), structural SVM (Tsochantaridis [2]). According to our empirical results on several image-annotation benchmark databases, our joint training strategy of SVMs can yield substantial improvements, in terms of both accuracy and efficiency, over training them independently. In particular, it compares favorably with many other state-of-the-art algorithms. We also develop a “perceptron-like” online learning scheme for joint SVM to enable it to scale up better to huge data in real-world practice.

A Study of Point Cloud Registration with Probability Product Kernel Functions

3D point cloud registration is an essential problem in 3D object and scene understanding. In many realistic circumstances, however, because of noise during data acquisition and large motion between two point clouds, most existing approaches can hardly work satisfactorily without good initial alignment or manually marked correspondences. Inspired by the popular kernel methods in machine learning community, this paper puts forward a general point cloud registration framework by constructing kernel functions over 3D point clouds. More specifically, Gaussian mixtures Based on the point clouds are established and probability product kernel functions are exploited for the registration. To enhance the generality of the framework, SE(3) on-manifold optimization scheme is employed to compute the optimal motion. Experimental results show that our registration framework works robustly when many outliers are presented and motion between point clouds is relatively large, and compares favorably to related methods.

Multi-label Object Categorization Using Histograms of Global Relations

In this paper, we present an object categorization system capable of assigning multiple and related categories for novel objects using multi-label learning. In this system, objects are described using global geometric relations of 3D features. We propose using the Joint SVM method for learning and we investigate the extraction of hierarchical clusters as a higher-level description of objects to assist the learning. We make comparisons with other multi-label learning approaches as well as single-label approaches (including a state-of-the-art methods using different object descriptors). The experiments are carried out on a dataset of 100 objects belonging to 13 visual and action-related categories. The results indicate that multi-label methods are able to identify the relation between the dependent categories and hence perform categorization accordingly. It is also found that extracting hierarchical clusters does not lead to gain in the systems performance. The results also show that using histograms of global relations to describe objects leads to fast learning in terms of the number of samples required for training.

Efficient, General Point Cloud Registration with Kernel Feature Maps

This paper proposes a novel and efficient point cloud registration algorithm based on the kernel-induced feature map. Point clouds are mapped to a high-dimensional (Hilbert) feature space, where they are modeled with Gaussian distributions. A rigid transformation is first computed in feature space by elegantly computing and aligning a small number of eigenvectors with kernel PCA (KPCA) and is then projected back to 3D space by minimizing a consistency error. SE(3) on-manifold optimization is employed to search for the optimal rotation and translation. This is very efficient; once the object-specific eigenvectors have been computed, registration is performed in linear time. Because of the generality of KPCA and SE(N) on-manifold method, the proposed algorithm can be easily extended to registration in any number of dimensions (although we only focus on 3D case). The experimental results show that the proposed algorithm is comparably accurate but much faster than state-of-the-art methods in various challenging registration tasks.

Learning V4 Curvature Cell Populations from Sparse Endstopped Cells

We investigate in this paper the capabilities of learning sparse representations from model cells that respond to curvatures. Sparse coding has been successful at generating receptive fields similar to those of simples cells in area V1 from natural images. We are interested here in neurons from intermediate areas, such as V2 and V4. Neurons on those areas are known to respond to corners and curvatures. Endstopped cells (also known as hypercomplex) are hypothesized to be selective to curvatures and are greatly represented in area V2. We propose here a sparse coding learning approach where the input is not images, nor simple cells, but curvature selective cells. We show that by learning a sparse code of endstopped cells we can obtain different degrees of curvature representations.

Learning undirected graphical models using persistent sequential Monte Carlo

Along with the popular use of algorithms such as persistent contrastive divergence, tempered transition and parallel tempering, the past decade has witnessed a revival of learning undirected graphical models (UGMs) with sampling-based approximations. In this paper, based upon the analogy between Robbins-Monro’s stochastic approximation procedure and sequential Monte Carlo (SMC), we analyze the strengths and limitations of state-of-the-art learning algorithms from an SMC point of view. Moreover, we apply the rationale further in sampling at each iteration, and propose to learn UGMs using persistent sequential Monte Carlo (PSMC). The whole learning procedure is based on the samples from a long, persistent sequence of distributions which are actively constructed. Compared to the above-mentioned algorithms, one critical strength of PSMC-based learning is that it can explore the sampling space more effectively. In particular, it is robust when learning rates are large or model distributions are high-dimensional and thus multi-modal, which often causes other algorithms to deteriorate. We tested PSMC learning, comparing it with related methods, on carefully designed experiments with both synthetic and real-world data. Our empirical results demonstrate that PSMC compares favorably with the state of the art by consistently yielding the highest (or among the highest) likelihoods. We also evaluated PSMC on two practical tasks, multi-label classification and image segmentation, in which PSMC displays promising applicability by outperforming others.