Training Neural Networks with Implicit Variance

This paper presents a new model for addressing the errorin-variables problem in Gaussian process autoregressive models by combining a Gaussian process with a deep neural network. The Gaussian process autoregressive model is a simple and effective method for modeling dynamical systems due to its nonlinear, nonparametric, and Bayesian nature. The analytic solution for the marginal log-likelihood is obtained by considering the training input-output relationship to be static, with dynamics modeled through the inclusion of lagged observations in the input regressor. The limitation of this autoregressive method is that both the outputs and inputs are affected by noise. The simulation is obtained iteratively by propagating the Gaussian distribution through a nonlinear function. This results in a costly estimation of the simulated response with Monte Carlo integration. We propose an alternative approach in which a pre-filtering step is performed using a deep neural network to approximate the intractable recurrent filtering of the latent states. The proposed model improves the autoregressive approach while reducing the computational time of the simulation. The proposed model is validated on two case studies: a synthetic example and a real-world problem.

arXiv (Cornell University), 2023

This paper focusses on the optimal implementation of a Mean Variance Estimation network (MVE network) (Nix and Weigend, 1994). This type of network is often used as a building block for uncertainty estimation methods in a regression setting, for instance Concrete dropout (Gal et al., 2017) and Deep Ensembles (Lakshminarayanan et al., 2017). Specifically, an MVE network assumes that the data is produced from a normal distribution with a mean function and variance function. The MVE network outputs a mean and variance estimate and optimizes the network parameters by minimizing the negative loglikelihood. In our paper, we present two significant insights. Firstly, the convergence difficulties reported in recent work can be relatively easily prevented by following the simple yet often overlooked recommendation from the original authors that a warm-up period should be used. During this period, only the mean is optimized with a fixed variance. We demonstrate the effectiveness of this step through experimentation, highlighting that it should be standard practice. As a sidenote, we examine whether, after the warm-up, it is beneficial to fix the mean while optimizing the variance or to optimize both simultaneously. Here, we do not observe a substantial difference. Secondly, we introduce a novel improvement of the MVE network: separate regularization of the mean and the variance estimate. We demonstrate, both on toy examples and on a number of benchmark UCI regression data sets, that following the original recommendations and the novel separate regularization can lead to significant improvements.

International Journal of Neural Systems, 2006

Gaussian processes compare favourably with backpropagation neural networks as a tool for regression, and Bayesian neural networks have Gaussian process behaviour when the number of hidden neurons tends to infinity. We describe a simple recurrent neural network with connection weights trained by one-shot Hebbian learning. This network amounts to a dynamical system which relaxes to a stable state in which it generates predictions identical to those of Gaussian process regression. In effect an infinite number of hidden units in a feed-forward architecture can be replaced by a merely finite number, together with recurrent connections.

2017

We propose a new approach to inverse reinforcement learning (IRL) based on the deep Gaussian process (deep GP) model, which is capable of learning complicated reward structures with few demonstrations. Our model stacks multiple latent GP layers to learn abstract representations of the state feature space, which is linked to the demonstrations through the Maximum Entropy learning framework. Incorporating the IRL engine into the nonlinear latent structure renders existing deep GP inference approaches intractable. To tackle this, we develop a non-standard variational approximation framework which extends previous inference schemes. This allows for approximate Bayesian treatment of the feature space and guards against overfitting. Carrying out representation and inverse reinforcement learning simultaneously within our model outperforms state-of-the-art approaches, as we demonstrate with experiments on standard benchmarks ("object world","highway driving") and a new b...

Gaussian processes have proved to be useful and powerful constructs for the purposes of regression. The classical method proceeds by parameterising a covariance function, and then infers the parameters given the training data. In this thesis, the classical approach is augmented by interpreting Gaussian processes as the outputs of linear filters excited by white noise. This enables a straightforward definition of dependent Gaussian processes as the outputs of a multiple output linear filter excited by multiple noise sources. We show how dependent Gaussian processes defined in this way can also be used for the purposes of system identification.

In recent years there was a tremendous progress in robotic systems, and however also increased expectations: A robot should be easy to program and reliable in task execution. Learning from Demonstration (LfD) offers a very promising alternative to classical engineering approaches. LfD is a very natural way for humans to interact with robots and will be an essential part of future service robots. In this work we first review heteroscedastic Gaussian processes and show how these can be used to encode a task. We then introduce a new Gaussian process regression model that clusters the input space into smaller subsets similar to the work in . In the next step we show how these approaches fit into the Learning by Demonstration framework of [2], . At the end we present an experiment on a real robot arm that shows how all these approaches interact.

arXiv (Cornell University), 2022

It is challenging to guide neural network (NN) learning with prior knowledge. In contrast, many known properties, such as spatial smoothness or seasonality, are straightforward to model by choosing an appropriate kernel in a Gaussian process (GP). Many deep learning applications could be enhanced by modeling such known properties. For example, convolutional neural networks (CNNs) are frequently used in remote sensing, which is subject to strong seasonal effects. We propose to blend the strengths of deep learning and the clear modeling capabilities of GPs by using a composite kernel that combines a kernel implicitly defined by a neural network with a second kernel function chosen to model known properties (e.g., seasonality). We implement this idea by combining a deep network and an efficient mapping based on the Nyström approximation, which we call Implicit Composite Kernel (ICK). We then adopt a sample-then-optimize approach to approximate the full GP posterior distribution. We demonstrate that ICK has superior performance and flexibility on both synthetic and real-world data sets. The ICK framework can be used to include prior information into neural networks in many applications. Preprint. Under review.

2019

A multi-layer deep Gaussian process (DGP) model is a hierarchical composition of GP models with a greater expressive power. Exact DGP inference is intractable, which has motivated the recent development of deterministic and stochastic approximation methods. Unfortunately, the deterministic approximation methods yield a biased posterior belief while the stochastic one is computationally costly. This paper presents an implicit posterior variational inference (IPVI) framework for DGPs that can ideally recover an unbiased posterior belief and still preserve time efficiency. Inspired by generative adversarial networks, our IPVI framework achieves this by casting the DGP inference problem as a two-player game in which a Nash equilibrium, interestingly, coincides with an unbiased posterior belief. This consequently inspires us to devise a best-response dynamics algorithm to search for a Nash equilibrium (i.e., an unbiased posterior belief). Empirical evaluation shows that IPVI outperforms ...

2021 IEEE International Conference on Robotics and Automation (ICRA)

Probabilistic regression techniques in control and robotics applications have to fulfill different criteria of datadriven adaptability, computational efficiency, scalability to high dimensions, and the capacity to deal with different modalities in the data. Classical regressors usually fulfill only a subset of these properties. In this work, we extend seminal work on Bayesian nonparametric mixtures and derive an efficient variational Bayes inference technique for infinite mixtures of probabilistic local polynomial models with well-calibrated certainty quantification. We highlight the model's power in combining data-driven complexity adaptation, fast prediction and the ability to deal with discontinuous functions and heteroscedastic noise. We benchmark this technique on a range of large real inverse dynamics datasets, showing that the infinite mixture formulation is competitive with classical Local Learning methods and regularizes model complexity by adapting the number of components based on data and without relying on heuristics. Moreover, to showcase the practicality of the approach, we use the learned models for online inverse dynamics control of a Barret-WAM manipulator, significantly improving the trajectory tracking performance.

2021 International Joint Conference on Neural Networks (IJCNN), 2021

Deep neural networks tend to underestimate uncertainty and produce overly confident predictions. Recently proposed solutions, such as MC Dropout and SDENet, require complex training and/or auxiliary out-of-distribution data. We propose a simple solution by extending the time-tested iterative reweighted least square (IRLS) in generalised linear regression. We use two sub-networks to parametrise the prediction and uncertainty estimation, enabling easy handling of complex inputs and nonlinear response. The two sub-networks have shared representations and are trained via two complementary loss functions for the prediction and the uncertainty estimates, with interleaving steps as in a cooperative game. Compared with more complex models such as MC-Dropout or SDE-Net, our proposed network is simpler to implement and more robust (insensitive to varying aleatoric and epistemic uncertainty).