Approximate inference and Bayesian decision theory
Lead Research Organisation:
University of Cambridge
Department Name: Engineering
Abstract
Solving a decision problem is a two-step procedure: firstly, one performs inference to obtain the approximate posterior distribution, which, secondly, is used to solve said decision problem. The beauty of breaking down Bayesian decision theory in this way, it is usually lectured, is that the two steps can be implemented independently. In reality, however, the inference step can only be performed in an approximate manner; as a consequence, the second step will depend on properties of the approximation in the first step. hence the two steps become coupled. We propose to investigate this coupling.
The project aims to research approximate inference in the context of Bayesian decision theory. We hope to enable the application of Bayesian decision theory in combination with well-established approximate-inference schemes; to qualitatively and quantitatively identify biases in decisions due to performing inference approximately; and to develop techniques that alleviate these biases.
The project aims to research approximate inference in the context of Bayesian decision theory. We hope to enable the application of Bayesian decision theory in combination with well-established approximate-inference schemes; to qualitatively and quantitatively identify biases in decisions due to performing inference approximately; and to develop techniques that alleviate these biases.
Organisations
People |
ORCID iD |
Richard Turner (Primary Supervisor) | |
Wessel Bruinsma (Student) |
Studentship Projects
Project Reference | Relationship | Related To | Start | End | Student Name |
---|---|---|---|---|---|
EP/N509620/1 | 01/10/2016 | 30/09/2022 | |||
1971216 | Studentship | EP/N509620/1 | 05/01/2018 | 30/06/2021 | Wessel Bruinsma |
Description | The key findings so far are academic and can be summarised as follows: 1. A Gaussian process is a particular model for time series, which is typically expensive to apply to problems with many outputs. We found that they can be cheaply applied to many outputs, without requiring approximation. 2. It is possible to build in symmetry into a particular family of neural networks (conditional neural processes), which significantly improves their predictive performance. |
Exploitation Route | As the key findings are mostly academic, they may be taken forward by further research. Alternatively, the developed techniques can be used directly by companies for time series prediction problems. |
Sectors | Digital/Communication/Information Technologies (including Software),Other |
Description | Research on modelling time series with multiple outputs is in use by a company. |