Application of Deep Learning Techniques to Commercial Real Estate Appraisal

Researcher: Giuseppe Ughi
Academic Supervisors: Jared Tanner
Industrial Supervisors: Richard Hodgson

New Rock Capital

Background

Currently, obtaining the value of a commercial real estate is a task reliant upon a subjective review undertaken by an agent. This assessment process is slow, manually intensive, and expensive when compared to the evaluation of financial assets which are frequently exchanged on a central market. New Rock Capital Management is a firm focused on real estate analytics, portfolio analysis, and risk management. They are interested in employing publicly available data and machine learning techniques to develop a computerized appraisal system, from which fast, objective, and accurate valuations can be obtained. The implementation of such a system requires the development of a method to objectively determine the quality of the interiors of properties. We aim to deal with this task by applying deep learning techniques on the available images of the estate.

Progress

The first part of this project has consisted of the development of algorithms which construct adversarial attacks in the form of imperceptible perturbations to an image which result in misclassification. For example, in Figure 1, the addition of a small perturbation results in a squirrel being misclassified as a parking meter. Our aim is to develop an extensive testing suite which can be used to compare how susceptible the algorithms we develop are to attacks. By developing these attack routines, we have also become aware of key vulnerabilities which will inform design decisions in the development of our algorithms. We have succeeded in generating and comparing state-of-the-art attack routines [1,2] and we found some defenses that would make attacks highly unlikely.

Figure 1: Illustration of an imperceptible perturbation to an image.

Once we have established the robustness of our algorithms to adversarial attacks, we focused on how to improve their training by suggesting an optimal initialisation. As a matter of fact, our algorithms behave similarly to a regression where the weights are randomly intialised and then trained, i.e. optimised on our task. The initialisation in [3] guarantees that our algorithms have already some knowledge before starting the training.

Publications

[1] G. Ughi, V. Abrol, J. Tanner, A Model-Based Derivative-Free Approach to Black-Box Adversarial Examples: BOBYQA, NeurIPS 2019 Workshop in “Beyond Second order Optimization”. [https://arxiv.org/abs/2002.10349]
[2] G.Ughi, V. Abrol, J. Tanner, An Empirical Study of Derivative-Free-Optimization Algorithms for Targeted Black-Box Attacks in Deep Neural Networks, final review for the Optimisation and Engineering journal. [pre-print: https://arxiv.org/abs/2012.01901]
[3] J. Tanner, G. Ughi, Mutual Information of Neural Network Initialisations: Mean Field Approximations, submitted to ISIT2021 [pre-print: https://arxiv.org/abs/2102.04374]