Lorraine Chambers

Doctoral Researcher

School of Computing and Digital Technology - CDT

Email:: lorraine.chambers@mail.bcu.ac.uk

Lorraine is a skilled software engineer with fifteen years of experience designing and developing object-oriented software in the space industry. During this time, she gained an MSc in Software Engineering, for which she had a paper published on Sentiment Analysis.

Areas of Expertise

Qualifications

Memberships

Research

A Streaming Approach to Real-Time Data Discrepancy Detection in Deep Neural Networks

Deep neural networks have become popular due to the increased availability of data and computational power. They are also widely used in the data streaming field due to the volume of data available from sources such as Internet of Things sensors. Confidence is required in deep neural networks operating in a streaming environment so that unseen instances that vary from the training data will be captured and their classification initially manually analysed. This will provide information that can be fed back into the deep neural network to correctly identify similar future instances automatically and provide an evolving system that can capture suspicious data and stay up to date with changing input data in a timely manner for data stream processing.

Deep neural networks consist of many individual linear functions which operate together hierarchically through non-linear activation functions to create a complex non-linear function. During the training process each individual function is autonomously assigned weights and biases. There are likely to be millions of these values and it is the complexity and enormity of these networks which makes it difficult for humans to determine how a deep neural network arrives at its classification decision. If the new unseen instances are different from that which the deep neural network was trained on, they will activate different neurons within the network, possibly resulting in incorrect classification. This can happen as a result of `Adversarial Attacks' ranging from where as little as one pixel is changed causing miss-classification, to where noise is added to the entire image; `Concept Drift' where a background change in data causes incorrect classification (this could be incremental, recurrent, gradual or abrupt); `Concept Evolution' where a completely new class is seen that the deep neural network was not trained on. These issues affect the designers of deep neural networks and their customers. The magnitude of the impact this has also depends on the application domain of the deep neural network. For instance, in application domains such as automated cars or health care systems, incorrect classification could lead to safety issues. Deep Neural Network Inspection is a research area that has gained much interest over recent years. The applicable part with respect to this research is the method that has been used to determine which neurons in the hidden layers have been activated. Once this is known, the activated neurons for different classes at run-time can be compared. Research has focused on this area with regards to static data, but not in a streaming environment. Progression is required in this area to provide greater real-time confidence in deep neural networks in order to progress them to the future.

This project will examine the deep neural network to provide a representation of the hidden layers and determine what units were activated for an unseen instance and compare that to units activated during training. If it is detected that the instance’s activations are not similar to that activated during training for that same classification, the instance will be flagged as suspicious. This will allow the user to analyse the instance and label it as `Adversarial Attack', `Concept Drift' or `Concept Evolution'. This data can be fed back into the system so that other similar instances can be automatically classified in a timely manner for streaming data.