Embedding deep neural networks within Cognitive AI for machine reasoning and automatic parameter tuning
Abstract
Cognitive AI is broadly defined as enabling human level reasoning and intelligence, but specific implementations have been lacking. In this tutorial we will demonstrate practical techniques for embedding deep neural networks within a Cognitive AI framework for medical image segmentation. The increased level of intelligence brings two major benefits: (1) making it easier for a data scientist to teach, train, and optimize the AI, (2) better segmentation accuracy and reliability, with common sense reasoning to avoid obvious mistakes.
Currently the work of a data scientist training deep neural networks involves hand tuning of parameters and applying knowledge ad hoc in pre/post processing algorithms. The result is application-specific Narrow AI, that is typically suboptimal in terms of parameter search and limited as to the level of knowledge and reasoning applied. These shortfalls impact the performance of AI systems and leaves them vulnerable to errors that are obvious to a human, thereby limiting real-world clinical application and adoption.
In this tutorial, participants will learn about Cognitive AI techniques and an open source software framework that supports deep neural networks with higher level intelligence. The new Cognitive AI framework is general-purpose and extensible, supporting AI development tasks that currently require human intelligence, to accomplish them more efficiently and optimally by:
- adding a knowledge base and methods for reasoning about scene content such as spatial inferencing and conditional reasoning to check neural network outputs,
- adding general-purpose process knowledge, in the form of software agents, that can be chained together to accomplish image pre-processing, outputs, neural network prediction, and result post-processing,
- performing automatic end-to-end automatic optimization of all agent parameters and learning hyper parameters to specific medical image segmentation problems.
Description
Deep learning segmentation algorithms are data driven, but rely on human knowledge for image pre-processing, tuning of deep neural network (DNN) architectures and learning hyper parameters, and post-processing of segmented regions. This knowledge is coded ad hoc in scripts, with limited application of common sense reasoning and limited hand tuning of parameters (both in learning and pre/post processing). The result is application-specific Narrow AI, that is typically suboptimal in terms of parameter search and limited as to the level of knowledge and reasoning applied. These shortfalls impact the performance of AI systems and leaves them vulnerable to errors that are obvious to a human, resulting in a loss of trust and limiting real-world clinical application and adoption.
This tutorial will cover the use of Cognitive AI in medical image segmentation to tackle limitations of current Narrow AI approaches. Cognitive AI is broadly defined as enabling human-level reasoning and intelligence. It can provide a layer of machine reasoning atop DNNs, applying knowledge where representative training data may be limited and using reasoning to avoid common sense mistakes. In this tutorial, we demonstrate practical techniques for embedding DNNs within a Cognitive AI framework, including both image content knowledge (e.g., structural and spatial relationships) and processing agent knowledge (e.g., image enhancement and morphological operations). Automatic parameter tuning is applied to discover unexpected high-performing parameter combinations beyond those evaluated by a data scientist (usually constrained due to manual limitations and bias from past experience).
Attendees will gain hands-on experience using a new open source Cognitive AI software framework that supports DNNs with higher level intelligence by:
- adding a knowledge base and methods for reasoning about scene content (e.g. spatial inferencing, conditional reasoning) to check DNN results,
- adding general-purpose process knowledge, in the form of software agents, that can be chained together to accomplish image pre-processing, DNN prediction, and result post-processing,
- performing automatic end-to-end automatic optimization of all agent parameters and learning hyper parameters to specific medical image segmentation problems.
The tutorial will consist of a series of lectures with Q&A, followed by an interactive hands-on session. The lectures will introduce Cognitive AI and its role in supporting DNNs and provide practical training. Case studies in a variety of medical imaging applications and modalities will be presented, showing promising real-world examples of improving the accuracy and reliability of DNNs for image segmentation. The tutorial will be capped with a hands-on demonstration of a new open source tool allowing users to embed DNNs within a Cognitive AI framework.
Audience
The target audience for this tutorial is participants involved in image segmentation using deep neural networks and seeking to improve their performance, reliability, and ease/optimality of training. It will accommodate a range of levels of experience, from those in the early stages of development seeking a framework to speed and support their efforts and also the more advanced who would like to improve the accuracy and reliability of their neural networks. Cognitive AI will be assumed to be a new topic and the tutorial will introduce a variety of techniques as well as a software framework for rapid application.
Materials
Presenter slides and tutorial notes will be distributed to attendees as PDF and will be submitted for publication in the MICCAI Satellite Events joint proceedings.
The open source Cognitive AI software framework demonstrated in the tutorial will be available on GitLab. It will include the agent-based framework for machine reasoning and automatic parameter tuning. It will also provide sample segmentation models derived from public data sets.
Attendees are asked to bring their own laptops for the hands-on session.