MARS 2020

Our exciting e-Conference focused on topics that have emerged as part of this Future Science Platform’s development and had a blend of novel computer science and application.

Topics throughout MARS 2020 included:

Re-imagining biological prediction through the development, extension and leveraging of machine learning and artificial intelligence approaches to advance analytical frontiers in areas with direct applications to animal and plant breeding.

• MARS 2020 provided the opportunity to explore the analytical challenges further including data integration, algorithm development and biological interference.

Spatio-Temporal application discussions around characterising complex systems and predicting their change through space and time. A key requirement for identifying new resources and adapting to change.

• MARS 2020 provided discussions points to explore challenges including spatio-temporal encoding to ensure generalisation and transfer-ability, approaches to handle data sparsity, noise or errors in data, uncertainty, and data integration given the diversity of data types and scales at which data is collected, learning spatio-temporal co-variance structures from complex S-T data sets.

Decision making – Human in the loop – hear more about algorithms to provide recommendations for decision-making under uncertainty in human operated systems. So far, machine learning and artificial intelligence research has mainly focused on developing near optimal algorithms for decision-making assuming an autonomous agent would learn, plan and implement the best decisions over time. However, many big impact decision problems cannot be effectively solved by humans or machine alone – they are best tackled with both.

• MARS 2020 examined the question ‘How to build machine learning and artificial intelligence with human-in-the-loop solutions, creating the highest performance system, leveraging existing institutional knowledge? Providing best decisions in human operated systems raise interesting challenges such as providing interpret-able, explainable and trusted solutions that remain near optimal.

Object Detection aims at developing new algorithms and tools better able to extract information from various types of images and image-like data. Some of these datasets could be small, or have limited labels, the information could be spread in the time-domain, or across complementary data streams – all these problems require different approaches.

• MARS 2020 provided an opportunity to discuss further the scientific challenge of how to best approach these problems. For example, how to encode and optimise against such non-trivial side constraints. The standard way of encoding simple constraints in the loss function will not be sufficient. Since the scientific core of the work is how to train deep models with such non-trivial constraints, it is very general and applicable across a range of artificial intelligence problems.