Month: October 2022

JGI Seed Corn Funding Project Blog 2021: Michael Rumbelow and Alf Coles

Introduction

This seed corn project aimed to explore and extend uses of AI-generated data in an educational context. We have worked on an AI-based app to recognise, gather data on and respond to children’s arrangements of wooden blocks in mathematical block play. The project was inter-disciplinary in two ways. Firstly, the people involved crossed disciplines (teachers, academics, programmers) and, secondly, the app itself provokes engagement in creative activities involving music, chemistry and mathematics.

Developing an app to recognise real-world block play

Block play is a popular activity among children. And in schools there has also been a resurgence in the use of physical blocks in primary mathematics classrooms, particularly in the teaching of maths (drawing on some East Asian practices of using physical blocks as concrete models of abstract mathematical concepts). We were interested in researching children’s interactions with physical blocks, with the aim of supporting their learning across the curriculum, and one of the key challenges was how to capture data on children’s interactions with blocks for analysis.

Previous studies of block play have focused on gathering data variously through sketching or taking photos or videos of children’s block constructions, or embedding radio transmitters in blocks which could transmit their positions and orientations. Recently developments in computer vision technology offer novel ways of capturing data on block play. For example, photogrammetry apps such as 3D Scanner can now create 3D digital models from images or video of objects taken on mobile phones, and AI-based object recognition apps are increasingly able to detect objects they have been trained to ‘see’.

With funding from the JGI we were able to form a small project team of two researchers in the School of Education, a software developer and the head of a local primary school, in order to develop an app to trial with children in the school (see Figure 1).

Figure 1. The experimental set-up as used in the initial trial in a primary school

Technical Developments

Over the course of the JGI project we have developed the app in the following ways:

• We have rebuilt the app architecture robustly around the Detectron-2 AI algorithm, to facilitate reliable data gathering, training and feature development.
• We have developed a new mode to enable gathering of data on mathematical block play around proportion (ie detection of relative block sizes as well as adjacency) and carbon chemistry modelling (ie detection of multiple row block adjacency).
• We have made improvements to the user interface (eg removal of text from the screen for when testing with pre-literate users).
• We have tested the app with 5-6 year olds in a primary school.
• The new version generates snap shots of children’s block arrangements and exports data on their positions to a spreadsheet which allows further analysis.

Lessons Learned

As well as developing a prototype, we have been able to trial this phase of development in school, giving us several valuable insights into both the technical development of AI computer vision apps for gathering anonymous data on block play in schools, as well as the usability and potential of apps controlled by children via the arrangement of physical blocks on a tabletop. In particular we have found:

• Benefits of using platforms available to the target audience as and when feasible. Our aim was to develop an app which is ultimately usable by schools. At the time of development, the AI algorithms used required processing power beyond standard laptops to run at reasonable speeds, and dedicated AI processing hardware such as the Nvidia Jetson NX offered sufficient processing power at a fraction of the cost of higher-end GPU equipped laptops. However, during development, due largely to global chip shortages, this price difference disappeared, Jetson NX’s became scarce, and we decided to switch to higher-end GPU-equipped Windows laptops. This has simplified installation and portability of the app without the need for specialist hardware and opened a route to incremental optimization for the types of standard lower-spec laptops used in schools, as well as easing technical maintenance, and sharing and processing of the data gathered in standard apps such as spreadsheets.
• The resilience of trained artificial neural network algorithms in practice, as well as the importance of responsively optimising training image datasets. The app was trained to recognise blocks using training datasets originally gathered with a specific higher-spec webcam at a fixed distance from the table, which required a separate support apparatus. In practice when we tried using low-cost webcams with their own built-in gooseneck support these worked relatively well, at a variety of heights, and in a variety of lighting and tabletop environments in the field, and were much more practical to set up. However, dips in reliability became apparent in certain lighting conditions, for instance in distinguishing red and pink blocks, which highlighted the need for fresh training datasets using the new webcam, focusing on these areas of ambiguity apparent in field-testing.
• Children’s patience, curiosity and creativity in using novel technology. We had minimised the textual buttons in the interface designed for the researcher, to change modes etc, in the assumption that young children would not want to have to bother with them and that their presence might be confusing. In practice children, having seen the buttons used during set-up, were curious to bring up and explore all of the interface buttons themselves. They were also patient when the app occasionally did not immediately detect a block, ‘helping’ it to ‘see’ the block by nudging its position or re-laying it. And rather than copying what they had seen researchers or other children doing, the children were creative in exploring the affordances of the app, for example trying laying blocks horizontally rather than vertically, or reversing the order of a melody played by placing blocks.

Above all, this phase of development and trialling has provided evidence of the feasibility of producing an app which can use AI to detect and respond to block placements by young children in the field, and highlighted several of the key challenges for next steps.

Future Challenges

The potential uses of the app are extensive and, following on from the successes of this JGI project, we now want to:

• Develop our app, which is currently a prototype, into something potentially ready to move into production.
• Engage with Research Software Engineering (RSE) at the University of Bristol, to support further app development.
• Trial and hone the tools and games to support learning using the app
• Extend the dataset of images used to train the app from several hundreds to several thousands, aligned with the diverse webcams and conditions likely in the field
• Pilot the app with visually impaired and blind children
• Pilot the app with teachers interested in teaching climate chemistry
• Develop an anonymised dataset of children’s block play, including creative free play and guided mathematical block play (inspired by the UoB’s EPIC-KITCHENS data set https://epic-kitchens.github.io/2020-100)
• Enable upload, storage and visualisation of data on block arrangements on a server, for potential research analysis using AI to detect patterns
• Extend the app to recognise stacked as well as laid-flat block constructions, making use of LIDAR technology.

We are currently taking part in a training programme (SHAPE “Pre-accelerator” course) to help us plan the next stages of development.

Bayesian methods in Neuroscience – Conor Houghton

For the last century science has relied on a statistical framework based on hypothesis testing and frequentist inference. Despite its convenience in simple contexts this approach has proved to be intricate, obtuse and sometimes misleading when applied to more difficult problems, particularly problems with the sort of large, complex and untidy datasets that are vital for applications like climate modelling, finance, bioinformatics, epidemiology and neuroscience.

Bayesian inference solves this; the Bayesian approach is easy to interpret and returns science to its traditional reliance on evidence and description rather than a false notion of significance and truth. With a rigorous handling of uncertainty Bayesian inference can dramatically improve statistical efficiency, allowing us to squeeze more insight out of finite, hard-won data which in turn reduces animal and biological tissue use and reduces costs for scientific projects.

With support from the Jean Golding Institute we ran a workshop about Bayesian Modelling: our workshop had lots of different elements, a tutorial for people unfamiliar with the approach, short talks by people in the University who use these methods, a few talks by external speakers and a data study group. In retrospect, we did try to do too much, but the workshop was very helpful, the short talks brought together the local community around Bayesian Modelling and the two external speakers, Hong Ge and Mike Peardon, were excellent and provided real unexpected insight into the current and potential future state of Bayesian Modelling.

We hope to next host a workshop on Hybrid / Hamiltonian Monte Carlo; HMC has quickly become a very useful tool in data science, allowing us to perform Bayesian inference for a host of real world problems that would not have been tractable a few years ago. Perhaps surprisingly, HMC has its origins in high energy particle physics and was invented to perform the high-dimensional integrals involved in Quantum Chromodynamics, the calculations required to predict the results of collider experiments in CERN.

We believe that there is a still a lot these two communities  – particle physics and applied data science – can learn from each other when exploring and developing the power and scope of HMC.