Projects funded under national AI programme AISG focus on good training datasets

Sorting recyclable plastics from a pile of rubbish is usually a tedious manual activity.

Plastics need to be sorted into seven categories, ranging from disposable bottles to tougher plastics, in what would appear to be the perfect job for machines powered by artificial intelligence (AI).

But, as a research team collaborating with Sembcorp’s waste management business has learnt, training AI systems with just pictures was not good enough for the accurate sorting of plastic waste.

The data set needed to be supplemented by infrared images to teach AI systems to recognise the seven different types of plastics.

Following this, plastic waste was accurately sorted by the research team’s sensors 95 per cent of the time, up from 85 per cent previously when the system relied on pictures of what needed to be sorted.

Trials at SembWaste’s facility are expected to commence once the project is launched and will pave the way for broader implementation, SembWaste’s head of material recovery facility and special projects Jimmy Fu told The Straits Times, without providing a start date.

The new plastic-sorting system was jointly developed by a team of researchers from the Agency for Science, Technology and Research (A*STAR) and Singapore University of Technology and Design (SUTD), in collaboration with SembWaste.

It is among a growing number of projects being backed by national programme AI Singapore (AISG), which promotes AI adoption here by providing funding and expertise.

In Singapore, there is no shortage of firms chasing the AI dream for better productivity, but not all are ready with the right data sets that will create reliable AI, said AISG’s innovation director Laurence Liew, who oversees collaborations with partnering companies.

The SembWaste project, which comes under one of AISG’s industry technology challenges, seeks to improve the recycling rate of plastics, which stands at 5 per cent.

One of the reasons for the dismal recycling rate is the low accuracy of sorting plastic rubbish by type. For instance, polyethylene terephthalate, used to make disposable bottles, has to be recycled differently from polystyrene, which is used to make styrofoam containers.

Infrared imaging scientist Andrew Ngo from A*STAR said that by refining the data sets, the system’s improved results relied on only 50 samples of hyperspectral images of the seven plastic types, instead of some 6,500 pictures of plastic waste that were used previously.

Dr Ngo said: “The machine can now see the extended wavelength of different plastics, and that’s much more than our eyes. It’s more accurate, and from that we can build a better data set.”

AISG’s Mr Liew mainly oversees a scheme called AISG’s 100 Experiments (100E), an initiative that aims to help 100 local partners deploy AI tools.

The 100E programme has, since 2020, expanded its goal to help up to 200 companies – by providing grants between S$180,000 and S$330,000 – to build and deploy AI systems.

So far, 118 projects have been approved for development, and 90 have been deployed.

Once complete, the AI tool must be deployed within 12 months, otherwise AISG may reclaim the intellectual property and release it under an open-source licence to benefit other companies.

A team of roughly 50 engineers, AI apprentices, and project managers from AISG support these projects.

The initiative has funded projects such as the World Wildlife Fund’s use of AI to improve the accuracy of spotting illegal wildlife trade online, compiling tax-related policies across the globe for sovereign wealth fund GIC, and speeding up the labelling of teeth for dentists at Q&M Dental Group.

Among the most common roadblocks for companies is a lack of training data, which has led to some firms being urged to return to the drawing board, said Mr Liew.

Without good data, AI risks becoming inaccurate, he added.

At Q&M Dental Group, an AI model it developed with MedTech company EM2AI can analyse X-ray scans of a patient’s jaws and label teeth and potential dental problems, as well as generate a recommended treatment plan.

It had to label some 20,000 X-ray scans to train the AI system to achieve between 85 per cent and 95 per cent accuracy.

The AI system, which has since been approved by the Health Sciences Authority, has been successfully rolled out to 150 clinics here and in Malaysia, in what Mr Liew said is among the most successful projects under the 100E initiative so far.

Companies keen to adopt AI can assess their readiness through AISG’s AI Readiness Index, which helps them to gauge whether they have the systems, data, and means to integrate AI models, said Mr Liew.

Inaccuracy is the key reason why people here do not trust AI, according to software company Salesforce’s AI Trust Quotient survey of 545 Singapore workers and 6,000 globally.

Most respondents said they do not trust AI as it lacks the information needed to be useful, or because it relies on outdated data. Four in five respondents said they need to see AI consistently deliver accurate results for them to trust it.

Increasing the accuracy of AI models is not simply about increasing the volume of data, like more images of used plastics, said SUTD Associate Professor Ngai-Man Cheung from its information systems technology and design team. For AI to work well, the right data sets are needed.
 

Source: The Straits Times ©️ SPH Media Limited. Permission required for reproduction.