In November 2022, StageZero CEO and co-founder Dr. Thomas Forss was invited as a panel speaker at Technology Day - Connected Ink 2022, an event hosted by Wacom, a long-term customer of StageZero. Connected Ink 2022 gathered creatives and thinkers from different industries and cultures to share ideas and inspiration of humanity’s evolution and innovation.
Together with Heidi Wang - Senior Vice President Ink Division and Dr. Markus Weber – Principal Ink Technologist at Wacom for the session “Understanding the meaning of ink”, Dr. Thomas Forss explained how StageZero’s solutions help Wacom’s projects and mission of making digital ink meaningful for artificial intelligence (AI) algorithms. You can see below an interesting recap of what has been discussed during the talk:
Wacom provides cutting-edge digital ink solutions for a wide range of partners who use or produce smartphones, tablets, and digital stationery. By offering innovative products in the education field, Wacom’s digital ink solution helps teachers and students in their daily workflow. Especially, the solution provides teachers with insights that they normally would lack due to emerging remote teaching and learning.
To enhance their digital ink solution, Wacom needed an AI partner who could use Wacom’s own data model for describing the contents of digital ink. Usually, in this case, enterprises have two options: the first option is to start building an in-house team of hundreds of people who deal with data manually; the second option is to work with a reliable partner that can handle data, so that they can put their focus on their core business.
Obviously, option one involves very dull tasks for many employees, high costs for the company to onboard said employees, and multiple risk factors regarding the ramp up time of the project and more. In Wacom’s case, this is when StageZero came in to help them to overcome this challenge. As a partner, StageZero assists Wacom by handling data annotation.
For Wacom’s projects, StageZero collect and label complex digital handwriting samples, including notes, drawings, and scientific formulas with a scalable approach. With a network of over 10 million global contributors for annotation, the resulting datasets offered by StageZero are fully labeled, authentic and diverse. This helps develop the next stages of AI automation in the digital ink space and contributes to Wacom’s improvement of their AI algorithms for semantic content recognition.
“Ink is definitely different from images or text”, Dr. Forss emphasizes, “On a technical level, for instance, an image is represented by many pixels and every pixel has its value, or characters or a group of characters come together to constitute a meaning. But, with digital ink, it’s a different dimension of text: there’s not only characters but every stroke that is drawn is taken into account. Even the sequence that a person draws this character to create a word or the pressure they use are recorded as well. In conclusion, there are so many dimensions to ink that differ from texts that make it a much more complex problem – and that’s the problem that StageZero is helping to solve.”
StageZero’s unique solution involves a crowdsourcing method for which we partner with over 10 million contributors. This diverse group executes bite-sized AI training tasks in exchange for different types of rewards. Hence, instead of having hundreds of thousands of in-house laborers, StageZero provides you with a pool of global users who you can access 24/7 for a significantly quick turnaround time.
StageZero specializes in two main fields: Conversational AI and Natural Language Processing (NLP), with a focus on data sourcing and data annotation. Within Conversational AI, StageZero mostly serves voice assistance, speech recognition, and chatbot. Within NLP, our strengths include text and handwriting, speech and audio, and digital ink.
When asked to forecast emerging technologies in the future that also use AI to solve ink data challenges, Dr. Forss states: “In the text segment, there’s these large language models – for example, you might have heard of GPT3 and other types of underlying models that understand how people use languages either in text or speech. Something similar could be built within digital ink: you’ll train the algorithm first with a large set of hundreds of thousands of notes so it can understand how people write and use this. On top of that, in some cases, you’ll utilize other methods with smaller samples at some point to, for instance, recognize math or something else, and then use that to start predicting different types of labels. After that, you only need people to verify those labels, instead of having to do all things manually.”
“Regarding data augmentation – meaning producing synthetic data on specific existing databases to train an AI model to produce handwriting, there has been some work done on this.", adds Dr. Weber. “You can use existing layouts to create new documents with just existing databases. For instance, the concept of ransom notes – the same thing can be done with ink, then produce more content based on an initial dataset which you have.
Another trend is annotation-free training. An example of this would be using an artificial dataset to train your initial model, and then label some unseen content with pseudo labels. If the network is confident with this model, you’ll use them again to train your system so it can improve itself. The interesting part is that you’ll still need some ground truth data, like an initial labeled dataset for kick-starting.”
Dr. Weber emphasizes that to execute these future technologies, the industry needs more contributors, especially more developers who are interested in exploring new domains such as ink. “That’s why we also try to open up our services to the community. We really want to jointly work on some services – for instance, sharing some initial datasets so you can license data from the beginning. This aims to tackle the barrier of starting out and to help enterprises that need specific targeted data collection, such as Wacom. There are many factors: education is different in different countries, students at a younger age may have different handwritings than the ones in higher grades… - different target groups have different handwritings, so we need really targeted datasets. That’s why we want to help the job get easier by providing the tool because there’s no good tool out there. That’s why we started. We had the universe ink model, we tried to standardize the content schemes where we define the results from handwriting recognition, sketch recognition, and so forth.”
Wacom and StageZero both have the vision of creating a developer community in the future – for Wacom, it will be used to build their own deep learning solutions fueled with ink content data.
Answering the question “What’s your expectation of further scale-up of the adoption of ink annotation technologies that we are building together?” Dr. Forss stresses the need for more data, “…especially the need of getting more data from the community - different stakeholders joining, partners who help build databases of content that can be labeled and used. My ambition for the future is to even build open-source datasets, so stakeholders can participate without having to invest too much in the beginning. When they know what they want to do with them, they can start investing more and so they can just experiment freely at the start basically.”
Dr. Markus Weber gave an example of an initiative generated by a computer graphics community – ImageNet which has millions of annotated image datasets that are available for research. “I can envision something similar for the future, maybe an 'InkNet', contributed by research communities. With such a big public dataset where people can provide their data collection and let those open to be available for other research, we’ll create such an engaging community and hence we can do much more and much better technologies for ink than has ever been seen.”
Dr. Weber also highlights that in the long run, Wacom is going to need much more data and in research, a lot of initiatives start from communities.
Regarding the chance to obtain handwriting data from the crowdsourcing approach, Dr. Forss notes that this depends on the use case. Using this crowdsourcing approach, StageZero specializes in speech data. “The main issue [for obtaining handwriting with this approach] would probably be that the ‘crowd’ needs to have their tablets. Otherwise, it is easy to collect people’s handwriting when they write with their fingers on smartphone screens. In the future, I think we’ll come up with more innovations to solve that problem.”
Link to watch the session: https://www.youtube.com/watch?v=RkuZpd2PmeQ&t=17072s