Investing in Innovation

"Innovation is not simply inventing something new" - How AI-driven projects could change Bangladesh

Bangladesh's oldest university wants artificial intelligence built for the country's own healthcare, farming and education problems rather than imported from richer economies. One project in its robotics depa9rtment, a sign-language translation model paid for by a Bangladeshi government grant, shows what that produces and what currently limits it.

Curzon Hall on the University of Dhaka campus.

Key Points

The University of Dhaka wants AI research aimed at Bangladesh's healthcare, agriculture, climate and education problems, drawing on faculties well beyond computer science. Its vice-chancellor argues the country should be creating technology rather than consuming it.
One project shows what that produces. SignFormer-GCN, published in PLOS ONE in February 2025, translates continuous sign language video into written text for languages with very little training data available, with Bangla Sign Language as the target.
The model uses 9.43 million parameters against 115.41 million for the leading comparable method, and trained in roughly six hours on one graphics card where the comparison needed 60 hours across four. That efficiency is what lets a university without a data centre attempt the work at all.
Limited research funding, weak industry R&D spending and thin infrastructure cap how much of this Bangladesh can currently do. The university is asking international partners for joint laboratories and shared supervision rather than memoranda of understanding.

"Innovation is not simply inventing something new," says Prof A.B.M. Obaidul Islam, Vice-Chancellor of the University of Dhaka. "It is turning knowledge and ideas into solutions that create meaningful value for society. A new idea becomes innovation when it can solve a real problem, when it can be implemented, and when it can create sustainable impact." Islam, a physicist, took office in March 2026 with a stated ambition to move Bangladesh "from a culture of consuming technology to a culture of creating knowledge and technology."

The university wants AI built across medicine, law and agriculture

"AI should not be limited to computer science only," Islam says. "We have expertise across science, medicine, pharmacy, business, law, social sciences, humanities and engineering, and an opportunity to bring these disciplines together to develop responsible AI solutions for Bangladesh." The targets he names are healthcare, climate and environmental monitoring, agriculture and education. In a country with a young, dense population and a shortage of specialists, he argues AI answers a gap rather than creating one. "We don't have that many skilled people," he says. "From that point of view, AI will help us."

He is candid about the other side of it. "When you build dynamite, it has some positive impact and also some negative impact," he says. "If AI is everywhere, our people will become lazy, or the innovative mentality will become less."

A sign language project shows what that looks like in practice

The clearest working example sits in the university's Department of Robotics and Mechatronics Engineering. In PLOS ONE in February 2025, Safaeid Hossain Arib, Rabeya Akter and Dr Sejuti Rahman at Dhaka, working with Shafin Rahman at North South University, published SignFormer-GCN, a model that converts continuous sign language video directly into written sentences. The work was supported by a Special Innovation Grant from Bangladesh's Ministry of Science and Technology for the 2023-24 financial year, awarded to Dr Rahman, whose research covers AI in healthcare, medical robotics and machine learning.

Most translation systems depend on gloss annotation, a written intermediate labelling of each sign, which requires trained sign language professionals and is expensive to produce. SignFormer-GCN skips that step and translates video straight to text. The paper states it is the first to combine both video and body-keypoint features for gloss-free translation.

How It Works

The system reads the video twice. One stream takes the raw picture and passes it through a transformer, the architecture behind most modern language models, which handles broad context across a whole sequence.
The second stream tracks the signer's skeleton, plotting the positions of joints frame by frame into a graph of how the body moves.
That graph passes through a spatio-temporal graph convolutional network, which picks up fine detail a transformer misses, such as the precise relationship between one joint and another during a sign.
The two streams are added together and passed to a decoder that generates the sentence word by word.

The model runs on a twelfth of the parameters of its rival

The efficiency figures are the part that matters outside a computer science department. SignFormer-GCN carries 9.43 million parameters. GFSLT-VLP, the method it is measured against, carries 115.41 million. Training took about six hours across 200 epochs on a single NVIDIA RTX 3090; the comparison method required 60 hours across two training phases on four of the same cards. At inference the Dhaka model answers in 0.03 to 0.034 seconds against 0.8 to one second.

Those numbers describe what a research group can attempt when it does not have a data centre, and they are the reason a project like this is reproducible elsewhere in Bangladesh rather than confined to one well-equipped lab.

The University of Dhaka is ready to engage with the world. We invite scientists and tech companies to work with us as friends and genuine partners."

Bangla Sign Language data does not look like German or American data

The team tested the model on three datasets: a German corpus drawn from television weather forecasts, an American corpus of instructional videos, and BornilDB v1.0, a Bangla Sign Language dataset. The Bangla set is the hard one, and the paper says why. It contains three signers, two men and one woman, filmed with varying camera quality, changing backgrounds and inconsistent lighting. With so few people signing, a model risks learning those three individuals rather than the language, which the authors identify as a source of bias limiting real-world use.

This is the practical version of the argument for building locally. A model trained on European and North American sign language corpora does not transfer to a language with three signers on file, and no international group has an incentive to fix that. The Dhaka team set the first published benchmark on BornilDB v1.0 and released their code on GitHub and their processed features on Kaggle, so the next group working on Bangla does not start from nothing.

Thin funding and weak industry R&D cap how much of this happens

Islam is direct about the constraints. "We have less research funding. We don't have advanced infrastructure," he says, adding that Bangladeshi industry "cannot support research and development the way we are asking for." The sign language work reached publication on a domestic government grant and a single graphics card, which is both the encouraging part and the limit: the same team attempting something that needed sustained computing time, or a dataset built from scratch with hundreds of signers, would not get there on the same budget.

Dhaka asks for joint labs rather than another MOU

"We are particularly interested in partnerships that go beyond signing an MOU," Islam says. "Sometimes it stays in the paperwork, but it doesn't work effectively." What the university wants is specific: joint research, joint laboratories, student and faculty exchanges, PhD supervision, joint publications, and technology transfer on externally funded projects, with joint degree and certification programmes as the longer-term goal. The same robotics department already holds a contract with Imperial College London and University College Dublin under Wellcome Leap's In Utero programme, which is the shape being asked for.

Founded in 1921, the University of Dhaka is marking its 106th year. "The University of Dhaka is ready to engage with the world," Islam says. "We invite universities, scientists, research organisations, industry and technology companies to work with us not only as donors or collaborators, but as genuine research partners, as friends, as development partners."

Sources

Innovation Report interview with Prof A.B.M. Obaidul Islam, Vice-Chancellor, University of Dhaka, 12 August 2026.
Arib, S.H., Akter, R., Rahman, S., Rahman, S., "SignFormer-GCN: Continuous sign language translation using spatio-temporal graph convolutional networks", PLOS ONE 20(2): e0316298, 14 February 2025.
SignFormer-GCN source code, GitHub.
Department of Robotics and Mechatronics Engineering, University of Dhaka.