Computational Intelligence and Operations Lab - CIOL

🎉 Huge Congratulations to the CIOL Team! 🎉

We are incredibly proud to announce that our researchers at the Computational Intelligence and Operations Laboratory (CIOL) have been awarded 2nd Place for the 2026 Health Systems Best Track Paper at the IISE Annual Conference & Expo 2026 in Arlington, TX! 🏆

Please join us in congratulating our outstanding team members on this massive achievement. The work was led by Azmine Toushik Wasi , with contributions from Mahfuz Ahmed Anik and MD Shafikul Islam Sohan , and guidance from Manjurul Ahsan! Congratulations to the entire team on this remarkable accomplishment.

Their award-winning paper, "Multimodal Vision-Language Models for Automated and Explainable Postoperative Complication Risk Stratification," introduces GROVE (GROunded Vision–languagE risk model for ICU).

About the Research:
Postoperative deterioration is a major challenge in intensive care, often hidden by alert fatigue and the cognitive burden of tracking diverse clinical data. The team's GROVE framework tackles this by acting as a multimodal vision-language risk model that bridges the gap between unstructured clinical notes and structured physiological measurements (like heart rate and oxygen saturation mapped as temporal plots).
By jointly modeling these quantitative and qualitative signals, GROVE generates interpretable, natural language assessments that achieved an AUROC of 0.92 for early detection of complications like sepsis and hemorrhage—with 94% of outputs rated as evidence-aligned by clinicians!
This is a massive step forward for human-centric clinical informatics and actionable decision support in high-stakes medical environments. We are incredibly proud of the dedication and innovation our members continue to bring to the field of biomedical AI! 🚀🏥

🚨 Excited to share that our paper, 𝐀𝐈 𝐂𝐨𝐥𝐥𝐚𝐛𝐨𝐫𝐚𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐭𝐡𝐞 𝐃𝐞𝐜𝐞𝐧𝐭𝐞𝐫𝐢𝐧𝐠 𝐨𝐟 𝐇𝐮𝐦𝐚𝐧 𝐂𝐫𝐞𝐚𝐭𝐢𝐯𝐢𝐭𝐲, has been published in 𝐀𝐂𝐌 𝐀𝐈 𝐋𝐞𝐭𝐭𝐞𝐫𝐬! 🇧🇩

Congratulations to Azmine Toushik Wasi and Sadia Tasnim Meem from Shahjalal University of Science and Technology for this contribution exploring the evolving relationship between human creativity and generative AI.

For centuries, Western philosophical traditions have framed creativity as an exclusively human capacity, grounding authorship and artistic agency in human consciousness, rationality, and intention. The rapid advancement of generative AI now challenges these assumptions by introducing computational systems capable of contributing meaningfully to creative production.

In this work, the authors examine how human–AI collaboration destabilizes traditional anthropocentric models of creativity. Drawing on posthumanist and new materialist perspectives, the paper argues that creativity should be understood as a distributed process emerging through interactions among human and non-human actors rather than as an exclusively human attribute.

The paper further explores how contemporary AI systems increasingly participate in creative workflows through idea generation, variation exploration, and iterative collaboration. By analyzing different levels of human–AI co-creation, ranging from AI as a digital tool to AI as a co-creator, the work highlights how creative agency is becoming increasingly distributed across computational and human networks.

The study also addresses broader philosophical and practical questions surrounding authorship, originality, aesthetic value, and responsibility in AI-mediated creative systems. Through a post-anthropocentric lens, the paper proposes a framework that repositions creativity as an emergent property of interactions among datasets, algorithms, interfaces, environments, and human interpretation.

This work contributes to ongoing discussions at the intersection of AI, philosophy, creativity, and human–computer interaction, highlighting how generative systems are reshaping established understandings of artistic production and creative agency.

🚨 Excited to share that our journal article 𝐒𝐓-𝐑𝐞𝐬𝐆𝐀𝐓: 𝐄𝐱𝐩𝐥𝐚𝐢𝐧𝐚𝐛𝐥𝐞 𝐒𝐩𝐚𝐭𝐢𝐨-𝐓𝐞𝐦𝐩𝐨𝐫𝐚𝐥 𝐆𝐫𝐚𝐩𝐡 𝐍𝐞𝐮𝐫𝐚𝐥 𝐍𝐞𝐭𝐰𝐨𝐫𝐤 𝐟𝐨𝐫 𝐑𝐨𝐚𝐝 𝐂𝐨𝐧𝐝𝐢𝐭𝐢𝐨𝐧 𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐏𝐫𝐢𝐨𝐫𝐢𝐭𝐲-𝐃𝐫𝐢𝐯𝐞𝐧 𝐌𝐚𝐢𝐧𝐭𝐞𝐧𝐚𝐧𝐜𝐞 has been published in Intelligent Transportation Infrastructure! 🇧🇩

Huge congratulations to Mohsin Mahmud Topu for leading the ideation and experiments; Azmine Toushik Wasi for continuous support and guidance throughout the project; Mahfuz Ahmed Anik for his valuable contributions; and Md Manjurul Ahsan for his support in this work!

Road infrastructure in climate-vulnerable regions is increasingly exposed to rapid deterioration, while maintenance resources remain limited. Traditional “fix-on-failure” strategies are often reactive, costly, and inefficient for large transportation networks.

In this work, we introduce ST-ResGAT, a novel explainable spatio-temporal graph neural network that combines graph attention learning with temporal modeling to forecast pavement deterioration across connected road networks. Unlike conventional approaches, the framework directly maps predicted pavement conditions into ASTM-compliant maintenance priority levels, making the system more practical for real-world deployment.
Using a real-world dataset of 750 road segments collected in Sylhet, Bangladesh between 2021–2024, our experiments show that incorporating spatial road-network topology substantially improves prediction reliability over traditional non-spatial machine learning baselines. We further integrate GNNExplainer to ensure that the model’s learned decision patterns remain interpretable and aligned with established engineering principles.
This work highlights how explainable AI and graph neural networks can support safer, more efficient, and resource-aware infrastructure management in high-risk and developing regions.

🔗 Read it here: https://doi.org/10.1093/iti/liag006

বাংলাদেশের তরুণ গবেষকদের আন্তর্জাতিক AI গবেষণায় আরেকটি উল্লেখযোগ্য সাফল্য!

Machine Learning এবং Artificial Intelligence গবেষণার অন্যতম শীর্ষ আন্তর্জাতিক সম্মেলন 𝐈𝐂𝐌𝐋 𝟐𝟎𝟐𝟔-এর main track-এ regular paper হিসেবে accepted হয়েছে Bangladesh-led student research collaboration-এর পেপার, “𝐓𝐢𝐦𝐞𝐒𝐩𝐨𝐭: 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐢𝐧𝐠 𝐆𝐞𝐨-𝐓𝐞𝐦𝐩𝐨𝐫𝐚𝐥 𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠 𝐢𝐧 𝐕𝐢𝐬𝐢𝐨𝐧–𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 𝐢𝐧 𝐑𝐞𝐚𝐥-𝐖𝐨𝐫𝐥𝐝 𝐒𝐞𝐭𝐭𝐢𝐧𝐠𝐬”। ICML 2026 অনুষ্ঠিত হবে South Korea-এর Seoul-এ।

এই পেপারে co-lead হিসেবে আছেন 𝐀𝐳𝐦𝐢𝐧𝐞 𝐓𝐨𝐮𝐬𝐡𝐢𝐤 𝐖𝐚𝐬𝐢 (IPE, 𝐒𝐔𝐒𝐓) এবং 𝐒𝐡𝐚𝐡𝐫𝐢𝐲𝐚𝐫 𝐙𝐚𝐦𝐚𝐧 𝐑𝐢𝐝𝐨𝐲 (CSE, 𝐍𝐒𝐔)। Co-author হিসেবে আছেন 𝐊𝐨𝐮𝐬𝐡𝐢𝐤 𝐀𝐡𝐚𝐦𝐞𝐝 𝐓𝐨𝐧𝐦𝐨𝐲 (𝐍𝐒𝐔), 𝐊𝐢𝐧𝐠𝐚 𝐓𝐬𝐡𝐞𝐫𝐢𝐧𝐠 (𝐍𝐒𝐔), 𝐒. 𝐌. 𝐌𝐮𝐡𝐭𝐚𝐬𝐢𝐦𝐮𝐥 𝐇𝐚𝐬𝐚𝐧 (𝐍𝐒𝐔), এবং 𝐖𝐚𝐡𝐢𝐝 𝐅𝐚𝐢𝐬𝐚𝐥 (𝐒𝐔𝐒𝐓)। কাজটির supervisors ছিলেন 𝐃𝐫. 𝐓𝐚𝐬𝐧𝐢𝐦 𝐌𝐨𝐡𝐢𝐮𝐝𝐝𝐢𝐧 এবং 𝐃𝐫. 𝐌𝐝 𝐑𝐢𝐳𝐰𝐚𝐧 𝐏𝐚𝐫𝐯𝐞𝐳, দুজনেই 𝐐𝐚𝐭𝐚𝐫 𝐂𝐨𝐦𝐩𝐮𝐭𝐢𝐧𝐠 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐈𝐧𝐬𝐭𝐢𝐭𝐮𝐭𝐞 (𝐐𝐂𝐑𝐈)-এর Research Scientist।

বর্তমান Vision–Language Models বা VLMs ছবি দেখে object, scene, landmark, এবং অনেক সময় location চিনতে বেশ ভালো করছে। তবে real-world intelligence শুধু “ছবিতে কী আছে” বোঝার মধ্যে সীমাবদ্ধ নয়। একটি intelligent multimodal system-এর বুঝতে পারা উচিত ছবিটি কোথায় তোলা হয়েছে, কখন তোলা হয়েছে, দিনের কোন সময়, আলো-ছায়া কী বলছে, এবং visual evidence কীভাবে physical world-এর সাথে connected।

এই গবেষণায় introduce করা হয়েছে 𝐓𝐢𝐦𝐞𝐒𝐩𝐨𝐭, একটি benchmark for evaluating 𝐠𝐞𝐨-𝐭𝐞𝐦𝐩𝐨𝐫𝐚𝐥 𝐮𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠 in Vision–Language Models under real-world settings। Benchmark-টিতে রয়েছে 𝟏,𝟒𝟓𝟓 ground-level images from 𝟖𝟎 countries, যেখানে models-কে test করা হয় তারা visual evidence alone থেকে ছবিটি 𝐰𝐡𝐞𝐫𝐞 এবং 𝐰𝐡𝐞𝐧 captured হয়েছে তা infer করতে পারে কি না।

তাদের evaluation দেখায়, strong spatial recognition থাকা মানেই reliable temporal reasoning থাকা নয়। অনেক state-of-the-art open-source এবং closed-source VLM image-এর place বা visual context চিনতে পারলেও 𝐭𝐢𝐦𝐞-𝐨𝐟-𝐝𝐚𝐲 𝐩𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧, 𝐠𝐞𝐨𝐝𝐞𝐬𝐢𝐜 𝐩𝐫𝐞𝐜𝐢𝐬𝐢𝐨𝐧, এবং 𝐠𝐞𝐨-𝐭𝐞𝐦𝐩𝐨𝐫𝐚𝐥 𝐜𝐨𝐧𝐬𝐢𝐬𝐭𝐞𝐧𝐜𝐲-তে গুরুত্বপূর্ণ limitation দেখায়।

এই কাজটি গুরুত্বপূর্ণ কারণ future AI systems যদি navigation, robotics, disaster response, climate monitoring, geospatial intelligence, বা real-world decision-making-এ ব্যবহার করতে হয়, তাহলে শুধু static image recognition যথেষ্ট হবে না। Models-কে পৃথিবীর জায়গা, সময়, আলো, environment, এবং physical constraints একসাথে বুঝতে হবে। 𝐓𝐢𝐦𝐞𝐒𝐩𝐨𝐭 সেই gap systematically measure করার জন্য একটি benchmark তৈরি করেছে।

এই অর্জন দেখায় যে বাংলাদেশের student researchers এবং young AI community future multimodal AI systems কীভাবে real world বুঝবে, reason করবে, এবং grounded intelligence অর্জন করবে—সেই দিকেও গুরুত্বপূর্ণ অবদান রাখছে।

Happy to see our work highlighted in Short Stories!

Artificial Intelligence and Machine Learning এ বিশ্বের most prestigious and most selective এবং Top 3 grand-slam conference এর একটি হলো International Conference on Machine Learning (ICML), যার 2026 সালের এডিশনটি South Koreaএর Seoul এ অনুষ্ঠিত হবে! ICML 2026 কনফারেন্সের পজিশন পেপার ট্র্যাকে Spotlight হিসেবে জায়াগা পেয়েছে বাংলাদেশের student-led, community-based AI রিসার্চ ল্যাব Computational Intelligence and Operations Laboratory (CIOL)-এর পেপার, "Position: AI Governance Needs ISO-like Interoperability Protocols, Not Just Laws"। ICML 2026-এ Position Paper ট্র্যাকে মাত্র ৩৯টি পেপারকে Spotlight এর মর্যাদা দেয়া হয়েছে!

পেপারে lead-author হিসেবে আছেন আজমাইন তৌসিক ওয়াসি (SUST - Industrial and Production Engineering থেকে)। এছাড়াও কো-অথর হিসেবে আছেন রাফিয়া ইসলাম (IUB থেকে) এবং মাহফুজ আহমেদ অনিক (SUST - Industrial and Production Engineering থেকে)। এডভাইজর ছিলেন Hanyang University থেকে Prof. Dong-Kyu Chae, Univerrsity of Oklahoma থেকে Prof. Md Manjurul Ahsan, এবং Hanyang University-থেকে Taki Hasan Rafi! ওয়াসি, রাফিয়া, অনিক এবং মঞ্জুরুল আহসান - সবাই CIOL ল্যাবের সদস্য!

বিশ্বব্যাপী AI দ্রুত ছড়িয়ে পড়ার সাথে সাথে বিভিন্ন অঞ্চলে আলাদা আলাদা নিয়ম তৈরি হচ্ছে। Europe, USA, এবং China—প্রত্যেকের নিজস্ব AI governance framework রয়েছে, কিন্তু এগুলো একে অপরের সাথে aligned না। ফলে cross-border AI deployment হয়ে যাচ্ছে slow, expensive এবং অনেক ক্ষেত্রে impractical। বড় কোম্পানিগুলো এই complexity manage করতে পারলেও ছোট team এবং startups পিছিয়ে পড়ছে।

এই গবেষণায় তারা দেখান যে, AI systems-এর জন্য দরকার standardized, machine-readable disclosures; শুধু legal framework-এর উপর নির্ভর করে এখানে প্রোপার Control and Governance পাওয়া সম্ভব না। এজন্য তারা প্রতিটি মডেলে "nutrition label" দেয়ার প্রস্তাব করেন। এটাকে food-এর nutrition label-এর মতোই। যেমন food label থেকে আমরা জানতে পারি ভিতরে কী আছে, ঠিক তেমনি একটি “AI label” জানাবে: model কতটা biased, energy consumption কত, training data কোথা থেকে এসেছে, কী ধরনের risks আছে, এবং আরো অনেক তথ্য।

ফলে regulators আলাদা করে আবার সব ব্যাখ্যা করতে হবে না, একই structured information সবাই পড়তে পারবে, companies-এর compliance burden কমবে, এনং small teams ও startups-এর জন্য global deployment সহজ হবে। এভাবে এটি একটি common language of trust তৈরি করবে AI governance-এর জন্য।

এটা দেখায় যে বাংলাদেশের low-resource environment থেকে স্টুডেন্টরা শুধু AI/ML অ্যালগরিদম নিয়েই গবেষণা করছে না, বরং ভবিষ্যতে AI কীভাবে governed হবে এবং পরিচালিত হবে, সেই দিকেও গুরুত্বপূর্ণ গবেষণায় অবদান রাখছে।

🚨 Excited to share our work: 𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧: 𝐓𝐢𝐦𝐞-𝐒𝐞𝐫𝐢𝐞𝐬 𝐅𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧 𝐌𝐨𝐝𝐞𝐥𝐬 𝐑𝐞𝐪𝐮𝐢𝐫𝐞 𝐄𝐱𝐩𝐥𝐢𝐜𝐢𝐭 𝐃𝐨𝐦𝐚𝐢𝐧-𝐋𝐞𝐯𝐞𝐥 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐬 has been accepted as a Position Paper at International Conference on Machine Learning (ICML) 2026! 🇧🇩

Huge congratulations to Md Asif Bin Syed, Md Younus Ahamed, and Azmine Toushik Wasi for this work!

Time-series foundation models are now widely used across domains like healthcare, finance, energy, transport, and retail. However, current evaluation benchmarks mix all these domains together, which can hide critical weaknesses in real-world settings.
Our study shows that model performance changes significantly across domains due to differences in data structure, noise levels, and temporal patterns. Experiments across 7 models and 72 datasets confirm that global leaderboard scores often do not reflect domain-specific reliability.
We argue that trustworthy deployment requires domain-level benchmarks, so models are evaluated where they are actually used, not only on aggregated global scores.

ICML is an A*-tier conference and widely regarded as one of the most prestigious venues in artificial intelligence and machine learning. ICML is also considered part of the AI/ML “grand slam” (the “big three”), alongside ICLR and NeurIPS!

🚨 A defining moment for Bangladesh in global AI 🇧🇩

Proud to see this milestone featured in The Daily Star, one of the country’s most widely read and respected news outlets.
🔗 https://www.thedailystar.net/campus/news/bangladeshi-team-earns-spotlight-icml-ai-governance-research-4165646

Our work, Position: AI Governance Needs ISO-like Interoperability Protocols, Not Just Laws, has been accepted as a Spotlight Position Paper at ICML 2026, marking a rare and historic achievement for Bangladesh in top-tier AI research.

Grateful to the CIOL team, collaborators, and mentors for making this possible. This recognition reflects not just a paper, but Bangladesh’s growing presence in shaping global AI governance discourse.

Bangladeshi team earns spotlight at ICML for AI governance research The paper, “Position: AI Governance Needs ISO-like Interoperability Protocols, Not Just Laws,” is scheduled to be presented on May 11 in South Korea's Seoul

🚨 Excited to share our work: 𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧: 𝐀𝐈 𝐆𝐨𝐯𝐞𝐫𝐧𝐚𝐧𝐜𝐞 𝐍𝐞𝐞𝐝𝐬 𝐈𝐒𝐎-𝐥𝐢𝐤𝐞 𝐈𝐧𝐭𝐞𝐫𝐨𝐩𝐞𝐫𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥𝐬, 𝐍𝐨𝐭 𝐉𝐮𝐬𝐭 𝐋𝐚𝐰𝐬 has been accepted as a 𝐒𝐩𝐨𝐭𝐥𝐢𝐠𝐡𝐭 Position Paper at ICML2026!!

Huge congratulations to Azmine Toushik Wasi, Mst Rafia Islam, Mahfuz Ahmed Anik for their contributions and to Dr. Manjurul Ahsan and Prof. Dong-Kyu Chae for their guidance and support!

As AI systems become deeply embedded in global infrastructure, regulation is currently split across regions, with frameworks like the EU AI Act, US NIST guidelines, and China’s governance policies evolving independently. This fragmentation makes cross-border deployment complex and costly.
Our work argues that laws alone are not sufficient. Instead, we propose ISO-like interoperability protocols that allow AI systems to communicate risk and compliance in a standardized, machine-readable form across jurisdictions.
At the core of this idea are AI “nutrition labels”: structured, unified reports describing key properties such as bias, energy usage, data provenance, and safety characteristics. These enable consistent verification without repeated legal reinterpretation.
We further propose modular, versioned standards inspired by GDPR-era technical compliance (such as ISO 27001), designed to evolve alongside AI systems while maintaining global compatibility.
The goal is a shift from fragmented legal compliance to interoperable technical governance, enabling a shared global language for responsible AI deployment.

ICML is an A*-tier conference and widely regarded as one of the most prestigious venues in artificial intelligence and machine learning. ICML is also considered part of the AI/ML “grand slam” (the “big three”), alongside ICLR and NeurIPS! This paper marks only the second time a Bangladeshi contribution has received Spotlight recognition at ICML and the firstinstance in the Position Paper track!

🚨 Excited to share our work: 𝐓𝐢𝐦𝐞𝐒𝐩𝐨𝐭: 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐢𝐧𝐠 𝐆𝐞𝐨-𝐓𝐞𝐦𝐩𝐨𝐫𝐚𝐥 𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝𝐢𝐧𝐠 𝐢𝐧 𝐕𝐢𝐬𝐢𝐨𝐧–𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 𝐢𝐧 𝐑𝐞𝐚𝐥-𝐖𝐨𝐫𝐥𝐝 𝐒𝐞𝐭𝐭𝐢𝐧𝐠𝐬 accepted at the International Conference on Machine Learning 2026 (ICML 2026)!

Huge congratulations to Azmine Toushik Wasi and Shahriyar Zaman Ridoy for co-leading this work. Thanks to Koushik Ahamed Tonmoy, Kinga Tshering, S. M. Muhtasimul Hasan, and Wahid Faisal for the contributions and to Md Rizwan Parvez and Tasnim Mohiuddin for their guidance.

Vision-language models are increasingly used to interpret the physical world, but their ability to reason about where and when an image was captured remains limited. While current systems can often detect landmarks or obvious visual cues, they struggle with deeper geo-temporal understanding, such as estimating season, time of day, climate context, or geographic region from visual evidence alone.
We introduce TimeSpot, a benchmark designed to evaluate this gap in real-world settings. It contains 1,455 ground-level images from 80 countries and tests models on structured prediction of both spatial attributes (continent, country, climate zone, environment type, coordinates) and temporal attributes (season, month, time of day, daylight phase), along with reasoning tasks under uncertainty.
Evaluations show that state-of-the-art vision-language models perform poorly on temporal reasoning and struggle with consistent geo-temporal inference. Even fine-tuning improves performance only modestly, indicating that current approaches are not yet capable of robust physically grounded understanding.
TimeSpot provides a structured framework to measure how well models connect visual signals to real-world space and time, highlighting a key limitation in current multimodal AI systems.

ICML is an A*-tier conference and widely regarded as one of the most prestigious venues in artificial intelligence and machine learning. ICML is also considered part of the AI/ML “grand slam” (the “big three”), alongside ICLR and NeurIPS!