Research

Overview

Study whether structured background knowledge helps explain sentiment decisions.

Motivation

Ambiguous wording and domain context can make sentiment classification brittle.

Objectives

• Study whether structured background knowledge helps explain sentiment decisions.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Compare a text-only classifier with retrieval of small, curated concept graphs; document how knowledge is selected.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure macro-F1 and calibration on a held-out, licensed corpus; inspect negation and conflicting-context errors.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A licensed sentiment corpus, a small curated knowledge graph and a CPU workstation with optional GPU access.

Potential contributor roles

• Nlp prototyping
• Knowledge curation and evaluation review

Illustrative stage and team requirement

active · Proposed contributor roles: NLP prototyping, knowledge curation and evaluation review. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Knowledge-Enhanced Sentiment Analysis” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore uncertainty-aware fusion while keeping emotion labels descriptive rather than diagnostic.

Motivation

Emotion cues in speech, text and visual signals may disagree, and labels depend on context.

Objectives

• Explore uncertainty-aware fusion while keeping emotion labels descriptive rather than diagnostic.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Build separate text and audio baselines before testing gated fusion; use permitted visual features only where consent allows.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Compare per-class recall and calibration with speaker-separated splits; test missing modalities and disagreement cases.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• Consented multimodal samples, annotation guidance, audio processing tools and optional GPU compute.

Potential contributor roles

• Audio feature design
• Consent-aware data preparation and multimodal evaluation

Illustrative stage and team requirement

planned · Proposed contributor roles: Audio feature design, consent-aware data preparation and multimodal evaluation. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Multimodal Emotion Understanding” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore the tradeoff between perception quality, latency and energy on a bounded robot task.

Motivation

Robots may need timely perception when network access or onboard compute is limited.

Objectives

• Explore the tradeoff between perception quality, latency and energy on a bounded robot task.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Benchmark a compact detector on recorded robot inputs; compare quantization and local execution in simulation before hardware trials.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Report detection metrics, tail latency, memory and energy estimates under repeatable workloads; log missed detections.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• An edge development board, permitted robot recordings, a simulator and a power measurement device.

Potential contributor roles

• Embedded integration
• Robotics simulation and performance measurement

Illustrative stage and team requirement

active · Proposed contributor roles: Embedded integration, robotics simulation and performance measurement. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Edge AI for Robotics” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Prototype understandable feedback for a bounded interaction and explore perceived clarity.

Motivation

Robot feedback can be difficult to interpret when users have different expectations or access needs.

Objectives

• Prototype understandable feedback for a bounded interaction and explore perceived clarity.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Develop a simulated task with alternative feedback designs; prepare a voluntary participant protocol before any user study.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Use task completion, misunderstanding counts and optional usability ratings; examine accessibility differences without claiming generalization.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A robot simulator, interface prototyping tools, accessible study materials and an approved consent process if participants are involved.

Potential contributor roles

• Interaction design
• Accessible interface development and study planning

Illustrative stage and team requirement

planned · Proposed contributor roles: Interaction design, accessible interface development and study planning. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Human–Robot Interaction” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore lightweight confidence checks and graceful fallback when inputs differ from training examples.

Motivation

Edge models may face changing inputs while having limited resources for monitoring.

Objectives

• Explore lightweight confidence checks and graceful fallback when inputs differ from training examples.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Implement uncertainty and input-shift baselines alongside a small edge model; define safe fallback behavior for a simulated task.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure calibration, shift-detection false alarms, abstention coverage and runtime overhead across controlled changes.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• An edge board or emulator, shareable benchmark inputs and profiling tools.

Potential contributor roles

• Model monitoring
• Embedded optimization and robustness review

Illustrative stage and team requirement

planned · Proposed contributor roles: Model monitoring, embedded optimization and robustness review. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Trustworthy Edge Intelligence” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore context-sensitive toxicity classification with transparent label definitions and an appeals-oriented error analysis.

Motivation

Bengali moderation can be affected by spelling variation, code mixing and contextual language use.

Objectives

• Explore context-sensitive toxicity classification with transparent label definitions and an appeals-oriented error analysis.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Use licensed, de-identified text; compare character and multilingual encoder baselines and document annotation disagreement.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Evaluate macro-F1 and false positives on dialect and code-mixed subsets; review errors without exposing identifying text.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A licensed Bengali corpus, annotation guidance, privacy-preserving review tools and optional GPU compute.

Potential contributor roles

• Bengali annotation design
• Nlp modeling and bias evaluation

Illustrative stage and team requirement

active · Proposed contributor roles: Bengali annotation design, NLP modeling and bias evaluation. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Bengali Toxicity Detection” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore explanation stability and failure modes on a permitted image benchmark; no clinical tool or diagnostic claim is proposed.

Motivation

Visual explanations can appear plausible even when a model relies on irrelevant image features.

Objectives

• Explore explanation stability and failure modes on a permitted image benchmark; no clinical tool or diagnostic claim is proposed.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Compare a small image classifier with saliency and perturbation-based explanations using de-identified benchmark data.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Use patient-separated splits where available; test localization consistency, perturbation sensitivity and classifier uncertainty.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A permitted de-identified image dataset, image processing tools and suitable GPU compute.

Potential contributor roles

• Image modeling
• Explanation analysis and dataset governance review

Illustrative stage and team requirement

planned · Proposed contributor roles: Image modeling, explanation analysis and dataset governance review. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Explainable Medical Image Analysis” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Prototype bounded row-following and obstacle avoidance in simulated agricultural scenes.

Motivation

Uneven terrain and changing foliage can complicate navigation in agricultural environments.

Objectives

• Prototype bounded row-following and obstacle avoidance in simulated agricultural scenes.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Combine vision and odometry baselines; vary row geometry and lighting in simulation before considering supervised physical trials.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure cross-track error, collision counts and completion rates; document failures across terrain and lighting variations.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A navigation simulator, permitted field imagery and optional sensors for supervised prototype trials.

Potential contributor roles

• Navigation software
• Simulation scene design and safety-oriented testing

Illustrative stage and team requirement

planned · Proposed contributor roles: Navigation software, simulation scene design and safety-oriented testing. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Agricultural Robot Navigation” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore a reproducible monitoring pipeline that reports uncertainty and missing observations.

Motivation

Low-cost environmental sensors may drift or lose connectivity, making readings difficult to compare.

Objectives

• Explore a reproducible monitoring pipeline that reports uncertainty and missing observations.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Prototype sensor ingestion, timestamp checks and reference-based calibration; compare local buffering with intermittent transmission.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure missing-data rates, reference agreement, timing error and power use; avoid health or regulatory compliance claims.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• Environmental sensors, a microcontroller, a reference instrument and a local storage service.

Potential contributor roles

• Sensor integration
• Data pipeline development and calibration analysis

Illustrative stage and team requirement

active · Proposed contributor roles: Sensor integration, data pipeline development and calibration analysis. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “IoT-Based Environmental Monitoring” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore evidence-linked responses and explicit abstention on a bounded collection of permitted documents.

Motivation

Research assistance can produce unsupported statements when retrieved evidence is weak or irrelevant.

Objectives

• Explore evidence-linked responses and explicit abstention on a bounded collection of permitted documents.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Create a licensed document index; compare keyword and embedding retrieval and require source-grounded answer generation.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Assess retrieval recall, citation support and abstention quality using a manually reviewed question set.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A permitted document collection, indexing tools, a local model or approved inference access and a review rubric.

Potential contributor roles

• Document curation
• Retrieval engineering and evidence verification

Illustrative stage and team requirement

planned · Proposed contributor roles: Document curation, retrieval engineering and evidence verification. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Retrieval-Augmented Research Assistance” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Design a reproducible comparison of small-data adaptation strategies before obtaining any additional data.

Motivation

Limited labeled data can make language models unstable across tasks and writing styles.

Objectives

• Design a reproducible comparison of small-data adaptation strategies before obtaining any additional data.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Compare multilingual transfer, parameter-efficient tuning and simple lexical baselines with a fixed annotation budget.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Track macro-F1, variance across seeds and performance by input length; audit leakage between training and test examples.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A licensed low-resource text corpus, clear label guidance and optional GPU compute.

Potential contributor roles

• Dataset preparation
• Multilingual modeling and reproducibility review

Illustrative stage and team requirement

paused · Proposed contributor roles: Dataset preparation, multilingual modeling and reproducibility review. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Low-Resource Language Understanding” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore anomaly detection that communicates uncertainty and supports human review.

Motivation

Rare defects and changing capture conditions can challenge automated visual inspection.

Objectives

• Explore anomaly detection that communicates uncertainty and supports human review.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Compare supervised and normal-only baselines on a licensed inspection benchmark; introduce controlled lighting changes.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure defect recall, false alarms and localization quality with separate capture conditions; review missed defects.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A licensed defect image dataset, image preprocessing tools and optional GPU compute.

Potential contributor roles

• Vision modeling
• Capture-condition analysis and inspection evaluation

Illustrative stage and team requirement

active · Proposed contributor roles: Vision modeling, capture-condition analysis and inspection evaluation. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Vision-Based Industrial Inspection” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore federated training behavior under heterogeneous data and limited connectivity.

Motivation

Connected devices can hold uneven local datasets while communication and energy budgets differ.

Objectives

• Explore federated training behavior under heterogeneous data and limited connectivity.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Simulate local clients with unequal data partitions; compare centralized, local-only and federated baselines without collecting personal data.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure accuracy, communication volume and worst-client performance; simulate dropouts and assess stated privacy limitations.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A client simulator, a licensed benchmark dataset and local compute for controlled experiments.

Potential contributor roles

• Distributed training
• Client simulation and privacy threat modeling

Illustrative stage and team requirement

planned · Proposed contributor roles: Distributed training, client simulation and privacy threat modeling. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Federated Learning for Connected Devices” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore multimodal classification while documenting uncertainty and the limits of automated moderation.

Motivation

Harmful intent can depend on text-image relationships, context and annotation assumptions.

Objectives

• Explore multimodal classification while documenting uncertainty and the limits of automated moderation.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Use licensed de-identified text-image pairs; compare unimodal baselines with cross-modal fusion and context ablations.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Evaluate macro-F1 and false-positive rates across annotation categories; review disagreement and avoid person-level labeling.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A permitted multimodal corpus, annotation rules, secure review tools and optional GPU compute.

Potential contributor roles

• Multimodal modeling
• Annotation review and fairness analysis

Illustrative stage and team requirement

planned · Proposed contributor roles: Multimodal modeling, annotation review and fairness analysis. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Multimodal Cyberbullying Detection” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Compare practical model compression options under a fixed device budget.

Motivation

Model accuracy alone may not justify memory and energy costs on embedded hardware.

Objectives

• Compare practical model compression options under a fixed device budget.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Benchmark pruning, quantization and compact architectures on an identical recorded workload; pin toolchain and device settings.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure accuracy, peak memory, tail latency and energy per inference; report measurement uncertainty and warm-up effects.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• An embedded board, a power meter, a licensed benchmark and profiling software.

Potential contributor roles

• Firmware integration
• Model optimization and energy measurement

Illustrative stage and team requirement

active · Proposed contributor roles: Firmware integration, model optimization and energy measurement. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Energy-Efficient Embedded Intelligence” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore feedback-based grasp adjustment for a small, explicitly bounded set of objects in simulation.

Motivation

Object shape and placement changes can make fixed grasp strategies unreliable.

Objectives

• Explore feedback-based grasp adjustment for a small, explicitly bounded set of objects in simulation.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Compare fixed and adaptive grasp policies using simulated contact feedback; define safety constraints before physical prototyping.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure grasp success, slip events and recovery attempts on held-out poses; document simulator limitations.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A grasping simulator, shareable object models and optional supervised access to a low-force gripper.

Potential contributor roles

• Grasp planning
• Simulation tooling and constrained control design

Illustrative stage and team requirement

planned · Proposed contributor roles: Grasp planning, simulation tooling and constrained control design. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Adaptive Robotic Grasping” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore interpretable answer paths with explicit handling of missing knowledge.

Motivation

Structured question answering can fail when entity linking is ambiguous or graph coverage is incomplete.

Objectives

• Explore interpretable answer paths with explicit handling of missing knowledge.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Build a small licensed graph; compare template-based querying with learned entity linking and constrained traversal.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure answer accuracy, entity-linking errors and abstention on unsupported questions; inspect path correctness.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A curated licensed graph, a question set, a graph query engine and a CPU workstation.

Potential contributor roles

• Graph curation
• Semantic parsing and answer-path evaluation

Illustrative stage and team requirement

planned · Proposed contributor roles: Graph curation, semantic parsing and answer-path evaluation. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Knowledge Graph Question Answering” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore resilient state estimation under bounded sensor dropouts and misalignment.

Motivation

Sensors have different noise and timing characteristics that can destabilize state estimates.

Objectives

• Explore resilient state estimation under bounded sensor dropouts and misalignment.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Compare a filtering baseline with learned fusion on permitted recorded or simulated sensor streams; model timing offsets explicitly.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure trajectory error, uncertainty consistency and recovery after dropouts; avoid claims of deployment readiness.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A sensor simulator or licensed logs, timestamped IMU and visual inputs and analysis tools.

Potential contributor roles

• State estimation
• Sensor synchronization and uncertainty evaluation

Illustrative stage and team requirement

active · Proposed contributor roles: State estimation, sensor synchronization and uncertainty evaluation. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Sensor Fusion for Autonomous Systems” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Prototype accessible explanations and alternative interaction paths for a bounded technical lesson.

Motivation

Technical learning materials may assume visual, language or device access that learners do not share.

Objectives

• Prototype accessible explanations and alternative interaction paths for a bounded technical lesson.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Review materials against accessibility guidance; develop text and audio alternatives and prepare a voluntary feedback study.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Assess keyboard access, reading clarity and task completion; report feedback only if an approved study is actually conducted.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• Accessible authoring tools, assistive-technology testing software and an optional approved participant process.

Potential contributor roles

• Accessible content design
• Frontend accessibility and learning evaluation

Illustrative stage and team requirement

planned · Proposed contributor roles: Accessible content design, frontend accessibility and learning evaluation. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Accessible Technology Learning” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.

Overview

Explore robust classification with explicit uncertainty and no inference about mental health or personal traits.

Motivation

Noise, speaker differences and label ambiguity can reduce the reliability of speech-emotion classifiers.

Objectives

• Explore robust classification with explicit uncertainty and no inference about mental health or personal traits.
• Document assumptions, limitations and reproducible procedures without claiming completed findings.

Proposed methodology

• Compare acoustic and pretrained speech baselines on consented audio; add controlled noise and use speaker-separated partitions.
• Keep any experiments within approved data permissions and record configuration choices.

Evaluation plan

• Measure macro-F1, calibration and performance degradation by noise level; document label disagreement and language coverage.
• Define a baseline and validation split before comparing approaches; publish no results until work is actually performed.

Required resources

• A licensed consented speech corpus, audio augmentation tools and optional GPU compute.

Potential contributor roles

• Speech feature engineering
• Audio data review and robustness testing

Illustrative stage and team requirement

paused · Proposed contributor roles: Speech feature engineering, audio data review and robustness testing. These are proposed layout details, not evidence of ongoing work or an actual team opening.

Collaboration inquiries

Mention “Robust Speech Emotion Recognition” as an illustrative preview in your inquiry. No researchers, supervisors, results or publications are claimed.