AI Tech Inspire

DeepModAI 2025 CFP: Multimodal Deep Learning at ICONIP, Nov 24 in Okinawa

Multimodal AI is quietly reshaping how systems see, hear, read, and act. If your work mixes text, images, audio, video, or sensor streams, this call is likely worth a calendar block. At AI Tech Inspire, we spotted an upcoming workshop that squarely targets the practical and theoretical edges of multimodal learning—where alignment, robustness, and efficiency collide.

Quick facts: what’s being announced

  • Event: DeepModAI 2025 — International Workshop on Deep learning for Multimodal Data.
  • When/Where: November 24, 2025, Okinawa, Japan (co-located with ICONIP 2025).
  • Focus: Advanced approaches (unsupervised, self-supervised, weakly supervised) to learn transferable latent representations across modalities; beyond unimodal and static setups.
  • Applications encouraged: Multimodal document analysis, health monitoring, autonomous systems, robotics, environmental modeling.
  • Topics include: Multi-view/multimodal architectures; cross-modal alignment/translation; attention for dynamic fusion; diversity-aware and ensemble methods; explainable and collaborative frameworks; adaptability to dynamic/incomplete/context-dependent data; scalable deployment and computational efficiency.
  • Submissions: Extended abstracts (2 pages) or regular papers (any length). Regular papers should have a preprint (e.g., arXiv, Jxiv) before submission. Accepted works appear as talks or posters and will be published on the workshop website.
  • Deadlines: Submission by September 30, 2025; workshop on November 24, 2025.
  • Program: Invited keynotes, technical talks, posters, and an interactive panel with international experts.
  • Details and contact: Workshop website, email deepmodai@sciencesconf.org.

Why this CFP matters to builders, not just academics

Multimodal systems have jumped from research demos to production: think text-to-image models like Stable Diffusion, vision-language assistants based on GPT, and perception stacks in robotics. But many teams still struggle with practical issues: fusing signals that arrive out-of-sync, dealing with missing sensors, aligning representations learned in noisy real-world settings, and deploying bulky models under tight latency and memory constraints.

This workshop’s emphasis on unsupervised, self-supervised, and weakly supervised methods signals a push toward learning from partially labeled or unlabeled data—often the reality in domains like healthcare and autonomy. For engineers who’ve bolted together modality-specific models in PyTorch or TensorFlow, the call here is to build shared latent spaces that transfer knowledge across modalities and tasks.

Key takeaway: The center of gravity is shifting from unimodal silos to shared, transferable representations that can adapt to context, data gaps, and deployment constraints.

From alignment to action: themes worth exploring

Cross-modal alignment and translation is everywhere in this CFP—think CLIP-style contrastive loss for matching images and text, or sequence-to-sequence translation between audio and video. The alignment question isn’t just academic: better alignment means more reliable downstream results, fewer brittle failures, and easier transfer to new tasks. Engineers might ask: can a single learned space handle OCR text plus page layout signals and then generalize to chart understanding? What about aligning noisy wearable data with clinical notes?

Attention for dynamic modality fusion matters for streaming contexts. Imagine a robot that must reweight audio vs. vision when motors get noisy, or an autonomous drone balancing LiDAR with RGB in low light. Attention-driven fusion allows a system to dynamically scale contributions from each channel—making it an attractive alternative to naive concatenation. You could experiment with transformer-style cross-attention or gated fusion layers to handle missing or delayed inputs.

Diversity-aware and ensemble methods suggest a recognition that no single model nails every context. For production, ensemble thinking can reduce variance, and diversity-aware training can prevent overfitting to dominant modalities.

Explainable and collaborative frameworks are especially relevant in regulated and safety-critical domains. If a model flags an anomaly using both ultrasound and EHR notes, clinicians need visibility into which modality drove the decision and why. Expect discussions around attribution for multimodal architectures, modality-specific saliency, and human-in-the-loop review interfaces.

Adaptability and incomplete data: the messy real world

Real deployments rarely get pristine, synchronized data. The workshop calls out adaptability to dynamic, incomplete, or context-dependent data. For developers, that could translate to architectures that gracefully degrade—operating under a missingness mask without collapsing. Techniques like cross-modal imputation, mixture-of-experts for routing, and curriculum strategies for missing modalities are all ripe for exploration.

Consider environmental modeling: fusing satellite imagery, weather station data, and social signals. Sensors go offline; clouds block views; sampling frequency varies. Can your model continue producing stable forecasts with confidence estimates, then snap back when data quality improves?

Scalable deployment: when research meets production

On-device robotics and resource-constrained settings bring computational efficiency front and center. If you’ve wrestled with GPU memory while fine-tuning, you know the score. Expect conversations about LoRA adapters, sparse attention, mixed precision, and graph optimizations. On the runtime side, CUDA-level tuning and operator fusion can make or break latency targets. The workshop’s nod to scalability invites contributions that turn heavy multimodal models into something you can ship—whether through quantization-aware training, distillation, or clever caching.

For practitioners publishing code, hosting models on Hugging Face with lightweight demo spaces is a practical way to make results reproducible and community-friendly. Bonus points for clean dataset wrappers, DataLoader scripts, and evaluation harnesses that mimic operational chaos—dropped frames, off-by-one timestamps, or blank sensor packets.

Real-world angles the CFP encourages

  • Multimodal document analysis: Combine OCR text, page layout, and visual cues to parse complex tables and charts. Can alignment methods reduce hallucinated numbers?
  • Health monitoring: Fuse wearables, clinical notes, and imaging. What’s the minimum viable modality set for robust early warning without overwhelming clinicians?
  • Autonomous systems and robotics: Balance camera, LiDAR, IMU, and audio for navigation and interaction. How should fusion policies shift under sensor failures?
  • Environmental modeling: Blend satellite imagery, weather, and ground observations. Can weak supervision unlock scale when labels are sparse?

Submission mechanics (and a few tips)

There are two tracks: extended abstracts (2 pages) and regular papers (any length). A notable detail: regular papers should first be posted to a preprint server (e.g., arXiv, Jxiv) before submission to the workshop. Accepted works will be presented as talks or posters and will appear on the workshop website—not a heavy proceedings overhead, which may suit ongoing or exploratory projects.

“Bring your in-progress ideas.” The format favors sharing early insights, collecting feedback, and iterating fast.

Practical tips observed by the AI Tech Inspire audience over the years:

  • Show the failure modes. A strong negative result (e.g., when cross-modal attention collapses) is valuable to the community.
  • Benchmark against unimodal baselines. Demonstrate that each modality pulls its weight.
  • Report latency, memory, and throughput alongside accuracy. If you can’t deploy it, you can’t use it.
  • Release a minimal, runnable README with a few test samples; even better, a small demo.py and a notebook.

Idea sparks you could spin into a submission

  • Doc triage assistant: Text+layout+diagram embeddings to auto-route insurance claims; explore contrastive alignment vs. late fusion.
  • Wearable + EHR risk scoring: Self-supervised pretraining on unlabeled time series; weakly supervised fine-tuning with limited chart labels.
  • Audio-visual anomaly detection for robots: Cross-attention fusion that falls back gracefully when microphones saturate.
  • Satellite + weather forecasting: Imputation strategies for cloud-obscured regions; uncertainty estimates at the pixel level.

The logistics: mark your calendar

Submission deadline is September 30, 2025. The workshop runs on November 24, 2025 in Okinawa, Japan, alongside ICONIP 2025. Details, topics, and updates live on the official page: deepmodai.sciencesconf.org. Questions can be sent to deepmodai@sciencesconf.org. The organizers also encourage sharing the call with colleagues, PhD students, and postdocs.

If multimodal learning is on your roadmap, this looks like the right room to test ideas, compare notes, and pressure-test deployment plans. The themes—alignment, dynamic fusion, explainability, and efficiency—line up closely with what practitioners are wrestling with right now. Consider sketching a 2-page abstract this week. Future-you may thank you when the panel Q&A surfaces that missing angle.

Recommended Resources

As an Amazon Associate, I earn from qualifying purchases.