TRON Pilot Experiment at Gemma 4 Kaggle Hackathon 2026
Gemma 4 Kaggle Hackathon 2026 took place from April 3 to May 19, 2026. The goal was to create solutions based on Gemma 4 multimodal models to address real-world global challenges. TRON startup participated in the Global Resilience direction - building systems for disaster prediction and response. 1,610 teams participated out of 14,523 registered participants. TRON technology author, Boris D. Yarovoy competed as a solo team of one person, writing all the software code for the pilot experiment independently.
Hypothesis
Multimodal neural networks, such as Gemma 4 from Google DeepMind, can detect changes in animal behavior by analyzing video recordings. This enables the transition from manual animal observation by humans to a fully automated system.
Why Gemma 4
The multimodal Gemma 4 model is capable of working with video in zero-shot mode (without fine-tuning). This allows overcoming the limitation of lacking labeled datasets of animal behavior before earthquakes.
How the Pilot Experiment Worked
The pipeline tested on short videos:
Video → ffmpeg (frames) → OpenCV → Gemma 4 → JSON verdict
Test Results
- 🐘 Elephants (earthquake, YouTube) → 🔴 ANOMALY (high confidence)
- 🐱 Cat (normal, YouTube) → 🟢 NORMAL (confidence 1.0)
- 🐻 Bear (earthquake, YouTube) → 🔴 ANOMALY (confidence 1.0)
Result: 3 out of 3 correct detections (100%).
The pilot correctly identified the anomaly in the bear's behavior even with strong camera shaking during the earthquake.
Pilot Limitations
- Small test sample (3 videos) - proof of concept only.
- No real-time streaming - only pre-recorded videos.
- Missing analytical layer correlating behavioral anomalies with seismic data.
Conclusion
This pilot experiment demonstrates that multimodal models can correctly classify animal behavior as normal or anomalous in zero-shot mode, without fine-tuning, across any species.