On-Device Training is the training or fine-tuning of ML models directly on edge devices — enabling continuous learning and personalization without sending data to a server, keeping all training data private and adapting the model to local conditions in real time.

On-Device Training Challenges

• Memory: Training requires storing activations for backpropagation — typically 10× more memory than inference.
• Compute: Gradient computation is expensive — MCUs and edge GPUs have limited floating-point throughput.
• Techniques: Sparse updates (freeze most layers, fine-tune only the last few), quantized training, memory-efficient backprop.
• Frameworks: TensorFlow Lite On-Device Training, PaddlePaddle Lite, custom implementations.

Why It Matters

• Personalization: Models adapt to local conditions (specific tool, specific product) without data transmission.
• Privacy: Training data never leaves the device — strongest possible privacy guarantee.
• Continual Adaptation: Models continuously update as conditions change, preventing performance degradation over time.

On-Device Training is learning where the data lives — fine-tuning models directly on edge devices for privacy-preserving, continuous adaptation.