Continuous training (CT) of artificial intelligence (AI) models refers to the ongoing process of fine tuning pre-trained AI models with new data, allowing them to continually adapt and improve. This approach is used to keep models up to date with the latest information, trends, and patterns. This results in more accurate predictions and decisions from the model, especially as the environment in which it is deployed changes.
In contrast to traditional approaches, where a model is trained once on a fixed dataset, CT allows the model to learn from real-time data streams and evolve to deliver operationally reliable results. This is especially useful in dynamic environments where the underlying data distribution can change rapidly, such as in intrusion and fraud detection.
With intrusion detection, for example, models are trained to detect objects or people within a zone of interest and create an alarm for subsequent action. These models are typically deployed at the edge, processing data from cameras in real-time. Since the environment in which these models are deployed varies from site to site, often several false-positive alarms are created.
Megh’s CT framework allows models to adapt to the environment quickly and dynamically, virtually eliminating false positives.
The framework can be summarized by the following steps:
- Inferencing service: Deploy Nimble Framework with inference engine on-premises to process video frames in real time
- Data collection service: Selected frames with lower confidence are forwarded to our data lake for storage, performance monitoring, and annotation
- Continuous training: Either retrain existing or develop new models on-demand using the data lake and evaluate historical performance for potential deployment
- Model staging: Stage new models for deployment and evaluate their performance on real-time data vs. existing models
The inference and data collection services are implemented at the edge, while the model repository, training, and registry are implemented in the cloud. The Megh CT frontend is designed to simplify the process of image review and annotation, automating the majority of the cloud data management process. Megh’s “out-of-the-box” model repository is trained on large proprietary datasets specifically designed for video surveillance use cases. The continuous training framework then builds on these models using the data stream from the deployed site to construct a model that represents the target environment. Our models are trained with a slight skew to reduce false negatives, ensuring that all potential events are detected.
There are key benefits to using Megh’s CT framework:
- Fast Modeling: Optimized ML-ops pipeline implemented with data collected on site and trained in the cloud, resulting in new models in four to six weeks
- Accuracy: Training data collected from inferencing service dramatically improves model accuracy and enables continuous monitoring to ensure robust model performance
- Efficiency: Models implemented at the edge for inferencing using a range of hardware architectures: CPUs, GPUs, FPGAs, or SOCs
The framework enables Megh to support a comprehensive range of use-case classes for its partners and end users:
- Standard use cases are out-of-the-box scenarios using pre-trained models (e.g., intrusion and loiter detection for physical security, monitoring areas for worker safety, vehicle tracking and counting for traffic management, etc.). Megh’s CT framework enables continuous optimization of these models, resulting in virtual elimination of false positives.
- Advanced use cases are scenarios that need model updates (e.g., PPE compliance for worker safety, proximity tracking for physical security, vehicle access control for traffic management, etc.). These uses cases start with pre-trained models, but require updates based on deployment environments. The CT framework quickly updates the models to deliver operationally reliable results.
- Custom use cases are scenarios that need unique models (e.g., defect inspection for operational efficiency, and box and pallet counting or shelf restocking for inventory management). The CT framework can be deployed on site to collect data for training of new models to solve business problems and deliver compelling ROI.
Continuous training is a key feature of Megh’s Open Analytics platform, allowing it to fulfill the promise of AI for intelligent video analytics.
