Practitioners in Object Detection 

Synoptix specialise as AI practitioners in Object Detection (focussed on Neural Networks). We deliver: 

• Integrated AI systems that effectively interface with existing systems. We are able to effectively integrate with or provide APIs to allow for smooth communication and data handling between systems. We deliver SAPIENT compliant systems and our classification hierarchies can be also be SAPIENT integrated.  

• Edge Processing to operate in a data-connection- and compute-starved environment. Our lightweight models can perform data processing at or near to the source of the data generation. They are optimised for usage on industry-standard edge hardware, and there is no requirement for or dependence upon cloud computing.  

• Learning on the Edge to adapt and improve responses to new situations. Whilst models are pre-trained before deployment, each individual situation that data generators and sensors are deployed in will present different angles of view, lighting conditions, environmental setup, etc. By allowing for adaptive learning on the edge, we optimise our models for operation in their respective deployed environments.  

• Built-In Error Correctors, using our proprietary and patent-pending technologies. Discussed in more detail below, our AI Governors allow for adaptive threshold adjustment to reduce rates of false positives.  

  
  

Operationally-realistic Synthetic Datasets 

Synoptix are experts in building operationally-realistic synthetic environments, that we use to: 

• Generate high-quality training datasets for machine learning models. Our hybrid training workflows integrate both natural and synthetic data to provide optimal balance between availability of data and realism.  

• Simulate operating environments and integrate operational perspectives. We utilise our domain expertise from across the business to introduce realism into the environments that we create. This helps create operational advantage for our customers.  

  
  

In particular, this allows us to create data of “rare events”, which increases the resilience and robustness of Object Detection systems, without the cost, difficulty, and risk of real-world data collection. 

AI Governors 

Through our research with the University of Leicester, we’ve developed our patent-pending AI Governor technology, which allows for on-the-edge adaptive error correction in object detection neural networks.  

Natural environments are inherently stochastic – they’re random and unpredictable. On the other hand, neural networks need to produce defined outcomes – for example, they might produce a classification based on a confidence threshold set within the model.  

By monitoring the distributions of confidence scores within the model, our technology dynamically adjusts these thresholds in real-time, so that it adaptively calibrates the results to minimise the number of false positives.  This delivers a real-world improvement in accuracy and performance of the model, and also allows the model to continue to adapt to operating conditions on the edge.