Machine Learning Model Development

Our Machine Learning and Training Data team develops and implements AI solutions for a wide variety of applications across a range of vertical markets. Our engineers train models to solve several business use cases, including image and sound recognition, recommendation systems, natural language processing, and more. We specialize in creating synthetic datasets and digital twins for training, testing and optimizing machine learning and deep learning algorithms.

Industrial Edge

Industrial Edge processing is an AI deployment strategy where data capture, processing and storage are implemented at or near the source, as opposed to delayed centralized processing. It's often utilized when system inference needs to be near real-time, such as inline manufacturing quality control, logistics and fulfillment centers, and other industrial processes where intelligent automation is needed.

In the example below, an object detection system was developed and optimized within a digital twin (left image) and then deployed and tested physically (right image). The resulting model was then tested and verified in the physical space using an NVIDIA Xavier Jetson NX embedded system and video camera. Because the system was completely developed virtually within the digital twin, physical edge inference worked accurately without additional model modification.

Labeled Cereal Labels on Conveyor
Real Cereal Boxes on Conveyor

Custom Machine Learning Solutions

In the rapidly evolving field of machine learning, oftentimes a specific use case exists where no commercial solution is available. Our machine learning model development team is experienced with developing custom TensorFlow and PyTorch implementations of state-of-the-art models that are not readily available within standard commercial machine learning products. These implementations can be suited to fit your unique data and functionality requirements, from dataset curation to model training and deployment.

The Kinetic Vision machine learning model development process includes an optimization feedback loop that iteratively improves the AI model by automatically adjusting the dataset configuration and model hyperparameters while training. The feedback process is further augmented by AIVision Generate which automatically renders high fidelity, annotated synthetic training datasets at scale. Models are deployed and evaluated in a physically realistic simulation capable of testing rare conditions in a controlled environment.

 

Synthetic Data Generator Data Flow Chart

Our machine learning model development pipeline scales across multiple GPUs and GPU clusters. The distributed training pipeline includes dataset optimization and evaluation. Multi-GPU architectures enable the simultaneous training and evaluation of unique job configurations, which accelerates the iterative experimentation and optimization process.

 

Scalable Optimization Framework Chart

Synthetic Data

Kinetic Vision’s AiVision® synthetic training data development platform utilizes real-time game engine technology and high-fidelity 3D digital assets to create photorealistic imagery and simulations that represent authentic data. The platform enables users to engineer synthetic datasets that meet specific conditions that are impossible to obtain through traditional data collection methods. This expedites our machine learning model development pipeline and allows us to treat the training dataset as a “parameter” that can be tuned and optimized. The training data is not limited to just vision systems, but can also represent infrared, LiDAR, X-ray, and other types of sensor information.

In the example below, synthetic drone and field imagery were utilized in this example to train machine learning systems to classify specific types of military hardware and identify other tactical events.

CRG Vehicle Composite
Segmentation

NVIDIA Partnership

Kinetic Vision is proud to be an authorized NVIDIA Service Delivery Partner. This partnership provides our machine learning model development team with exclusive access to NVIDIA AI software platforms and the respective development teams. This collaboration optimizes system performance, and enables us to quickly train, test, and deploy models that utilize NVIDIA GPU hardware. Our models have been deployed on a wide range of devices, including cloud-based enterprise GPUs and edge devices such as the Jetson Xavier NX. Kinetic Vision has been invited to present AI research at multiple NVIDIA GPU Technology Conferences (GTC). View our presentations at NVIDIA GTC below:   

NVIDIA Preferred Partner Badge