DEEP LEARNING

 

NEURAL NETWORKS

Full-stack Neural Network Services

Data Labeling

Neural networks can require vast amounts of training data, but often data does not come labeled or tagged. MulticoreWare has a dedicated team of in-house data analysts to perform any type of labeling task.  Tagging can be performed confidentially and the team can be trained for complex labeling tasks that cannot typically be done by untrained eyes.  We use our own custom tools to perform labeling quickly and accurately and also provide these tools to our customers.

Design & Training

Choosing a neural network architecture for your target hardware and applications and training it is no trivial task.  Our team of 50+ experts can leverage their research knowledge to design and train a network for your task and provide a comprehensive solution using our own proprietary technology or using open-source technology such as Torch7, Caffe, CUDA Convnet, Theano and others.  We can also provide consulting services based around these technologies.

Platform Optimization

The platform for running the prediction path of neural networks can include server-class hardware with accelerators, mobile device SoCs, FPGAs, and custom embedded platforms.  Based on your target platform, compute resources, memory footprint, and mathematical precision must be chosen very carefully.  We’re experienced in optimizing a neural network path for a number of architectures and platforms to improve performance and power efficiency while retaining its accuracy and predictive capability.

Incremental Updates and Maintenance

Updating and maintaining a neural network solution is essential to its continued viability.  We can update and incrementally train a network as more data becomes available to continuously improve its accuracy.  As the target hardware for the network advances, we can also upgrade the network’s predictive capabilities by using additional computer power, expanded memory, and changing power profiles.  We can also deliver solutions that allow the end user to easily add data and incrementally train and improve the system.

Accelerated Training Appliances

We’ve taken Torch7 and Caffe, two of the most popular Neural Network packages, and modified them to improve training speed by an order of magnitude.  Using our combined expertise in neural networks and heterogeneous computing, our versions of Torch7 and Caffe exhibit best-in-class scaling to multi-GPU servers with eight or more GPGPUs.

Hardware Agnostic

Our accelerated training appliances can target Nvidia GPU platforms with CUDA capability or AMD GPU platforms with OpenCL compute capability. This allows you the flexibility to use your existing hardware or purchase the most cost effective solution.

Best In-class Scaling

Accelerate your training to systems with eight or more GPUs to cut down training time by an order of magnitude.  Openly available versions of Torch7 and Caffe only scale to 2-4 GPGPUs but begin to lose parallel efficiency beyond that.  Break through that limit with our custom, accelerated versions of these packages.

Framework Agnostic & Interoperability

MulticoreWare has experience in using, optimizing, and porting the most popular and high performance Neural Network frameworks available for building NN applications. We also develop tools for customers to support interoperability and conversions between the unique Neural Network representations used by each framework.

Applications

Machine Learning and Neural Network techniques can be applied to large range of problems across multiple industries.  We’ve worked on many areas and can bring that experience to your specific problem.

Automotive (ADAS)

  • Traffic Sign Recognition

  • Lane Marker Detection

  • Pedestrian Detection

  • Driver Monitoring

Time/Power Constrained

  • Real-Time Detection

  • Low Power Platforms

  • Pruning & Compression

Video Quality

  • Audio/Video Synchronization

  • Artifact Detection

  • Banding Detection

  • Macro Block Detection

Action Detection

  • Facial Expressions

  • Pose Estimation

  • Gestures

LipSync: Audio-Video Synchronization Detection

MulticoreWare’s LipSync technology uses Deep Neural Networks to “watch” and “listen” to videos and determine if content is correctly synchronized.  Trained classifiers are used to find and match human faces to human speech, allowing LipSync to detect errors that are not found by current automated video quality control systems.