deep Learning

Deep Learning is a Machine Learning and AI approach based on Artificial Neural Networks, particularly with the use of Convolutional Neural Networks

Thousands of examples of successful uses in visual understanding, image recognition, object detection and classification / recognition

is an obstacle when using deep learning in applications where large amounts of images and videos need to be processed in real time (or near real time)

deep learning visual

Experience in various CV / ML tasks and models

Deep CNN based solutions
  • Image classification and segmentation.
  •  Object Detection 2D/3D, Tracking, segmentation.
  •  Novelty/anomaly detection.
  • Image enhancement (denoising, super resolution).
  • Generative models.
Frameworks for Deep Learning R&D
  • PyTorch, TensorFlow, Caffe, Caffe2
Frameworks for Deployment
  •  DeepAPI, Caffe, OpenVINO,
    Qualcomm SNPE, TensorRT
Train everywhere and deploy at the Edge
  • Edge is any device that is close to the data!

  • Models need to fit to the edge  device and meet the specificationsv(FPS, power consumption, etc.)

  • Deep CNN Models complexity
    can scale w.r.t. the Edge device
    computing capabilities

Deep Learning Technology

DeepAPI is a software suite consisting of high-performance deep-learning models, optimized for embedded computing systems and SoCs

  • Implements deep learning in embedded devices in order to accelerate adoption of AI
  • It takes full advantage of all computational units including CPU, GPU and DSPs.
  • Supports commonly used architectures SqueezeNet, MobileNet, AlexNet, VGG.
  • Provides an Interface for creating new optimized networks.
  • Compatible with major Deep Learning FrameWorks including Caffe/Caffe2, TensorFlow and PyTorch.
  • Supporting new standard exchange formats (ONNX).
  • Capable of supporting a wide range of applications in end-markets such as IoT, surveillance, machine vision and automotive, and has already been adopted for applications in automation/product identification, medical imaging, and automotive

Deep API – The big picture of our work in DL

DeepAPI Suite automates the way towards efficient deep learning with emphasis on network level and embedded code optimizations 

  • Classic Approaches targeting on optimal code generation following the LLVM compiler Approach.
  • Optimizations can take place in code level and network (model) architecture.
  • IridaLabs presents an orthogonal approach for optimizing model, code and application performance in the long-term by presenting three axes of optimization.

Optimizations take into accound the target platform. Training is performed once by the user using a custom automated tool supporting Caffe/Caffe2, PyTorch and TensorFlow. The optimal model is also available for usage from the major frameworks. No need to access user data – privacy preserving approach.

DeepAPI generates optimal code for standard Networks or compiles a novel network. Optimization at this point is maximum, targeting low power inference.

Network is continuously learning from the target environment. This allows further optimizations in model level as well as in code level.