What are CV ML algorithms

Training goal

Artificial Intelligence - Specialization - Image Processing with Deep Learning (CV)

 

3-day iX workshop (+ optionally bookable 1-day preparation workshop directly before)

 

 

Artificial intelligence (AI) and neural networks are currently revolutionizing the automatic processing of images and videos. The recognition of faces, objects and characters in images and the automatic classification of images allow more efficient processes and new types of applications in the processing of image material. In this course you will learn the basics of using neural networks and deep learning with Tensorflow for image processing and understand and apply the latest developments in this dynamic field.

With this 3-day intensive workshop you enter the exciting world of computer vision (CV) with deep learning. After an introduction, we will deal with the basics of processing image data in neural networks and the application of developed models. You will then learn to understand and apply the latest developments in this area. You will also find out which ready-made software solutions are already available for various CV tasks.

The workshop contains numerous practical exercises with which you can test and deepen the knowledge you have acquired. You will be supported by the speakers. The knowledge is conveyed with clear explanations and exercises, not with large mountains of formulas. Thus, everyone interested has the opportunity to understand how the algorithms work.

After this course, you will not only know how to classify terms such as embeddings, object detection, image segmentation, attention mechanisms, GANs and many more, but you will also understand the underlying methods and be able to apply them in practice.

 

Preparation:

As preparation for the course and to refresh your machine learning and tensorflow knowledge, we recommend our 1-day preparatory workshop, which can also be booked. This takes place directly before the actual course.

If you have no previous knowledge of machine learning and Tensorflow, we recommend our 4-day AI intensive course with many exercises or, for a quick start, our 2-day Tensorflow course with exercises for at home afterwards.

 

Aims:

  • Basics and application of neural networks / deep learning techniques in image processing
  • Advanced deep learning applications in image processing
  • Independent programming with Tensorflow to apply deep learning to image data


 

requirements

  • Python knowledge
  • Basic knowledge of Tensorflow
  • Basic knowledge in the area of ​​machine / deep learning
  • If you need to refresh your knowledge, we recommend our 1-day preparation course, which can also be booked, and which takes place directly before the course
  • If you do not meet the requirements, we recommend our 4-day AI intensive course.
  • All that is required for programming is a common web browser. No additional software is required.
  • For this course we use the open source platform BigBlueButton. All you need is a microphone or headset and an up-to-date browser (Firefox / Chrome).
  • Please use an email address to which you have free access to purchase tickets.

target group

  • Software developers and AI enthusiasts with basic knowledge of Python and machine learning with Tensorflow.

Contents (click on the individual points for more details)

Optional preparation day that can be booked in addition (directly before the actual course)

  • Overview Artificial Intelligence & Machine Learning (ML)
  • Development process of ML models
  • Fast crash course in Python and the Scientific Stack (Numpy, Jupyter, Matplotlib, Pandas)
  • Deep learning basics
  • Deep learning in practice with Tensorflow

Machine vision with deep learning specialization

  • 1) Entry

    • Current exciting application examples

    • Areas of responsibility in image processing (Computer Vision = CV)

    • Human vision vs. computer vision

  • 2) Basics of Convolutional Neural Networks (CNNs)

    • Areas of responsibility and delimitation
    • How do I measure success and train?
    • Detailed explanation of the concept of convolution
    • AlexNet as an application example to get you started
      • Hyperparameters in CNNs: filters, kernel size, padding, stride
      • Properties of CNNs: Receptive Field, Invariances, Feature Maps
      • Other important layers: pooling, dropout, softmax
    • Interactive demo for experimenting yourself
    • Computer vision with Tensorflow and exercises
  • 3) Advanced models for image classification

    • Deeper models and their tricks
    • Stabilization of training - batch normalization and residual connections
    • Adjusting screw model complexity
    • Use and practice of advanced models with Tensorflow
    • Latest model developments and neural architecture search (e.g. Efficient Net)
    • Existing challenges
    • Data augmentation
    • Use of pre-trained models through transfer learning and fine tuning
    • Standards for model transfer between frameworks and optimization for special hardware (e.g. ONNX, Nvidia TensorRT)
    • More exercises with Tensorflow
  • 4) Models for object detection and segmentation

    • Motivation and existing data sets
    • Overview: Tools for labeling with object detection
    • Two main approaches and their origins (One Stage vs Two Stage Detectors)
    • From single to multiple objects
    • Development of the two-stage models (from R-CNN to Faster R-CNN)
    • Various key tricks of these models
    • One-stage models using the example of YOLO (You only look once)
    • Segmentation models and their special features (e.g. MaskRCNN)
    • Existing challenges
    • Creation of a complete object detection pipeline with Tensorflow
  • 5) trends

    • A new approach through similarity / contrastive learning
      • What is Similarity / Contrastive Learning and its advantages
      • Entry into Siamese Networks
      • Measure distances in vector space, various important loss functions and tricks
      • Visualization through UMAP
      • Extension library Tensorflow Addons and use of these
    • Attention in neural networks
      • What is the concept of attention for neural networks
      • Origin from word processing
      • From SE-Net to self-attention
      • Characteristics of attention
      • Application in image processing and possible future
    • Explainability for neural networks and image processing
      • Why is explainability important?
      • What does this mean in the context of ML and DL
      • Levels of explainability
      • Methods, approaches and their differences (e.g. SHAP)
      • Existing solutions in Python
    • GANs - Generative Adversarial Networks
      • Application examples and recent history
      • The opponents: generator vs. discriminator
      • Special about training and the loss functions
      • Special challenges
      • Various extended GAN architectures
  • 6) Outlook

    • Existing solutions
    • Other trends in computer vision

Trainer Philipp Braunhart

Philipp is a freelance machine learning engineer and trainer. At Siemens he developed scalable Artificial Intelligence (A.I.) solutions in the IoT environment. From physics (B.Sc.) he moved on to A.I. in particular deep and machine learning (M.Sc.) developed. Whether it's deep reinforcement learning in robotics, quality predictions in manufacturing processes or anomaly detection in the test results of printed circuit boards: he is always attracted to solving practical problems with the help of intelligent algorithms. With graphic, easily understandable explanations, he helps to overcome other hurdles in order to develop just as much enthusiasm for machine learning.

Trainer Moustapha Karaki

Moustapha is a machine learning engineer and consultant at IBM for the application of AI in companies. Among other things, at Fraunhofer IPK in the research group "Security Technologies", in projects in the automotive industry, in mechanical engineering (B.Sc.) and at IBM, he gained experience in deep and machine learning, computer vision and robotics. This was also the focus of his M.Sc. in Computational Engineering Science. His previous projects include person recognition for safety, product quality optimization based on machine sensors, and intelligent IoT systems in the automotive industry. Through his work in consulting, he became aware of how important it is for people to have an AI-oriented mindset. Spreading this knowledge is his contribution on the way to the coming transformation.

execution

If the event cannot be held due to force majeure, the prevention of a speaker, disruptions at the venue or insufficient number of participants (less than 50%), the participants will be informed by the heise events team at least 7 days in advance.