SO Basics of Neural Networks 2023

Europe/Madrid
Salón de Actos (IAA-CSIC)

Salón de Actos

IAA-CSIC

Description

Deep learning (DL) is a family of techniques widely used in multiple fields with excellent results. Unfortunately, due to its steep learning curve, its use is not as extended as desirable. Several domain-specific libraries have been developed to facilitate the use of these DL models with modest success. For the widespread adoption of these techniques, researchers should be able to design and use their own DL models. Image classification is one of the main problems in astrophysics. The de facto standard to tackle this problem is Convolutional Neural Networks (CNN), a concrete deep learning architecture. A good knowledge of this technique is essential for its proper application by researchers. These three sessions have been designed with a goal in mind: to gain confidence in using CNNs for image classification tasks.

The tutor of this school is Dr Francisco Eduardo Sanchez Karhunen (Universidad de Sevilla).

There will be coffee breaks available to participants.

Contents

  • Session 1: Deep Learning fundamentals
  • Session 2: Convolutional Neural Networks fundamentals
  • Session 3: Practical considerations in real-world CNNs

For more details, see the agenda below.

The repository with contents and materials can be found in Github  https://github.com/iaa-so-training/basic-neural-network-2023

Organizing Committee 

  • Rainer Schödel (Chair), IAA-CSIC, Spain 
  • Laura Darriba, IAA-CSIC, Spain  
  • Javier Moldón, IAA-CSIC, Spain 

Registration

This course will be an in-person event, and it is limited to 25 pariticipants in order to assure adequate interaction and individualised attention. Inscription will be handled on a first come, first serve basis. To formalize the participation you must fill the mandatory fields in the registration form. Payment can be made by bank transfer. 

 

The registration is currently closed because we have reached full capacity for the event. To be added to the waiting list, contact Laura Darriba.

 

The registration fee is 50 euros. It includes two coffee breaks on Thursday and one on Friday. 

 

 

 


This event is supported by the "Center of Excellence Severo Ochoa" award to the Instituto de Astrofísica de Andalucía. We acknowledge financial support from the Severo Ochoa grant CEX2021-001131-S funded by MCIN/AEI/ 10.13039/501100011033 from the Instituto de Astrofísica de Andalucía.

  • Thursday, November 16
    • 9:30 AM 12:00 PM
      Session 1: Deep Learning fundamentals 2h 30m
      • Roots of deep learning techniques. Reasons for layer stacking. Role of weights and activation functions. Layer as a map between representation spaces. Model parameters. Basic structure for classification tasks. Network training as an optimization problem. Weight initialization techniques. Typical loss functions and optimizers. Learning-rate scheduling.

      • Hands-on lab: Build from scratch a basic multilayer neural network using the Tensorflow-2 library. Sequential mode of layer stacking. Model training to tackle a classical basic image classification problem.

      Speaker: Dr Francisco Eduardo Sanchez Karhunen (Universidad de Sevilla)
    • 2:00 PM 6:30 PM
      Session 2: Convolutional Neural Networks fundamentals 4h 30m
      • Drawbacks of classical multilayer networks in image classification tasks. Human brain image handling. Convolutional layers: padding and stride. Types of pooling layer. Kernels for feature map extraction. Kernel stacking. CNNs as an extension of classical stacked layer models. Top layers in CNNs.

      • Hands-on lab. Build from scratch a basic CNN for image classification using the Galaxy10 dataset.

      Speaker: Francisco Eduardo Sanchez Karhunen (Universidad de Sevilla)
  • Friday, November 17
    • 9:30 AM 1:00 PM
      Session 3: Practical considerations in real-world CNNs 3h 30m
      • Overfitting in CNNs. Techniques for overfitting reduction: data augmentation and drop-out. Types of data augmentation. Drop-out rates. Transfer learning: concept and usage. Top Pre-Trained models for image classification. Handling large image datasets: ImageDataGenerators.

      • Hands-on lab: Use of ImageDataGenerators combined with realistic folder structures in image classification problems. Inclusion of data augmentation in our preprocessing pipelines. Add drop-out layers to the CNN design in session 2. Use of transfer learning in a real situation.

      Speaker: Francisco Eduardo Sanchez Karhunen (Universidad de Sevilla)