Tutoriales

Resumen:

In this presentation we make a brief review of the basic concepts of deep learning and applications. The topics that will be addressed are:

• Representation learning
• Feed-forward deep neural networks.
• Optimization of deep neural networks.
• Applications.

Resumen:

In this tutorial we will learn the basic concepts of a convolutional neuronal network, we will learn to recognize the different architectures, we will approach the most important concepts such as convolutional layers, pooling and a general vision of its operation.

Some of the topics are:

• History.
• Computer vision.
• Convolutional neuronal networks.
• Convolutional layer.
• Pooling layer.
• Fully Connected layer.
• Example of a convolutional neuronal network.
• Convolutional neuronal networks in operation.
• Ship detection.

Resumen:

In many real world applications, it is relatively easy to acquire a large amount of unlabeled data. For example, we have access to thousands of images, speech, audio, books, text and videos, which are available on the web. However, most of this data does not have a label that can help us to make predictions or find groupings. In order to obtain labels for this data, it is required a slow human annotation process and/or expensive devices. Therefore, being able to utilize the plentiful unlabeled data jointly with the scarce labeled data (if available) is desired. How can we develop statistical models that can discover underlying structure, cause, or statistical correlation from the data in unsupervised or semi-supervised way?

In this tutorial we will cover different unsupervised models that are the basis of many of the works we see today in the literature. We will talk about how each model works, giving priority to the mathematical part, as well as areas of application. Among some of the models that we will review, we have:

• Non-probabilistic Models: Autoencoders.
• Probabilistic (Generative) Models: Variational Autoencoders and Generative Adversarial Networks

Resumen:

In this tutorial, we will learn some concepts of deep learning for natural language processing. We will approach the most important concepts such as word embedding and recurrent neural networks.

Some of the topics are:

• Word Embedding.
• Recurrent Neural Network, Long Short Term Memory, Gated Recurrent Unit, Bi-RNN, Recursive Neural Networks, Stack RNN.
• Language Modeling.
• Sentiment Analysis.
• Named Entity Recognition(NER).
• Opinion Mining.
• Attention/self-Attention.
• Machine Translation.
• Question Answering.
• Visual Question Answering.
• Transfer Learning in NLP.

Resumen:

The object detection is achieved through machine learning models that carry out in three steps: feature extraction, training and prediction. Traditionally, an object detection model consists of a trained classifier and a feature extractor. Adaboost, Suport Vector Machines and Neural Networks that only execute the training step, these are feeded with the features extracted from another techniques as LBP, Haar, HOG, etc.

Recently, deep learning techniques have been growing and getting more visibility, the representative tool in deep learning for object detection is the Convolutional Neural Network (CNN) where the feature extraction and the training steps are carried out in one step only. The arquitecture of a CNN have convolutional, subsampling and training layers, one of the main hyperparameters is the activation function, it could be a lineal or no-lineal function, there are many choices for the activation function but, as many applications of object detections, it depends of the object characteristics to detect.

Some arquitectures are altered lightely getting many versions of the initial proposal. RCNN evolves to Fast-RCNN, Faster-RCNN and Mask-RCNN, in the same case you can find YOLO, it evolves to YOLOv2, YOLOv3 and so on. The objective of the modifications is to improve the detection performance.

In this tutorial we will see the main hyperparameters of a deep arquitecture for object detection and some models altered with the goal to get a better performance.

Resumen:

This tutorial gives a brief explanation of the basic of reinforcement learning, and provides a quick overview of Deep Reinforcement Learning. The topics to be covered include :

• Overview of the Reinforcement Learning problem.
• Mathematical formulation.
• Exact solutions using dynamic programming.
• Model-free prediction and control.
• Scaling up with function approximation.
• Policy optimization methods.
• Deep Reinforcement Learning overview and use cases.