Deep Learning: What it is, concepts and definitions

 


If you follow technology news and market trends, you have certainly heard the buzzword Deep Learning. But what is Deep Learning anyway? To answer this question, we must first get a sense of what Machine Learning is and how it works.

Machine Learning (ML) is an area of artificial intelligence where it is possible to create algorithms to teach a given machine to perform a task. How is this possible? Simple. It is necessary to have a set of data, and from these data, explore the correlation between them, discover patterns, apply algorithms, and generate models that can be generalized to a specific task.

This process is called model training. After having your model trained, it is able to generalize to new data (not presented in the training step), finding correlations and generating predictions for performing the specified task. Generally speaking, ML is divided into two major areas:

  • Supervised Learning: Consists of labeled data. The algorithm receives a set of labeled data, that is, data with the corresponding correct outputs and the algorithm learns by comparing the model output with the expected output, readjusting its parameters until it reaches an acceptable and pre-determined threshold.
  • Unsupervised learning: consists of unlabeled data. The algorithm receives a set of unlabeled data and seeks to find similarities between groups of data, generating clusters, or groups of data. (link to ML article).

Deep Learning – but, after all, where does Deep Learning fit into this universe?

Deep Learning, or Deep Learning, is a sub-area of ​​ML. More specifically, it deals with Artificial Neural Networks, an area that seeks to computationally simulate the brain as a learning machine. The first scientific records of the attempt to reproduce an artificial neuron date back to the 50s, when computational models were developed, but little explored due to the lack of processing power of the computers of the time.

With the development of the digital age, computers became more powerful from a processing point of view, which allowed significant advances in the area. As a Neural Network is a weighted connectionist paradigm, that is, the ability to develop intelligent models lies in the connections of a significant amount of artificial neurons (and not in the neurons themselves), and arithmetic operations grow exponentially with respect to the Deep Networks. It is clear at this point the importance of the Big Data concept since this type of architecture tends to work better with more input data.

Connectionist paradigm of an Artificial Neural Network
Figure 1: Connectionist Paradigm of an Artificial Neural Network

What differs an Artificial Neural Network from a Deep Neural Network is the number of neurons and connections, being significantly higher in the second case. A simple Neural Network has up to 5 layers, whereas a Deep Network has more than 5 layers. A common desktop computer no longer has enough processing capacity to train deep networks, depending on the network architecture used, which generates the need to advance parallel processing techniques and GPU (Graphics Processing Unit). There are companies currently focused on the development of dedicated hardware for graphics processing with applications in Deep Learning (NVIDIA link).

Deep Learning – what are the applications?

Deep Learning can be used in the most diverse applications, but the most common are:

  • Image Processing;
  • Natural Language Processing
  • Applications in Medicine:
    • Image Recognition Applications
    • Breast cancer
    • Alzheimer's disease
    • Cardiovascular Diagnosis
    • Skin cancer
    • Stroke
    • Applications in Drug Development
  • genomics

 Case Deep Learning: Image Recognition

To show the power of Deep Neural Networks, we made a small image recognition case.

Deep Learning – Database

Database items

Columbia Object Image Library 100 (COIL-100). It consists of a base with 7200 images of objects with the presence of rotation, divided into 5400 images to form the training set, and 1800 images for testing, totaling 100 distinct classes.Figure 2: Images present in the dataset

Deep Learning – Network structure and training strategy

A deep neural network was used in the following configuration:

  • Layer 1: Convolution (activation: Rectifier Linear Unit)
  • Layer 2: MaxPooling
  • Layer 3: Convolution (activation: Rectifier Linear Unit)
  • Layer 4: MaxPooling
  • Layer 5: Full Connect (activation: Rectifier Linear Unit, 128 neurons)
  • Layer 6: Full Connect (activation: sigmoid, 100 neurons)

Cost function: Crossentropy; Optimization Algorithm: Adam Small virtualizations of images were created to generate noise in the dataset, such as geometric transformations, random zooms, and small rotations.

Tests and results

After training, a model specialized in the COIL-100 image recognition task was obtained. The accuracy of the model in the training set was 98.64%, and in the test set, it was 98.78%. What's up? Convinced of what Deep Learning is and its power? Its applications are widely varied, and its results promise to revolutionize the way artificial intelligence has influenced decision-making in the modern world.

Post a Comment

Previous Post Next Post