COVER STORY u ANALOG DEVICES
Convolutional Neural Networks: What Is Machine Learning?
The world of artificial intelligence (AI) is rapidly evolving, and AI is increasingly enabling applications that were previously unattainable or very difficult to implement, says Ole Dreessen, staff engineer, field applications, Analog Devices
T
his article explains convolutional neural networks (CNNs) and their significance in machine learning within
AI systems. CNNs are powerful tools for extracting features from complex data. This includes, for example, complex pattern recognition in audio signals or images. This article discusses the advantages of CNNs vs. classic linear programming.
WHAT ARE CONVOLUTIONAL NEURAL NETWORKS? Neural networks are systems, or structures of neurons, that enable AI to better understand data, allowing it to solve complex problems. While there are numerous network types, this article will solely focus on convolutional neural networks (CNNs). The main application areas for CNNs are pattern recognition and classification of objects contained in input data. CNNs are a type of artificial neural network used in deep learning. Such networks are composed of an input layer, several convolutional layers, and an output layer. The convolutional layers are the most important components, as they use a unique set of weights and filters that allow the network to extract features from the input data. Data can come in many different forms, such as images, audio, and text. This feature extraction process enables the CNN to identify patterns in the data. By extracting features from data, CNNs enable engineers to create more effective and efficient applications. To better understand CNNs, we will first discuss classic linear programming.
10 October 2023 Irish Manufacturing
LINEAR PROGRAM EXECUTION IN CLASSIC CONTROL ENGINEERING In control engineering, the task lies in reading data from one or more sensors, processing it, responding to it according to rules, and displaying or forwarding the results. For example, a temperature regulator measures temperature every second through a microcontroller unit (MCU) that reads the data from the temperature sensor. The values derived from the sensor serve as input data for the closed-loop control system and are compared with the setpoint temperature in a loop.
This is an example of a linear execution, run
by the MCU. This technique delivers conclusive outcomes based on a set of preprogrammed and actual values. In contrast, probabilities play a role in the operation of AI systems.
COMPLEX PATTERN AND SIGNAL PROCESSING There are also numerous applications that work with input data that first must be interpreted by a pattern recognition system. Pattern recognition can be applied to different data structures. In our examples, we restrict ourselves to one- and two-dimensional data structures.
Some examples are as follows: audio signals,
electrocardiograms (ECGs), photoplethysmographs (PPGs), vibrations for one-dimensional data and images, thermal images, and waterfall charts for two-dimensional data.
In pattern recognition used for the cases mentioned, conversion of the application in
classic code for the MCU is extremely difficult. An example is the recognition of an object (for example, a cat) in an image. In this case, it doesn’t make a difference if the image to be analysed is from an earlier recording or one just read by a camera sensor. The analysis software performs a rules-based search for patterns that can be attributed to those of a cat: the typical pointed ears, the triangular nose, or the whiskers. If these features can be recognised in the image, the software reports a cat find. Some questions arise here: What would the pattern recognition system do if the cat is only shown from the back? What would happen if it didn’t have any whiskers or lost its legs in an accident? Despite the unlikelihood of these exceptions, the pattern recognition code would have to check a large number of additional rules covering all possible anomalies. Even in our simple example, the rules set by the software would quickly become extensive.
HOW MACHINE LEARNING REPLACES CLASSIC RULES The idea behind AI is to mimic human learning on a small scale. Instead of formulating a large number of if-then rules, we model a universal pattern recognition machine. The key difference between the two approaches is that AI, in contrast to a set of rules, does not deliver a clear result. Instead of reporting “I recognised a cat in the image,” machine learning yields the result “There is a 97.5% probability that the image shows a cat. It could also be a leopard (2.1%) or a tiger (0.4%).” This means that the developer of such an application must decide at the end of the
www.irish-manufacturing.com
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50
