Deep Learning is currently a buzzword in the world of smart solutions, but what is it, and how can it help create intelligent solutions?
In today’s advancing technological landscape, the terms ‘smart’ and ‘intelligent’ are often overused. References to AI are common, and another technology which regularly crops up is Deep Learning.
Deep Learning is often referenced as the ‘silver bullet’ of intelligent solutions, but different people have conflicting understandings of what this term means. It makes sense for any integrator or end user to have a good understanding of what the technology can offer, in order to ensure their expectations are met.
Deep Learning in a nutshell
Deep Learning is just one element of the growing use of Artifical Intelligence (AI). AI is an over-arching technology which includes several sub-sets. Within the AI arena sits Machine Learning. This is more flexible, and allows systems to be taught certain parameters and criteria which can be used in the decision making process. Deep Learning is a more advanced sub-set of Machine Learning.
It is important to understand this hierarchy, as while all Deep Learning is Machine Learning and AI, not all AI or Machine Learning delivers Deep Learning techniques.
Machine Learning uses large datasets to detect patterns, and tries to predict what might happen, and which actions are recommended to deal with the predicted event. This dataset-based approach is clearly more flexible than using explicit programmed instructions.
What makes Deep Learning different is the use of an artificial neural network with a number of computational layers. This allows data inputs to be processed to assess and collate information, without the need for extensive data sets to be used as a reference. Deep Learning solutions are trained with data inputs until the outputs achieve a high level of accuracy.
As an example, each deep learning level is trained to transform its input. In a vehicle detection application, the video input is passed to the first layer to transform pixels into edges, while the following layer might encode the arrangement of edges. The next layer can identify specific vehicular characteristics, and the process continues until the neural network identifies the object as a car.
As the neural network has been trained to recognise vehicles, it is aware of the core characteristics of various cars, vans, buses, etc.. This training allows it to accurately classify objects.
Because of the number of neural layers used, the process doesn’t impact on the image data. It uses the data, along with the knowledge of vehicles it has been trained to use, to identify them.
Deep Learning at the edge
Hanwha Techwin has introduced 4K Wisenet P AI cameras featuring license-free Deep Learning video analytics which offer a high level of detection accuracy, whilst minimising false alarms. The deep learning analytics detect and classify various objects, including people, vehicles, faces and numberplates, and can identify the attributes of objects or people, such as age group, gender, or the colour of the clothing.