Deep learning is a form of machine learning that can use either supervised or unsupervised algorithms, or both. It is based on the representation learning (or feature learning) branch of machine learning theory.
At the most simple level, it mimics the human brain in terms of structure. It processes data and creates patterns for use in decision making. But it requires sufficient high quality data for its performance to be accepted. If there is insufficient data, the model can fail quite badly. Test also shows when random noise is introduced, performance remarkably declines.
It really shines when it comes to complex problems such as image classification, natural language processing and speech recognition.
Why is Deep Learning Important?
Deep learning is better than a shallow model because it is able to build or extract better features with fewer parameters. By extracting high-level complex abstractions as data representations through a hierarchical learning process, deep learning models yield results more quickly than standard machine learning processes.
If used, deep learning is highly accurate with the right quality and amount of data.
There are three layers in a deep learning neural network (see diagram below):
Input layer - Holds the data the model will train on. Each neuron in the input layer represents a unique attribute you model will train on
Hidden layer - Sits between the input and output layers and applies an activation function before passing on the results. There are often multiple hidden layers in a network. Each neuron receives input from all the previous layer’s neurons and sends its output to every neuron in the next layer.
Output layer - Final layer in the network. It receives input from the previous hidden layer, optionally applies an activation function, and returns an output representing your model’s prediction.
But there are challenges. Interpretability is one of the biggest challenges. The model is a black box. The computer learns and it goes through many neurons. You cannot know the importance of each feature that goes through the hidden layers and explain the outputs you get from the model.
We can use the example of credit scoring. You will obtain the output whether the customer passes or fails the credit scoring. But you are unable to explain why the customer failed because what goes through the hidden layers of the deep learning neural network is extremely complex.
There are different weights in each neuron. A neuron's input equals the sum of weighted outputs from all neurons in the previous layer. Weight determines the strength of the connection of the neurons. Weightage is a learnable parameter of the deep learning model.
The values of weights before learning starts are initialised randomly. But with a large number of neurons in hidden layers and a large number of hidden layers, the random initialisation may cause problems.
When presented with data during training, weights are adjusted towards values that have correct output. If you have an already trained model and want to re-use it, you can set the values directly.
Weights are important for a model to have the correct output. Unlike linear models, deep neural networks are generally over-parameterized, capable of fitting training datasets to perfect accuracy.
Examples of What is Deep Learning Is Used For
A common example is how Facebook goes through million of images that users post on its site without much of a human intervention. It is able to identify the person in the picture and suggest name tags through image recognition. Or Apple Face Recognition when you use your face to unlock your phone.
Another area is Natural Language Processing (NLP) to recognise and classify accents with high accuracy. NLP is currently used to automatically analyze and represent human language. This is then used in virtual assistants such as Alexa or Siri or Cortana. It can also automatically translate between languages. This is one of the most commonly used area.
Deep learning is also used for chatbots and service bots. The bots handle all conversations by itself and removes the need for customer service team. It does not use any sort of predefined repository and can easily adapt its style.
But the area of deep learning for chatbots is still being perfected, as it can't yet imitate human interactions without mistakes.
Deep learning is very useful for complicated applications. It requires large amount of data and it learns through practice and experience. It can achieve tremendous amounts of accuracy, sometimes surpassing human like classifying objects in images.
Traditional neural networks might only contain 2-3 hidden layers, while deep networks can have as many as 150. Deep learning tools will continue to change the way people work, create and even design products. This is only the beginning.
What are your thoughts on deep learning for your business? Share with us by leaving us a comment. If you require more information or assistance on machine learning or deep learning, contact us. We want to be an extension of our clients. We help systematise process and decision-making. Subscribe to our newsletter for regular feeds.
Did you find this blog post helpful? Share the post! Have feedback or other ideas? We'd love to hear from you.
References
Forbes, What is Deep Learning and How is it Useful?, https://www.forbes.com/sites/kevinmurnane/2016/04/01/what-is-deep-learning-and-how-is-it-useful/#55c80f07d547, published on 1 April 2016
Intellipaat, Importance of Deep Learning and Its Implications, https://intellipaat.com/blog/importance-of-deep-learning-and-its-implications/, published on 20 June 2019
Interesting engineering, What is Deep Learning and Why is It More Relevant than Ever, https://interestingengineering.com/what-is-deep-learning-and-why-is-it-more-relevant-than-ever, published on 27 March 2019
ML Cheatsheet, Neural Network Concepts, https://ml-cheatsheet.readthedocs.io/en/latest/nn_concepts.html