Blog Post
Understanding the difference between Deep Learning and Machine Learning
If you think that deep learning (DL) and machine learning (ML) have a lot in common, you’re right. But if you hear someone using deep learning as a synonym for machine learning, it’s not quite as accurate.
Machine learning uses past data and statistical algorithms to create systems that recognize patterns and predict future observations. Deep learning is a machine learning technique that draws its inspiration from the human brain and how it thinks and extracts information.
In other words, deep learning is a subset of machine learning, but not vice versa. Both are types of artificial intelligence.
Machine learning solutions can automatically adapt over time based on the new and natural patterns within the data, rather than having to programmatically account for every possibility with logic and routing. There are 4 types of machine learning:
The most common form, supervised learning, refers to algorithms that learn from data labeled by humans. The most common supervised machine learning techniques include linear regression, decision trees or random forests.
For ML algorithms to make decisions, predict something or recognize a pattern, data scientists have to train them with properly collected, cleaned, engineered and labeled data. Based on a data team’s data pipeline, model accessibility and the model’s additional exposure to more data, the effectiveness of the algorithm can continue to improve.
Ultimately, a trained machine learning model is composed of the parameters that inform how much each variable affects the predicted value (for example, when predicting home prices, how much does the number of baths sway a home price).
In other words, the higher-quality the data the ML algorithm gets, and the more useful the features engineered within the dataset, the better, and more accurate the output will be.
Machine learning is the parent category for artificial intelligence methodologies that include deep learning, so the principles by which machine learning works give some understanding as to how all forms of artificial intelligence products work.
Machine learning relies on examples of past experiences in the form of data to offer predictions with varying levels of confidence in the future. If an experience in the future very closely resembles an example seen by a ML model in the past, it can likely predict an outcome (also referred to as a target) with fair confidence. That space in ML with an object to predict is called supervised learning.
Machine learning and deep learning derives results on cleaned data sets. In some more detail, now, supervised machine learning works by finding a function (just a combination of inputs and tuning parameters) to fit a dataset where the experiences (data) have a label.
The right parameters for the function are the ones that minimize the function’s error rate against that particular set of data; you can think of the error as the distance between a prediction and the actual value. The process of creating that function with minimized errors is also called training.
As an example, if a data scientist is trying to train a model to predict home prices, they may use a linear regression model whose parameters like the number of baths, the square footage of the house, and the proximity to a mass transit center affect the label: the home price.
By comparison, when machine learning isn't deployed to help predict an outcome, it’s usually used for identification or segmentation. Those approaches that lack a variable, or a target to predict, are a part of the sphere of unsupervised learning. That brand of machine learning works by grouping similar data points together during training and then determining which group a new and previously unseen sample should fall into.
As an example of unsupervised ML, if a data scientist is trying to train a model for anomaly detection, they may use a clustering model (like K-Means) to see which data points make sense to be clumped together. If there are data points that don’t fit into a cluster, those are the anomalies.
Deep learning works to recognize patterns and relationships in data and is often used for processing data formats like documents, photos, videos or audio.
“ Deep” is to indicate that learning happens in several layers. Say, the first layer learns to identify an “orange” and its basic features in an image. By going through the next layers, the model may add information about more features that make an orange an orange (texture, color, shape, etc). Each new layer has more information about the features of the orange based on the previous layer's knowledge so it can better detect the fruit and differentiate it, for example, from an orange ball.
How accurate it is at categorizing each item it sees (the orange), and therefore how quickly it’s learning is determined by comparing the predicted value to the correct value. Comparing the two values and then feeding the results back into the training process gives the neural network the chance to experiment with another set of weights and biases.
For another scenario, say you’d like to build an ML algorithm that can distinguish between an image of a cookie and an image of a dog. To do this, you’d need a lot of training data, in this case, images that present either a dog or a cookie and are labeled appropriately.
By training the ML algorithm on what features have dogs and what features belong to cookies, the algorithm learns to recognize whether it’s a dog or a cookie and uses this information to predict the correct label for the new image.
The truth is, your use cases should dictate whether you’ll use deep learning or simpler machine learning models. More traditional machine learning models are comparatively cheap and easy to train and deploy, while deep learning models help automate use cases for data types like images.
Say you have well-structured, clean numerical data and you’d like to predict customer churn in your insurance company or classify your customers and their lifetime value. In this case, building and training a simpler machine learning model like a logistic regression is a better choice.
But if you’re dealing with unstructured data types and need to do image recognition, deep learning will do the job here. Say you’re giving a DL algorithm an image of a dog but you’re not telling the model what the picture presents. Here, the neural network will recognize, step-by-step, features of a dog, until it classifies the image as a dog image.
Ushur’s Customer Experience Automation (CXA) Platform is designed to support the operations of complex businesses in the insurance, healthcare and financial industries. Ushur and Ushur AI Labs have already thought through the decisions an enterprise business needs to consider when evaluating AI projects in highly regulated industries. Using proprietary language and document services, Ushur’s AI solutions help businesses understand user intent, evaluate document content and drive seamless customer experiences so that customer support functions and business owners can focus on the highest-value projects in their queues.
If you want more information on the machine learning and deep learning capabilities in the Ushur platform, and to see how Ushur blends AI with customer experience technologies, visit ushur.com/platform.
