Type to SearchView Tags

Machine Learning and How it Differs from AI & Deep Learning
Kamal Anand Senior Enterprise Architect, Digital & Analytics | June 25, 2020

Artificial intelligence (AI), deep learning (DL), and machine learning (ML) are some of most common buzzwords that we hear and use in today’s discourse around technology. But do these terms describe the same thing? Simple answer is no! These terms are definitely related to one-another, but they have different meanings and are applied differently. Machine Learning is one of the branches of artificial intelligence whereas deep learning is subset of machine learning. So, what are these terms and how are they related. Figure 1 below illustrates this relationship.

Artificial intelligence (AI)

Artificial intelligence are machines that can act autonomously. AI mimics the cognitive functions of humans such as learning, problem-solving, reasoning etc. Programs that have reasoning, that can act or adapt and those which have sense. The purpose of AI is to replicate a human brain, the way humans think, act and function. It is synonymous to human intelligence incorporated in machines.

Are AI, ML, and DL the same? No. They overlap due to data but function independently with different applications

Machine learning is a way of using algorithms to use the data provided, learn from it and derive predictive insights which can help in making repeated decisions. Its purpose is to analyze, understand, and identify patterns in the data. Data is key for having successful ML models, the more suitable data we have the better we can train and improve our model. We need to identify when there is enough data being used for ML algorithms to be carried forward to a successful ML project. In the project we need to train, deploy, and evaluate the model and if there is not enough data then modeling results will not meet our expectations and thus model will not predict with a desired degree of accuracy. ML algorithms are trained on data and their performance improves over time as more and more data is incorporated into these models to improve predictability. Training a model means attempting to optimize based on the features of data and trying to predict the data which would be very close to actual value.

ML focuses on the ability of machines to receive data, learn by themselves, and change the algorithm by receiving and processing more data over time whereas AI refers to a machine that carry out tasks based on algorithms in a smart and intelligent manner.

Deep learning mimics the network of neurons in a brain. It is a subset of ML application that functions in almost the same way but has different capabilities. Data in a deep Learning model is analyzed continuously similar to functioning of a human brain i.e. with logic structure which will in turn draw conclusions. Algorithms in deep learning are trained in identifying the patterns and then classifying this information to provide the output basis the input it receives. Output accuracy in deep learning is dependent on data that is inputted.

Most of the analytics leveraged for business purposes are usually based on historical patterns as it looks at historical data for metrics calculations and identifying trends. However, can businesses make predictive decisions by leveraging data? This is where machine learning comes into play.

Solving business problems using ML application

Every organization faces business challenges but applying ML models for every problem is not the answer. A few things that need to be considered before using ML models are:

  • Problem complexity or difficulty: - First thing to consider is whether problem is complex or difficult and does any prediction or decision we are trying to make warrants ML. The goal to transform the business by using ML should be challenging but not impossible or simple. If a set of rules or if-then scenarios are able to handle the problem, then there is no need for ML. Similarly, if the expectation from the goal is almost impossible, for e.g. how perfect product can be delivered every time, then ML would also not be able to help.

  • Specificity: - If complexity or difficulty of a business goal is too specific or open-ended,  then ML may not be the right solution
  • Data: - ML depends on large amount of data but it’s also about accuracy to avoid invalid predictions. Similarly, if we have historical data from years ago but which is not of relevance today due to business process changes or mapping to current trends in business, then ML usage is irrelevant. Furthermore, it is important that the data is labeled so that machine understands and make sense of it.
  • Allowance of error: - Machine learning application can be incorrect sometimes as algorithms learn in the same way we do. As more and more data is incorporated, the model can keep on learning and deriving insights but if errors are not allowed for the particular task then ML is not right for that scenario
  • Biases: - Algorithms may have biases in trained data and we should avoid creating or reinforcing unfair bias. The encouraging part is that ML algorithms improve over time as models continue training and can provide more consistent answers with even with the same data.

Machine learning models are categorized into:

  1. Supervised learning: - The machine in supervised learning is trained by using the data which is ‘labeled’ i.e. tagging is already performed on the data. An algorithm in supervised learning learns from the training data which is labeled and using this data it helps to predict an outcome for unforeseen data and it uses offline analysis. For e.g. time of day, weather conditions etc.

  2. Unsupervised learning: - In unsupervised learning model works on its own and discovers the information. Data in unsupervised learning is ‘unlabeled’. Unknown patterns in data can be found using unsupervised learning and it uses real-time analysis. Anomaly detection is an example of unsupervised learning.

  3. Reinforcement learning: - Algorithms in reinforcement learning are trained using system of ‘rewards’ and ‘punishments'. There is no pre-defined data and models are trained to make decision sequentially. Reinforcement learning follows a trial and error methodology. A simple example of reinforcement learning is that when you applaud or give hi-five to a child when he/she identifies an alphabet character correctly or make a sad face when he/she identifies an alphabet character incorrectly. Over a period of time, the child begins to understand the meaning of facial expressions/actions, and this is how reinforced learning problem is described.

Now where does data science fit in all of the above. Data science is everything that is related to data whether it is cleansing of data, preparation of data or analysis of data. Data science is being used to make business decisions and the more data we have the more business insights can be generated. Patterns can be uncovered that we may not even have known to exist by using data science and recommendation engines are being build using data science. As AI, ML, DL and data science concepts deal with data so there is an overlap between all of them and Figure 2 below illustrates that:

Artificial intelligence (AI)

To conclude, AI is used by machines to carry out smart tasks by exhibiting the intelligence of humans. And ML is a subset of AI while DL is a subset of ML.