Skip to content Skip to navigation

Deep Learning and Traditional Machine Learning: What is the difference?

Karen Tay

Data Scientist - IBM

Karen Tay is currently a Data Scientist at IBM. She graduated from SMU’s Masters in Information Technology and Business (Analytics track) in 2017. The programme’s strong foundation in predictive, visual and text analytics, as well as its emphasis on practical school projects helps her add business value to the organisation.

The views presented in this article are my own, and do not necessarily represent IBM’s positions, strategies or opinions.

Deep Learning and Traditional Machine Learning: What is the difference?

In my course of work as a data scientist, I found that this role requires a good knowledge of statistical principles and machine learning algorithms, besides demonstrated capabilities in preparing data and building models. The end goal: to discover insights from massive amounts of structured and unstructured data to help shape or meet specific business needs.

Some Machine Learning Algorithms in Data Science

There are different categories of machine learning algorithms and in recent years, there has been an increased adoption of a particular method of machine learning known as Deep Learning. Simply put, deep learning models how variables interact and activate, as inspired by neurons in the human brains. They promise more accurate results compared to traditional machine learning algorithms, especially in situations involving huge amounts of data. One reason for this increased performance is the automatic nature of feature extraction. Feature extraction involves identifying which aspects of data are useful for analysis and modeling.

Deep learning has been widely employed in Natural Language Processing (NLP) as it learns nuanced features, in contrast to feature extraction techniques used in traditional machine learning, which rely on handcrafted rules such as identifying certain entities, parts-of-speech or frequent/infrequent used terms. The difference between deep learning and traditional machine learning is most apparent in Machine Translation, when a source text is automatically converted from one language to another. In the Deep Learning approach, the machine learns to map the input text directly to the associated output text, in an end-to-end fashion.

Another area where Deep Learning has made an impact is in image classification where the advancement of Convolutional Neural Network (CNN) has made revolutionary progress. Instead of manually extracting features (rather like in the approach to texts above), CNN automatically learns the nature of facial features, for example, evolving from low level ones like edge detectors to higher level features such as the eyes and noses, all the way up to the face of a specific person in facial recognition tasks.

The Downside of Deep Learning

In the process of building Deep Learning models, it is hard to interpret which features are necessary or important to include for desired results. This is a downside involving scenarios where end users need to acutely understand the key features included in the analytics. For example, it is important to understand potential contributing factors to how a machine might fail, so as to prevent it.

In building a Deep Learning model, much time is spent on refining the model, a process known as hyperparameter optimisation. Training the model through multiple iterations is also time-consuming. All this leads to increased computational costs and overall costs, which could be a huge deciding factor for a Deep Learning model or other Machine Learning algorithms.

Conclusion

Despite the concerns raised above, experts have commented that performance increases with Deep Learning models that can support and run through a huge amount of data, whereas traditional machine learning models will plateau in their performance.

In my opinion, I do not think that Deep Learning will replace other Machine Learning algorithms entirely, but the choice of whether to use a Deep Learning model or other Machine Learning algorithms would be dependent on and influenced by many factors in a project.

Last updated on 09 Sep 2020 .