Guide to Machine Learning and Deep Learning

Artificial Intelligence for the Business

McKinsey & Company recently published “An executive’s guide to AI” in which they discuss the business use cases for Machine Learning and Deep Learning. The article provides an informative and concise guide worth discussing in the context of text analytics.

Artificial Intelligence

Before diving into use cases and framing AI for the business, the author defines what is meant by AI. Even more interesting than defining what is and what is not AI is the discussion around why AI has suddenly become relevant now.

Artificial intelligence and artificial neural networks trace their modern methods back to 1943. Despite books, films, and academic discussion around artificial intelligence since then, there has been no real practical use cases or applications. What’s changed?

The environment for AI has come to the forefront at the confluence of advances in algorithms, very large data sets, computing power, and cheaper storage. While McKinsey details many milestones in these areas, the fundamental takeaway is that the components necessary for AI in the business now exist, as does the key use case: large repositories of un- or semi-structured data which likely holds valuable business insights, particularly data generated by consumers.

Machine Learning

Machine learning is probably the level of artificial intelligence most of us are familiar with. Examples of this are automatically routing emails into different categories based on email characteristics (unknown sender) or user actions (move to folder), recommending films based on prior watching habits, or serving up advertising based on a user’s browsing history or other online behavior. Some of these tasks are relatively simple machine learning functions while others can become extremely complex.

There are three major types of machine learning, each of which is explained in detail in the article. Supervised learning involves training an algorithm on labeled input data so the algorithm can be used on new, unlabeled data. Unsupervised learning places an algorithm against input without being guided to any specific output, inferring structure and patterns from the input data itself. Reinforcement learning uses an algorithm in an environment which is unknown and improves over time as it is rewarded for desirable outputs.

Deep Learning

Deep learning requires less human data preparation up front and requires very large data sets. It uses interconnected calculators to form a neural network to deal with very complex data features.

Deep learning works well with unstructured data such as images which do not have usable text and may or may not have usable descriptive metadata.

As with the section on machine learning, the author details each type of deep learning and associated business use cases.

Taxonomy & Text Analytics in AI

The McKinsey overview does a nice job of summarizing and simplifying the complex world of machine learning and deep learning. However, one of the underlying, unstated assumptions is a clean and relevant data set. While images can be difficult to parse, textual language in all of its variants and inherent complexity causes many more challenges.

Machine learning has to parse text and recognize the individual words and phrases within sentences to make sense of the text. While it is a relatively basic process to tokenize text and break it up into words and parts of speech, semantic understanding of the words and how they relate to each other is far more difficult.

For supervised learning, a clean, relevant, and accurately tagged data set is required for training. Using small, user-tagged content which has consistent metadata applied is a basic necessity for supervised learning. Taxonomies with preferred and non-preferred terms and relationships between concepts can act as the foundation for consistent tagging of this content. Once the supervised learning model has been trained to recognize the desired output, new content can be parsed and categorized.

Likewise, text analytics processes can turn un- and semi-structured text into data which is more useful for machine and deep learning algorithms to process.

The road to artificial intelligence includes human input in the form of provided consistency and semantics in taxonomy and text analytics solutions. When a particular viewpoint is assigned to words and text examples, this viewpoint can be recognized and expanded upon by machine and deep learning models.

%d bloggers like this: