Takeaways - Notes From the AI Frontier MGI (Part I)

[Article 5 of 50. please refer to my first article for some background on this personal project and some standard disclosures. This week, I will summarize a recent discussion paper from McKinsey on their insights on Artificial Intelligence (AI) use cases. Next week, we will take a deeper dive into the expected impact of AI on world economy.]

Here is the LINK to the full discussion paper, and here is the LINK to the summary article.

Introduction

These days we can't go a day without hearing about how AI is changing the world, the term has definitely entered the public conscious through mainstream media. From reports of AI being crowned the king of Jeopardy and Go to countless predictions about how much human jobs/tasks will be replaced by AI system in the next few decades, AI is definitely the buzzword of the moment. In term of Gartner's hype cycle, it's at the "peak of inflated expectation" (see where Deep Learning is on the curve as of 2018). But just like "Blockchains" in the last few years and "Big Data" a few years earlier than that, viewer fatigue will set in, the hype will pass. When that happens, people's attention will shift toward the "next big thing", but their personal and professional lives will continue to be shaped by the development in AI. This discussion paper is an attempt by McKinsey & Company's business and economics research arm, Mckinsey Global Institute (MGI), to understand the current use cases of AI. The context here is important, because technologies around AI are still evolving repidly, new use cases will continue to emerge. Industries not studied in this paper will find nobles ways to utilize AI and bad actors will continue to exploit AI for their personal gains.

Summary

The discussion paper is organized into four parts:

1. Mapping AI techniques to problem types. The authors looked at various analytics techniques and tried to map them to nine different problem types they've identified
2. Insights from use cases. Here the authors looked at the insights they gathered from more than 400 use cases across 9 industries.
3. Sizing the potential value of AI. The authors look at the potential economic value of AI, this section will be discussed in much more details in next week's report
4. Limitations and challenges of AI

Part I - Mapping AI techniques to problem types

It's important to be clear about what the authors mean by "AI" first. The definition is still fluid, given the continuous development in this space. The report's usage of AI is quite disciplined in my opinion: it only referred to deep learning techniques that use artificial neural networks. More specifically, the report only looked at the application and value of three type of deep learning techniques:

1. Feed forward neural network
2. Recurrent neural networks (RNN)
3. Convolutional neural networks (CNN)

The use cases also considered generative adversarial netowrks (GANs) and reinforcement learning, but as they are not yet widely applied in the business world, their impacts are more relevance in the future.

The authors then mapped these AI techniques, along with other more traditional analytics techniques, to nine different problem types they've identified.

Part II - Insights from use cases

The authors collated and analyzed more than 400 use cases across 19 industries and nine business functions. They created a nice heatmap below to show the relevance of different analytics techniques for different industries based on their research.

Coming from the insurance industry, it's good to see how heavily tree-based models, classifiers and regressions are use across the industry (beat Excel for sure!), but the applicability of deep learning techniques are much less. Across the industries, 69% of the opportunites to use AI are in improving performance of existing analytics use cases, only 16% of the use cases are "greenfield" cases where only neural networks can be used. In this sense, there are a lot of potentials for insurance companies to replace existing analytics framework with more advanced AI techniques.

Part III - Sizing the potential value of AI

The authors estimated that AI techniques quoted earlier could generate between $3.5 trillion and $5.8 trillion in additional economic value annually. The table below breaks down the potential impacts for each of the 19 industries. The biggest value opportunities for AI are in marketing and sales and in supply-chain management and manufacturing. (upper right corner of the table). Insurance and banking are more middle of the pack.

The authors then gave examples from three industries: retails, consumer packaged goods and banking, to help understand the biggest value drivers within each industry. The chart below shows the banking example.

Part IV - Limitations and Challenges

The major limitations of AI are technical in nature, such as the need for massive training data sets and the ability to process different types of data. There are also difficulties around model interpretability, generalizing learnings and potential bias in data and algorithms. Unlike the other limitations, which are likely to be solved by technical developments, the bias problem is a multi-faceted problem that's rooted in society norm and human nature, so it takes more effort to consciously overcome these norms and tendencies.

In addition, there are the usual organizational challenges around technology, processes and people that might slow down the adoption of AI. The recent negative news about Facebook and other tech companies also rightfully increased the societal concerns, governments across the global are reviewing or implementing regulations around data privacy and data security.