Four Characteristics of Wide Learning™

Social Implementation of AI Has Revealed its Problems

AI technology that is currently in the spotlight is deep learning, which reads huge amounts of data into a high-performance computer to learn. Deep learning simulates the function of the neural networks of a brain and can recognize images and sounds with high accuracy.

However, as AI continues to be implemented in society, some tricky aspects of deep learning have become apparent: "there are no clear reasons for judgements", "huge amount of data are needed to make the right judgements", and "it often involves high hardware costs, requiring high performance parallel computers for real-time processing".

When compared to deep learning, Wide Learning™ has four distinct features:.

  1. Reasons for judgements can be explained
  2. All Important Hypotheses are Found
  3. Small amounts of data can be used to make highly accurate judgements
  4. Action plans can be computed

1. Reasons for Judgements Can Be Explained

The first problem with deep learning is that it is a black box (*1), so the reasoning behind judgements is unclear.

If people do not understand the reasoning behind the AI's conclusions, they may feel intimidated when medical diagnostics or financial investments are involved.

Wide Learning™, on the other hand, uses hypotheses that can be interpreted by humans to derive its answers, so it is easy to understand and explain the intermediate computation process and the results of final judgements both logically and objectively.

Due to these features, Wide Learning™ is positioned as one of the XAI (Explainable AI) technologies which are the opposite of black-box AI technologies.

*1: A black box is a device or tool which can only be viewed in terms of its output results, without any knowledge of its internal workings.


2. All Important Hypotheses Are Found

So why can Wide Learning™ be explainable?

The reason is that in the hypothesis space for the input data, all the hypotheses are tested comprehensively, and important hypotheses (We refer to them as knowledge chunks) are discovered without fail.

Also, when there is no important hypothesis in the hypothesis space, we can prove "No" in a strict sense.

Fujitsu Research Developed Ultra-fast Combinatorial Computing Technology

Here, the hypothesis space that Wide Learning™ handles is any combinations of data items in the input data.

To validate all combinations of data items, you must efficiently enumerate astronomical numbers of combinations.

Fujitsu Research developed ultra-fast combinatorial computing technologies while working on discovery science (*2) research for many years.

*2: Discovery science is a research field proposed in the 1990s by Professor Setsuo Arikawa (currently a professor emeritus) of Kyushu University.
He has developed theoretical and practical research to discover hypotheses and knowledge using computers.


3. Small Amounts of Data Can Be Used to Make
Highly Accurate Judgements

The second problem with deep learning is that the computer needs a huge amount of data to make the correct judgements.

For example, it is said that Deep Learning requires at least one to tens of thousands of images and photos in order to identify animals, so if you can't provide this huge amount of data, you may not be satisfied with the results that Deep Learning will provide.

As mentioned earlier, the hypothesis space that Wide Learning™ explores and tests is any combinations of the data items in the input data.

Therefore, even if a large amount of data for learning cannot be prepared, a sufficiently large hypothesis space can be constructed from even a small amount of data, and important hypotheses can be comprehensively discovered. This feature is not available in Deep Learning.

For example, let's say you want to analyze the cause of defective products on a factory production line. However, in general, defective products rarely occur, so we find the problem that "it is difficult to collect a large amount of data when a defective product occurs".

However, even with only a few dozen or a few hundred pieces of data, Wide Learning™ can discover important hypotheses that apply only to defective products and start analyzing the causes of defective products.

[Note] The following comparisons are for people who have experience with AI, so if you're interested, please read through them.
Comparison of Wide Learning™ and Deep Learning
Comparison of Wide Learning™ and Other Major AI technologies


4. Action Plans Can Be Computed

The final feature is that based on the results of Wide Learning™, you are provided with actual actions you can perform.

As we've explained many times, Wide Learning™ performs comprehensive testing of all the hypotheses in the hypothesis space.

Using this information, for example, in digital marketing, it can compute the difference between "hypotheses about low purchasing rates" and "hypotheses about high purchasing rates" and present the most significant difference (= The biggest increase in the purchase rate) as an "action plan".

Experiments conducted by Fujitsu's marketing department have confirmed that action plans proposed by Wide Learning™ have higher customer coverage and higher average purchase expectations than those developed by marketing professionals.

In this way, Wide Learning™ can be a powerful tool for supporting policy planning and decision making in the field.

Summary: Comparison of Wide Learning™ and Deep Learning

Finally, the following table summarizes the differences between Wide Learning™ and deep Learning (Some points that we could not cover in this article are also included).

There are advantages and disadvantages due to the differing operating principles, but there are many situations where Wide Learning™ is easier to use.

  Wide Learning™ Deep Learning
Operating principle Process of scientific discovery Simulation of a neural network
Suitable data type Tabular data Images and sound
Amount of data From several dozen to several hundred records At least one thousand to several tens of thousands of records
Explainability XAI (explainable AI) Black box
Output Classifications, forecasting, and action plans Classifications and forecasting
Hardware requirements Computer with a generic CPU (Even a notebook can be used) Parallel computers such as GPUs

[Note] The following is a comparison between Wide Learning™ and Deep Learning, so if you are interested, please read through it.
Comparison of Wide Learning™ and Deep Learning


Let's Try Using the Trial Tool

This site provides the Wide Learning™ "Trial Tool".

Please try "Wide Learning™", AI inspired by the process of scientific discovery!