Hello everyone, My name is Kenshin Fujiwara and I am the CEO and founder of HACARUS Inc.
Through this series of blogs, I will discuss a wide range of topics of AI, from the history of AI to practical tips for the successful application of AI projects. I hope my blog posts will help you better understand AI and solve your business issues.
Previously, I introduced sparse modeling and its strong points. I also compared its performance with conventional deep learning methods. From today’s blog, I will share examples of success stories for sparse modeling and how they were introduced into the field. Before introducing the cases of sparse modeling implementation, let’s take a look at the differences between specialized and general-purpose AI.
Specialized & General-Purpose AI
Almost all of the AI currently used in the business world is categorized as specialized AI. This AI is also sometimes called Artificial General Intelligence (AGI). AGI is positioned as an all-purpose AI that thinks on its own and makes predictions or decisions for every situation based on the given data.
In a sense, this can be thought of as the ideal form of AI, which has a thought process similar to that of a human. Since AI development is a complex field, there are many conflicting opinions on the capabilities of AI. While some people think that AGI will be achievable within the next 20 years, others believe that it is still hundreds of years away.
A concept closely related to AGI is the term “singularity.” A technological singularity refers to the point in time when technology and intelligence will have advanced to the point that they can improve themselves through self-feedback. If this happens, some people believe that AI will replace humanity as the main driver of civilization’s progress.
While the pros and cons have swirled around whether such a science fiction reality will really arrive, the influence of Ray Kurzweil, a leading AI research authority who presented this idea, has led many to believe that it will arrive around 2045.
There are several opinions on whether the Singularity will come or not, it is only a matter of time before the “specialized AI” described will permeate various parts of society. With the age of AI quickly approaching, the ability of a company to create jobs that work alongside AI will act as a litmus test for the failure or success of the company.
When aiming for the successful implementation of AI, it is important to choose the right AI for the project. By doing this, companies can overcome many hurdles like issues with the amount of training data, black boxing, and a lack of computational resources.
Therefore, first, I will explain what challenges were faced, how AI aided in solving the problem, and how the challenges were solved in the case of Mitsubishi Tanabe Pharma. All of these examples are from actual projects that were jointly conducted at HACARUS.
Case Study 1: How to Use AI in Drug Discovery R&D – The Case of Mitsubishi Tanabe Pharma
Mitsubishi Tanabe Pharma Corporation is one of Japan’s leading pharmaceutical companies, with approximately 7000 employees. The main focus of its business is the manufacturing and sale of pharmaceutical products.
Dating back to its inception in 1678, the company was founded by Gohei Tanabeya. Currently, the company is engaged in research and development activities with a focus on the central nervous system and immuno-inflammatory fields.
In the drug discovery field, screenings are done to test which compounds are effective as drugs against diseased cells, such as those with cancer. In other words, it is a process of sifting through many compounds to see which ones are effective. This is not an easy task, as there are potentially hundreds of thousands of compounds that need to be tested.
In the past, the actual compound was administered manually to the cells to be tested. The condition of the cells was then visually inspected to determine whether the compound was effective or not.
Challenges with Deep learning Technology Prior to Implementation
When looking at AI that uses deep learning for drug discovery, it has no issues in terms of its functionality. While it is capable of providing accurate screenings, there are still several issues that emerge after its introduction.
- It takes a long time to obtain results
- The AI judgments are hidden behind a black box.
- It requires a huge number of prediction models to build each compound (requiring a large amount of data)
To start, one big issue of deep learning AI is the amount of time it requires to create and implement the model. This is because deep learning needs to create a different prediction model for every compound. This requires a significant amount of data, and it takes 15 to 40 minutes to evaluate the effects of a single drug.
To some people, 15 to 40 minutes might not seem like a long time, but it adds up when you consider that thousands of compounds need to be evaluated and identified. In some cases, this process might take even longer than manual inspection by humans.
Another issue is the lack of interpretability that results from black boxing. When looking at different compounds, the AI shows if any phenomenon occurred, but it failed to explain what mechanisms contributed to this result.
In drug discovery, not knowing the mechanism of how a compound works is a major problem. This is because it is important to be able to interpret and understand the test results when human life and health are on the line.
The third issue is related to predictive models. These models help predict a phenomenon that is likely to occur based on past data. In this case, we need to predict what kind of reaction will occur when a certain compound is administered to a cell and prepare a model that matches this prediction.
Since a predictive model needs to be created for each compound, it requires massive data. The process also demands a lot of computational resources and time.
In the next blog, I will show how the company utilized sparse modeling to solve the issues in AI implementation. For your updated knowledge and insight about AI technology, subscribe to our newsletter or visit HACARUS website https://hacarus.com.
