Editor's Picks

Do You Trust AI? Technology to Explain the Basis of AI-produced Inferences

Published in Nature, an international science journal

AI is rapidly becoming employed in daily life. Many forms of AI presently in use can make judgments about new elements and phenomena by learning a variety of data. However, such judgments are made intuitively like those of the right brain―why specific conclusions have been obtained is inexplicable; the judgment process is a black box. To solve this problem, Fujitsu Laboratories has fused two AI technologies and developed a system to help settle more complex issues.

Importance in Ensuring Wider AI Use―AI Smart Speakers, Automated Driving, and Medical Care

Recently, AI has been becoming more used in people's daily lives. For example, various AI smart speakers were released one after another in 2017. These speakers have built-in microphones to collect voice data. An AI in the cloud receives the data, understands what was said by searching and processing sentence structures, and returns appropriate verbal responses. Thus, AI analyzes speech and varying sentence structures.

In the field of autonomous cars, AI performs fast image processing to recognize objects ahead and determines whether going forward may result in a collision; AI may then reduce the vehicle's speed or change direction.

However, can we really leave decisions on critical matters up to AI? For example, doctors identify the causes of illnesses by analyzing a medical diagnosis, including a physical examination and questions asked to the patient. It is convincing if the doctor explains that the patient is suffering from a certain disease because the patient has developed some distinctive symptoms, and both the cause and optimal treatment are known. If the explanation is unconvincing, the patient will request that the doctor provide a logical explanation based on past experience, papers, and treatments.

Similar situations will continue to occur even after medical AI arrives. The explanation as to why a certain treatment is ideal for a given patient will be unconvincing if AI provides no basis for said treatment.

Machine Learning and Deep Learning Cannot Always Explain the Basis for Response

The AI technologies attracting the most attention recently are machine learning and deep learning.

Machine learning is an approach for learning from an immense amount of data and producing a solution when unknown data is input. Solutions are generated by making optimal decisions using combinations of large amounts of data with known solutions.

Deep learning combines several neural networks, which are models with simple roles inspired from the biological mechanism of the human brain, and attempts to uncover and prioritize connections to find solutions. Deep learning has drawn greater attention in many countries around the world since the successes of image recognition using deep learning in 2012.

Machine learning and deep learning can produce results based on training data even if additional new data is input; thus, they can be likened to the way inspiration works in the right brain. However, the problem is that since solutions are found by intuition, it may be inexplicable why conclusions have been obtained.

Two AI Technologies to Explain Why a Result Was Produced

To spread AI technology's use through society, a mechanism to logically explain the basis for decisions is necessary. Fujitsu has developed a solution that combines two AI technologies: Deep Tensor and Knowledge Graph.

Deep Tensor makes learning more effective by using a representation known as tensors in addition to the deep learning approach explained above.
It draws inferences while providing the reasons it produced specific inferences. For example, if it infers that some kind of cyber attack occurred by analyzing daily communication logs, Deep Tensor shows the inference factors―which parts of the logs imply an attack (such as IP addresses and port numbers); this is essentially equivalent to providing a reason.

Knowledge Graph semi-automatically collects an immense amount of data from the Internet and academic papers; it then organizes such data into a database according to relevance. It uses this database to produce data that may be a basis for inference upon inputting a reason (inference factor).

Combining these two AI technologies enables experts to clearly confirm the reasons and bases for AI-produced inferences. Learning effects and inference accuracy can be enhanced by feeding the results of expert studies into Knowledge Graph.

AI Reduces Experts' Workloads, Shortening Determination Periods in Genomic Medicine from Two Weeks to a Day

An example of applying this AI that can explain the basis is genomic medicine (for cancer treatment). A variety of symptoms may indicate cancer, and there are many kinds of treatments. In the conventional approach of using cancer drugs, doctors diagnose patients and adjust dosages.

The latest genomic medicine helps detect patients' genetic defects that have caused cancer and uses therapeutic drugs that affect cancer cells produced by such genetic defects. In other words, this enables use of therapeutic drugs that accurately affect patients' cancer cells. The downsides of genomic medicine today are that the diagnosis is not covered by Japan's national health insurance, it is very expensive, and it can identify only some causes of cancer; further, effective therapeutic drugs for certain types of genetic defects are not yet available, while other drugs are awaiting approval and thus also unavailable.

In genomic medicine today, a patient's normal and cancerous cells are analyzed with a next-generation sequencer; then, a medical team uses the obtained genetic data to identify a causal gene and determines the recommended treatment. It takes at least two weeks for the medical team to conduct an examination after completing genetic analysis. Unless the cost and time problems are solved, spreading this advantageous genomic medicine far and wide will be difficult.

This is why Fujitsu has worked on genomic medicine that employs this AI technology that can explain the reason. We trained Deep Tensor using 180,000 pieces of disease mutation data, successfully embedding more than 10 billion pieces of knowledge from 17 million medical articles and other materials into Knowledge Graph. Inputting genetic mutation data into this system enables Deep Tensor to infer cancer-causing factors and enables Knowledge Graph to find medical evidence to justify the obtained results. Medical specialists then simply need to review the flow of obtained inference logic, thereby reducing the period between analysis and report submission significantly―to a single day.

It is virtually impossible for any medical team to memorize all the synopses of the immense number of papers available. However, use of AI in genomic medicine enables inferences that encompass past findings.

To summarize, the Fujitsu AI technology that explains the basis for AI-produced inferences will help specialist teams perform time-consuming tasks and significantly reduce both the period and workload.

Currently, Fujitsu is planning to promote this technology in the healthcare, finance, and corporate fields. Going forward, Fujitsu will incorporate this technology into FUJITSU Human Centric AI Zinrai and offer it as a technology that helps experts make decisions.