• Columns
  • Realizing AGI with Machine Intelligence: Increased Generalization of Current AI has Greater Immediate Effect than Building AGI
COLUMNS Advanced Technology Megatrends Artificial General Intelligence
  • URLをコピー Copy
Artificial General Intelligence Part 2
A world where AI provides humans with close support through super-human intellect

Realizing AGI with Machine Intelligence: Increased Generalization of Current AI has Greater Immediate Effect than Building AGI

Tsuyoshi Inoue, Science and Safety Division 17 April 2020

While the definitions of "machine intelligence" are still varied, this article uses the term to describe the form intellectual capability created by emulating human intelligence and decision-making through predictive models using machine learning technology. Machine intelligence is a representative application of narrow artificial intelligence (AI), discussed in Part 1. Machine learning has recently seen remarkable advancements. If the current trend in development continues, there are projections for machine learning to well exceed the intellectual capacity of humans in the near future. Even more conservative forecasts in which machine learning does not make such progress hold that computers will come to substitute a large portion of human work. On the other hand, many experts believe that the realization of artificial general intelligence (AGI) as an extension of the current machine learning technology will be achieved only in the distant future. This article will highlight the reasons for such forecasts and important points for consideration when applying AGI to business.

Capabilities Required to Develop AGI Using Machine Learning

Research is being conducted based on various concepts to discern the capabilities required to develop AGI using machine learning. Research by GoodAI*1, a Czech startup, and research by the New York University*2 offer two prominent examples. Based on the concepts from these two projects, table 1 summarizes the capabilities required to realize AGI using machine learning technology. Each capability is explained in comparison to human capabilities.

Table 1

Five Capabilities Required for Realization of AGI Using Machine Learning

Capability 1 Learning with limited data
Capability 2 Solving various real-world problems
Capability 3 Accumulating and utilizing knowledge and experience
Capability 4 Searching for appropriate information sources to accumulate knowledge
Capability 5 Setting hypotheses, testing by trial and error, and obtaining answers

Source: Mitsubishi Research Institute, Inc.

Capability 1: Learning with limited data

Humans are able to identify animals rarely seen in their daily lives (e.g. lions, elephants, hippopotamuses, and zebras) with a high degree of accuracy by looking at just a few photographs. Such accuracy is not found through thousands of photographs*3 seen as a child, but it is instead the result of learning with limited data, or in this case a small number of photographs.

Capability 2: Solving various real-world problems

Humans are able to form an understanding of the type of animal through not only photographs but also picture-book illustrations and textual descriptions. Further, humans can link these distinctively different modes of expression to identify specific animals.

Capability 3: Accumulating and utilizing knowledge and experience

Humans are able to apply the skills and knowledge acquired through education and training to solve problems in their day-to-day lives. AI too can learn using the massive amounts of educational materials for humans that already exist and acquire an understanding of highly abstract concepts.

Capability 4: Searching for appropriate information sources to accumulate knowledge

Humans are able to search for information sources based on their awareness of a problem and use the newly acquired knowledge to solve problems all without receiving instruction from others. Although similar to Capability 3, this capability is considered more sophisticated. AI too will be able to gather massive amounts of information from the Internet, develop an understanding of that information, and solve problems.

Capability 5: Setting hypotheses, testing by trial and error, and obtaining answers

Humans are able to solve complex, real-world problems by establishing and testing hypotheses. AI too will be able to conduct work requiring creativity, such as research activities and development of new products.

Current Situation of Development of AGI Using Machine Learning

The following describes the degree to which current machine-learning technologies have achieved the Five Capabilities required for the realization of AGI.

Capability 1

Current forms of machine learning require massive amounts of training data. One major challenge in the development of machine learning today is achievement of the ability for a machine to learn with a small amount of data. Various methods exist that have seen a certain degree of efficacy in increasing the efficiency of learning by reducing the amount of data required. However, current technologies are still far from achieving levels comparable to humans.

Capability 2

Current forms of machine learning demonstrate a degree of flexibility in adapting to real-world problems of a relatively small scale. For example, with photographs AI can properly discern multiple variations of the same target object, daytime and nighttime settings, and differences in shooting angle. However, the flexibility of AI understanding is still far behind that of humans.

Capability 3–5

Research and development for capabilities 3 through 5 is still in a stage of infancy. Several approaches have been proposed by researchers, but none have proved conclusive. One such approach is called Lifelong Learning*4 and is being investigated by the US Defense Advanced Research Projects Agency (DARPA). Lifelong Learning is a form of machine learning that incorporates the memory mechanisms in Capability 3 and the curiosity mechanisms in Capability 4. However, this approach is still limited to the aspects of Capabilities 3 through 5 that are relatively easy to apply to machine learning. Concurrently, figure 1 describes the shortcomings of current technology for future development.

Figure 1

Development of AGI Using Machine Learning: Shortcomings of current technology

Source: Mitsubishi Research Institute, Inc.

Business Impact of AGI Research & Development

The realization of AGI using machine learning technology is still in the early stages of development and without definite technical leads. However, the application of AGI-related technologies to business would have an enormous impact despite the Five Capabilities having only been marginally achieved. For example, the development of a highly efficient method for learning with limited data, or Capability 1, will dramatically reduce costs required for labeling data and result in immediate business benefits. Major corporations have offered AGI-oriented businesses such as Google Brain and OpenAI with substantial funding to develop methods for the application of AGI-related technologies.

While AGI using machine intelligence is still far from realization, potential is growing for the application of machine learning technologies to business. To succeed in developing businesses based on machine learning, corporations should focus on forging and utilizing technologies that maximize the versatility of machine learning capabilities.

  • *1: GoodAI “Roadmap”
    https: //www.goodai.com/roadmap/ (Accessed: December 22, 2019)
  • *2: Lake et al. (2017) “Building machines that learn and think like people” Behavioral and Brain Sciences (Vol. 40, e253)
  • *3: In case of deep learning, which is a kind of machine learning method, thousands of learning data are required for image recognition
  • *4: Senator, T. “Lifelong Learning Machines (L2M)” DEFENSE ADVANCED RESEARCH PROJECTS AGENCY
    https://www.darpa.mil/program/lifelong-learning-machines (Accessed: December 22, 2019)
  • URLをコピー Copy