Next-generation AI implementing a Generative Model

1. What is a “generative model”?

The generative model proposed by researchers such as Mumford[1] and Kawato et al.[2] incorporates frameworks for both “bottom–up” processing, which produces high-level predictions based on low-level inputs, and “top–down” processing, which works from higher to lower levels. Top–down processing simulates outer-world subjects in our inner-world virtual reality, thereby incorporating models for our understanding and knowledge of those subjects, allowing us to improve the accuracy of our predictions, as in Hermann von Helmholtz’s view that “perception is unconscious inference”. Such generative models remain at the research stage (Tajima and Watanabe[3]), however, and are not yet realizable in a practical form.

2. Next-generation AI implementing a generative model

The company’s next-generation AI uses novel technologies for high-level implementation of top–down processing mechanisms to realize a generative model which results in fusion of “data oriented learning” and “logical reasoning”. Taking medical imaging as an example, this approach incorporates biological mechanisms of disease with physical mechanisms that capture lesions through measurement methods such as MRI and CT. At present, skilled physicians supplement their readings of medical diagnostic images with knowledge retained in their heads, but our AI is expected to provide significant contributions as a physicians’ diagnostic aid by incorporating subtleties of information far in excess of human cognitive abilities.

3. Status and outlook for business development

From the features described above, fields in which the Company’s AI is expected to be applied are rapidly growing markets in which high specialized expertise is required, including medicine and healthcare, smart cities, autonomous driving, and space and aviation. The Company has already partnered with Fujitsu Ltd., Nagoya University, and others.

[1] Mumford, David. “On the computational architecture of the neocortex.” Biological cybernetics 66.3 (1992): 241-251.

[2] Kawato, Mitsuo, Hideki Hayakawa, and Toshio Inui. “A forward-inverse optics model of reciprocal connections between visual cortical areas.” Network: Computation in Neural Systems 4.4 (1993): 415-422.

[3] Tajima, Satohiro and Masataka Watanabe. “Acquisition of nonlinear forward optics in generative models: two-stage ‘downside-up’ learning for occluded vision.” Neural Networks 24(2) (2011): 148-158.