Generating Internal Representations through Interactions with the Real-World Environments

Опубликовано 6 сентября 2016, 16:42

Human manually written programs do many (clean) tasks well, such as word processing or dexterous dancing. However, machines have done poorly for (muddy) tasks that the brain is good at, such as perceiving and behaving properly in open ended, complex human environments. Human hand-designed task-specific representations face great challenges in such environments. Inspired by neuroscienece, this talk presents general purpose architectures that constrain the types of representation to be generated. Further, inspired by the biological laminar cortical structure, the Multilayer In-place Learning Networks (MILN) is presented to show what the internal representation is, how it is automatically generated while the system interacts with the environments. MILN has several major advantages over existing networks, such as FFN, RBF, SOM, CCLA, SVM and IHDR. This new technology indicates a commercial potential for generating a new kind of programs --- epigenetically programmed through human-machine interactions --- to lead to seeing, hearing, thinking and behaving machines and robots. (A more technical talk can be arranged in another setting.)