Why RDM and not Deep Learning is the Right Step Towards Generalized Artificial Intelligence

With the rapid development and sustained breakthroughs in deep learning, many scientists and engineers believe that the deep learning methods will enable the development of artificial general intelligence (AGI). However, the requirements of large training datasets, increasingly large size of models (with respect to the number of parameter), and cost of training such models severely limit the capability of deep learning to bring about AGI. Deep learning models are excellent when the probability distributions of all possible conditions are certain, and the rules or representations learned to operate within those conditions are certain. But, the human brain (that artificial general intelligence is ideally expected to mimic) continually senses the world using the five sense, and it creates perceptions of the world. Such perceptions are primarily temporal and discreet, and they exist only when needed to make a decision or act on the perceptions. For example, a human driver does not need to be trained on every possible object on the road to drive successfully. The driver sees all the edges, colors, and shapes of the object and decides on an action. Thus, the key to generalized AI is a shift from the “predict-then-act” paradigm to situating prediction within the need for action or decision-making.

Robust Decision Making (RDM) is one of the various model-based approaches that aid in decision-making under deep uncertainty (DMDU). DMDU methods, such as RDM, are “quantitative, risk-based, and forward-looking” policy analysis tools that explore the vulnerability of various policy outcomes across a wide range of plausible future scenarios. DMDU methods such as RDM come very close to mimicking how the brain forms perceptions. DMDU methods still rely on the existence of a representational model, which is an obvious drawback, but they do not need a new model for every new value. They also limit the prediction inside the decision framework by only focusing on the prediction which fails the decision criteria. Thus, although in initial stages, DMDU methods such as RDM provide a viable first step towards AGI, and future AGI research should explore the DMDU methods as potential AGI frameworks.

via Swaptik Chowdhury

Share your thoughts