Recent technological breakthroughs in AI have the power to transform every aspect of business operations. This is particularly true for generative pretrained transformers (GPT) and reinforcement learning. Following is an overview of some of these breakthrough technologies and why they are so much more powerful than traditional AI approaches.
Transformers: Having clear historical context for the current business situation is fundamental to making good decisions. Traditional AI either: a) left this burden to people, forcing them to feature engineer the proper contextual data, or b) used recurrent systems to form a memory that could learn the context. For complex problems, having people feature engineer all the historical context for the optimal solution is not possible. Meanwhile recurrent systems left us with good short-term memory but fuzzy long-term memory. Transformers are game changing because they can learn historical context with unlimited memory depth and clarity. When given all the historical context for the current situations, transformers learn what to pay attention to when making a particular decision.
Pre-Training: Most AI applications require human data labeling that provide expert examples to learn from. Data labeling is typically slow, tedious, error prone and expensive; a fact exacerbated by the fact that AI requires an enormous amount of data to learn from. A form of learning called unsupervised learning has been used for decades for a small number of use cases. Next-gen AI now leverages unsupervised learning for all use cases by pre-training models using all of the available data in an unsupervised setting, then reusing these models to train much smaller use-case specific supervised models; which require significantly fewer labels.
Reinforcement Learning: From a business decision perspective, the biggest breakthrough in next-gen AI is a completely new approach to modeling. Reinforcement Learning (RL) is a new form of learning that fits somewhat between supervised and unsupervised learning. In supervised learning we train AI by giving it the answers, i.e., the labels. In RL, however, we give AI the problem to be solved without any information about how to solve it, nor examples of solutions. With RL, human domain experts focus exclusively on codifying their knowledge of the business context for a decision to be made, and then RL learns how to make the decision to maximize future return.
nDimensional combines these technologies to solve industrial business challenges of unprecedented complexity and dimensionality, and to optimize actions to deliver quantifiable business value.