Source Link
The enthusiastic embracing of AI as the go-to technology for solving specific problems is both undeniable and remarkable. But while there is still much progress being achieved every day through the most popular AI approaches like supervised learning or reinforcement learning, the often monolithic way in which those classic approaches are used may also be the very thing that holds AI back.
While AI is increasingly successful in a growing number of fields, it still operates primarily as a tool to execute narrow-focus tasks, or as a simple form of automation, rather than a supporting partner in a relationship with human users. It largely relies on carefully curated or annotated, mostly historical, data, and only very indirectly learns from human users. AI has remarkable predictive power in some cases, yet is incapable of the adaptive prowess routinely demonstrated by humans from their infancy. It simply is not (yet) able to extrapolate on data that it has never encountered quite like humans can. Additionally, the need for more accuracy is leading to ever larger and complex models, compute-intensive training, and engineering challenges that hinder the trustworthiness, transferability, and scalability we seek in AI-based solutions.
Achieving our AI goals requires a shift from the current data paradigm; it’s time to put humans at the center of the AI training process. You don’t have to take our word for it: the benefits of mixing human and AI resources from design to deployment are echoed in other independent studies like the MIT Sloan’s Findings from the 2020 Artificial Intelligence Global Executive Study and Research Project, and the resulting collaboration has even been dubbed “Super Teams” in this Deloitte Insights.
Designing, training, and deploying solutions mixing human users and AI agents provides new avenues for success compared to standard AI approaches. Imitation Learning, Curriculum Learning, and other more recent techniques already show other means to train AI by leveraging human expertise, feedback, and guidance. Instead of limiting ourselves to one approach, what if we could have it all and combine all those different approaches, alongside humans, to build new intelligent systems not bound to any specific method, model, or algorithm anymore? Considering human and AI respective strengths and weaknesses, such a human-AI partnership would yield more than the sum of its parts, leveraging complementary abilities towards results that would otherwise be impossible or very difficult to achieve with only one or the other. However, for AI agents to work as synergistically and closely as possible with humans in the loop, specific methods, approaches, and technologies are warranted. It calls, notably, for an architectural design that is naturally conducive to multi-agent, multi-human, tech-agnostic, distributed approaches, with fast and frictionless iterative back and forth between research, prototyping, and operationalization.
Those were the guiding principles behind Cogment, a novel open-source framework designed to enable such a partnership between human and AI agents. It’s already used today in complex contexts like man-machine teaming, adaptive learning, and critical decision-support systems, and it’s primed to tackle the challenges of tomorrow.
No comments:
Post a Comment