Enterprise AI Is The Antithesis To AI Feudalism
By: Mark Stadtmueller, VP Product Strategy and Russ Loignon, SVP Market Strategy
The value of data and the ways in which artificial intelligence can create or deliver enterprise wealth should not be controlled by the few.
Feudalism was a medieval political system in which a lord owned all the land while vassals and serfs farmed it. Today, as it pertains to AI, it has been suggested that large AI players are controlling the process and the system while smaller, more nimble companies are competing for a slice of the pie. As a result, some have suggested that 2018 is the year in which we have begun to see the emergence of ‘AI feudalism‘
While this is the most recent reference to potential concerns regarding AI feudalism, it is not the first. If one were to search the internet of AI feudalism, the results are staggering. These many articles speak of concerns around the impact and control of proliferating data lakes. Take healthcare, for example. Is health data owned by the providers, payers, patients, or other stakeholders? In the case of the Internet of Things, is the data owned by the data collector or the contributor? The ownership and subsequent benefaction of data through AI needs to be handled responsibly – and so far, the lack of progress on this issue is concerning.
Business results from AI can be impressive, but if most enterprises cannot overcome these challenges, then AI feudalism looms. The three key challenges contributing to this are: data size and compute requirements; data governance and accountability; as well as data science and engineering skillsets.
Concern is rightfully growing that only a handful of large organizations will be able to address these challenges. Yet, another view as to why these larger AI providers are perceived as the only solution may come down to their reputation, capabilities, or pure momentum. However, there are many innovative and more nimble companies capable of taking these challenges head-on as well. By investing in enterprise AI, these companies are enabling independence and freedom for businesses through distributed control of the technology.
DATA SIZE AND COMPUTE REQUIREMENTS
AI has led to groundbreaking results in computer vision, natural language processing, speech recognition, and other specialized capabilities that allow us to learn from data. However, data volumes are growing as a result, with compute requirements intensifying. Even with cloud computing, the resources necessary to support this growth is often too much for many businesses to completely harness.
To innovate in AI, businesses need to not just use models that others have trained, but also train their own models unique to their infonomics. This, argues author Douglas B. Laney, enables companies to monetize, manage, and measure their data as a competitive asset. As a result, scaling computer resources efficiently is a key and necessary capability.
One way to address these challenges is to make the training task easier. At Lucd, we are investing in something called reservoir computing – a computational framework where the data training is conducted at the readout portion of the process. It has been shown that AI accuracy is maintained while speed is increased and the cost is reduced. Modern deep learning training models and data are often very large with training challenges that often have trouble converging to a solution. RC is fast and computationally simpler compared to current deep learning methods. By streamlining the training task, businesses will have greater capability to handle the data size and compute requirements.
DATA GOVERNANCE AND ACCOUNTABILITY
If data training is left up to the large tech and AI players, so too will innovation. This raises some key concerns for enterprises. How do you know how the models were trained? Bias and accountability in data and training can unwittingly distort results. Additionally, if models are not unique to the enterprise and their customers, how can a business responsibly train from data?
Feudal AI lords, on the other hand, can throw resources at the problem, while businesses cannot afford to handle their unique data as if they were just learning to differentiate between cat and dog photos. For more businesses to differentiate with AI, they need to be able to handle data effectively and responsibly.
Our Unified Data Space is one example of an effective mechanism with which businesses can leverage both a persistent data store and a responsible data factory for training. Automated features can reduce costs to maintain compliance as well as an internal security audit framework. With a responsible data factory, businesses can... READ MORE