A recent article in GCN, “DARPA director clear eyed and cautious on AI”, by Mark Pomerleau, illustrated DARPA’s knowledge of artificial intelligence (AI) technology and a need to be cautious.
Key points from the article include:
AI needs machines to automatically ingest data and act upon it. Although this is of course possible already with some kinds of information, no machine is smart enough to integrate many divergent facts into a clear picture of the overall situation in a real-world environment such as a battle space. It seems to me this is still a uniquely human ability involving powerful correlation of enormous amounts of information by simultaneously applying a number of different mental “tools” — e.g. pattern recognition (not just of images and sound, but other examples include recognition of categories/groupings/sequences of events, categories/groupings/interactions of players, patterns in deployment of assets), if/then/else logic, high-speed parallel processing by very different mental agents operating and collaborating simultaneously, biological neural networks, etc.
Pomerleau points to a weakness in current AI/Machine learning, citing one high performance image processing algorithm decided that a baby with a tooth brush was actually a baby with a baseball bat. This illustrates a key weakness in many of the machine learning algorithms that I studied earlier in the year. Namely, while they can often deliver correct results, often stunningly, you usually have no idea of how they did it. Neural nets are a good example of this. You train the features in a neural net with thousands of elements of data, and it eventually can get spooky good at, for example, interpreting wildly different styles of hand printing. But this advanced form of modern machine learning is, after you peel away the cool math, a big regression machine, analogous to simple linear regression but using a more complex mathematical construct. Once a neural net, for example, learns how to interpret all kinds of hand-printed characters from all kinds of people pretty infallibly, you still have no idea exactly how it is doing this or how to fix an obvious errors (other than feeding in more and better training sets). One good area of research would be layering other classes of AI on the neural net outputs to apply common sense to proposed solutions. Representing common sense, by the way, is no simple proposition! But there are ways to do some of this.
They article also references the desire to Causal models – where the machine actually figures out and understands the chain of logic that led to the current situation. Then they’d love to take what is learned from one domain and use it in another. Again, this still the domain of humans, in no small part because deep and subtle knowledge of the world at large is required to figure stuff like this out, and no one has figured out how to give a machine the highly organized a massive and heterogeneous information base – plus processing algorithms – that is required to do this.
The conclusion of article is right on the money – current AI has its place, but it can’t do everything. Sometimes conventional programming is even the better option!
Getting a machine to do what DARPA wants involves learning how to represent massive volumes of heterogeneous and highly correlated information about the real world, and creating a large number of heterogeneous, parallel agents that collaborate to use this information. This involves knowledge processing that goes way beyond current-day notions of “big data” processing.
Knowledge processing is playing an integral role in new Digital Risk Management frameworks that provide a practical, analytical discipline for managing digital risk from a business perspective, by enabling business and technology leaders and their security partners to collaborate on business aligned decisions. Knowledge processing is enabling the quantification of the business impact of digital risk for the first time. It is beginning to allow business leaders to understand the risk profile of their operations from a business perspective and for the organization as a whole to make risk mitigation decisions based on the level of operational and financial risk.
Check out the website for more information about advancements in knowledge processing and how it is being used in new digital risk management frameworks.
