Skip to content

Posts from the ‘Recommendation explanations’ Category

Just the Facts Ma’am?

Now that Siri has a bona fide competitor, Google Voice Search, a bit of a kerfuffle has emerged with regard to personalities or lack thereof of these assistants. While Siri strives to project some personality by being conversational and peppering her responses with a bit of whimsy, Google Voice Search is all about just giving us the facts. Each approach has advantages and its vocal adherents. And as the systems’ capabilities leap-frog one another with each new version, the latest incarnation of Google Voice Search seems to have gained some speed and effectiveness advantages versus the current incarnation of Siri. Of course, both of these incarnations promise to be fleeting given the pace of the respective development cycles.

Although Google labels their product “search,” the functionality has clearly already morphed more generally into a recommender—i.e., providing suggestions given a context of various of kinds. This trend is a reflection of a generalization noted in The Learning Layer—plain old search is really best considered just a recommendation in which the context is of a particular type, i.e., a search term provided by the user. The inevitable next step in general-purpose recommender technology is delivering “meta-recommendations”—that is, explanations as to why the recommendation was provided, particularly when an explanation is specifically asked for by the recommendation recipient. A capacity for a limited degree of explanatory capability has already been incorporated into the Apple and Google gals to some degree.

Then comes the really interesting advance—making the recommendations and even the explanations adaptive to the user.  That is, learning from her experiences with us to adapt her recommendations and explanations accordingly. Which is followed by one more short step in which aspects of her overall personality become adaptable to us and our particular circumstances as well. A little humor when called for, a bit of sympathy at other times; and all the while learning as to what works best and when, and tuning accordingly. I’ve got a feeling that at that point, which at the current pace of innovation, is not far away, always just providing the facts will be perceived to be somewhat stilted behavior—coming off like a cheesy movie version of AI of the 1960s.

So my guess is that there are times when, indeed, we are all Joe Friday’s, but more often than not we’ll welcome more than just the facts.

Our Conceit of Consciousness

MIT recently held a symposium called “Brains, Minds, and Machines,” which took stock of the current state of cognitive science and machine learning, as well as debating directions for the next generation of advances. The symposium kicked-off with perspectives from some of the various lions of the field of cognitive science such as Marvin Minsky and Noam Chomsky.

A perspective by Chomsky lays down the gauntlet with regard to today’s competing schools of AI development and directions:

Chomsky derided researchers in machine learning who use purely statistical methods to produce behavior that mimics something in the world, but who don’t try to understand the meaning of that behavior. Chomsky compared such researchers to scientists who might study the dance made by a bee returning to the hive, and who could produce a statistically based simulation of such a dance without attempting to understand why the bee behaved that way. “That’s a notion of [scientific] success that’s very novel. I don’t know of anything like it in the history of science,” said Chomsky.

Of course, Chomsky is tacitly assuming that “meaning” and truly understanding natural language is something much more than just statistics. But is it really? That certainly seems like common sense. On the other hand, if we look closely at the brain, all we see are networks of neurons firing in statistical patterns. The meaning somehow emerges from the statistics.

I suspect what Chomsky really wants is an “explanation engine”–an explanatory facility that can convincingly explain itself to us, presenting to us many layers of richly nuanced reasoning. Patrick Winston at the same symposium said as much:

Winston speculated that the magic ingredient that makes humans unique is our ability to create and understand stories using the faculties that support language: “Once you have stories, you have the kind of creativity that makes the species different to any other.”

This capability has been the goal of AI from the beginning, and 50 years later, the fact that such a capability has still not been delivered has clearly not been for lack of trying. I would argue that this perceived failure is a consequence of the early AI community falling prey to the “conceit of consciousness” that we are all prone to. We humans believe that we are our language-based explanation engine, and we therefore literally tell and convince ourselves that true meaning is solely a product of the conscious, language-based reasoning capacity of our explanation engines.

But that ain’t exactly the way nature did it, as Winston implicitly acknowledges. Inarticulate inferencing and decision making capabilities evolved over the course of billions of years and work quite well, thank you. Only very recently did we humans, apparently uniquely, become endowed with a very powerful explanation engine that provides a rationale (and often a rationalization!) for the decisions already made by our unconscious intelligence—an endowment most probably for the primary purpose of delivering compact communications to others rather than for the purpose of improving our individual decision making.

So to focus first on the explanation engine is getting things exactly backward in trying to develop machine intelligence. To recapitulate evolution, we first need to build intelligent systems that generate good inferences and decisions from large amounts of data, just like we humans continuously and unconsciously do. And like it or not, we can only do so by applying those inscrutable, inarticulate, complex, messy, math-based methods. With this essential foundation in place, we can then (mostly for the sake of our own conceit of consciousness!) build useful explanatory engines on top of the highly intelligent unconsciousness.

So I agree with Chomsky and Winston that now is indeed a fruitful time to build explanation engines—not because AI directions have been misguided for the past decade or two, but rather because we have actually come such a long way with the much maligned data-driven, statistical approach to AI, and because there is a clear path to doing even more wonderful things with this approach. Unlike 50 years ago our systems are beginning to actually have something interesting to say to us; so by all means, let us help them begin to do so!

The Chinese Room Explanation

My favorite recommendation explanation, a real honest-to-goodness computer-generated recommendation, is the one pictured below. I was amazed when I received it even though I had a hand in designing and developing the system that generated the recommendation—the system had truly become complex and unpredictable enough to spark a delighted surprise in its creators!

The recommended item of content in the screen grab is about AI and the explanation engine whimsically mentions as an aside that it has been pondering the Chinese Room Argument—the explanation engine’s little joke being that, as AI aficionados know, the Chinese Room Argument is philosopher John Searle’s famous thought experiment that suggests computer programs can never really understand anything in the sense that we humans do. The argument is that all computers can ever do, even in theory, is just operate by rote and that computers will therefore always be relegated to the realm of zombies, albeit increasingly clever zombies.

The Chinese room thought experiment goes like this: you are placed in a room and questions in Chinese are transmitted to you. You don’t know Chinese at all—the symbols are just so many squiggles to you. But in the room is a very big book of instructions that enables you to select and arrange a set of response squiggles for every set of question squiggles you might receive. To the Chinese-speaking person outside the room who receives your answers, you appear to fully understand Chinese. Of course, you don’t understand Chinese at all—all you are doing is executing a mind-numbing series of rules and look-ups. So, even though in theory you could pass the test posed by that other famous AI thought experiment, the Turing test, you don’t really understand Chinese. And this seems to imply that since, most fundamentally, all computers really ever do is execute look-ups and shuffle around squiggles in prescribed ways, no matter how complex this shuffling is it can never really amount to truly understanding anything.

That’s all there is to it—the Chinese Room Argument is remarkably simple. But it is infuriatingly slippery with regard to what it actually tells us, if anything. It has been endlessly debated in computer science and philosophy circles for decades now. Some have asserted that it proves computers can really never truly understand anything in the sense that we do, and certainly could not ever be conscious—because, after all, it is intuitively obvious that manipulations of mere symbols, no matter how sophisticated, are not sufficient for true understanding. The counterarguments are many. Perhaps the most popular is the “systems reply”—which posits that yes, it is true that you don’t understand Chinese, but the room as a whole, including you and the book of response instructions, understands Chinese.

All of these counterarguments result from taking Searle’s bait that the Chinese Room Argument somehow proves that there is a limit on what computer programs can do. What seems most clear is that these armchair arguments and counterarguments can’t actually prove anything. In fact, the only thing the Chinese Room really seems to assure us of is that there is no apparent limit on how smart we can make our zombie systems. That is, we can build systems that are impressively clever using brute-force programming approaches—even more clever than us, at least in specific domains. Deep Blue beating Gary Kasparov is a good example of that.

As I have implied previously, understanding, as we humans appreciate it, is a product of our own, unique-in-the-natural-world explanation engine, not our powerful but inarticulate, unconscious, underlying neural network-based zombie system. The Chinese room doesn’t directly address explanation engines, or any capacity for learning, for that matter. There is no capacity in the thought experiment for recursion, self-inception, and for the system self-modification that occurs, for example, during our sleep. These are the capabilities core to our capacity for learning, understanding, creativity, and yes, even whimsy. We will have to search beyond our Chinese room to understand the machine-based art-of-the-possible for them.

And we will surely continue to search. We know we can build arbitrarily intelligent systems; but that is not enough for us to deign attributing true understanding to them. This conceit of our own explanation engines is summed up by the old adage that you only truly understand something if you can teach it to others—that is, only if you can thoroughly explain it. Plainly, for us, true understanding is in the explaining, not just the doing.

The Explanation Engine

Recommendation engines are now becoming quite familiar parts of our online life, although we are still in the early stages of the recommendation revolution. The ability to infer your preferences and deliver back to you valuable suggestions of content, products, and even other people is the wave of the future—another manifestation of the dominant IT theme of this decade, “adaptation.”

But what intrigues me even more than the intelligent recommendations themselves is a topic I discuss at length in The Learning Layer: the ability for the system to deliver to you an explanation of why you received a recommendation. You may have already experienced some simple examples of the “people that bought this item also bought these items” variety. But in more sophisticated systems that infer preferences and interests from a great many behavioral cues, there is an opportunity to provide much more detailed and nuanced explanations. Among other subtleties, these increasingly sophisticated explanations can provide you with not just the rationale for making the recommendation, but also any caveats or reservations the system might have, as well as a sense of the system’s degree of confidence in making the recommendation. Providing such a nuanced, human-like explanation is in many ways a more daunting technical achievement than providing the recommendation itself.

Among the many things that intrigue me about explanations for recommendations is how they can shed some light on the peculiarities of how our own brains work. This was brought home to me as we worked on designing explanatory capabilities for our recommender systems. At first we thought this would be pretty trivial—we would just have the system regurgitate the recommendation engine’s rules for recommending the particular item. But here’s the rub—the rules are not always or even typically “if-then” rules of the sort that we humans are going to easily understand. More typically, rating and ranking of potential items to recommend are going to be the product of various high-powered mathematical evaluations and manipulations of vectors and matrices. How can they possibly be compactly conveyed in an explanation to the recommendation recipient? After a while we realized that they really cannot—explanations necessarily have to be an approximation of the actual thought processes of the recommendation engine. A very useful approximation, but an approximation nevertheless. We also realized that to do them right, it was an architectural necessity to have a dedicated engine for explanations complementary to, but separate from, the powerful but inarticulate recommendation engine.

Turns out we are no different—we humans have a language-based explanation engine that explains to others, as well as ourselves, why we do the things we do. And it has become increasingly apparent from psychological studies over the past decade or two that our explanations are really only approximations of our underlying, unconscious decision making. In fact, it has been confirmed by recent brain imaging studies that although we tend to believe that our conscious and logical explanation engine is making the decisions, in reality it is just providing an after-the-fact explanation for a decision already unconsciously made elsewhere in our brains. Moreover, studies involving hypnosis have revealed that our explanation engine just makes up stories as required to provide a rationale for an action if the real motivation is not accessible to the conscious explanation engine!

Now this seems like a pretty weird state of affairs. But taking some lessons from building explanatory facilities for recommender systems suggests that this seemingly strange brain behavior is really the only way it can be. After all, our brain most fundamentally is nothing more than a vast, weighted network of neurons. Decisions are assessed and made by inscrutable interactions among this network. We humans, fairly recently in our evolutionary history, have developed a language-based explanation engine that has essentially been grafted on to this underlying network that enables us to effectively communicate with other humans. To achieve a reasonable compaction, the explanations we give must necessarily typically be extreme simplifications, and quite often will also have some degree of fabrication woven in. And our explanation engine continuously explains itself to us—so that we come to believe that we are our explanation engine, that there is a clear logic to what we do, that we have an explicit freedom to decide as we wish, and a variety of other explanatory conceits.

There is really no other way we could be—it’s an inevitable architectural solution. Turns out the explanation engine has a good deal to say about us!

Engineering Serendipity

Serendipity, the notion of unintentional but fortuitous discoveries, has become a trendy concept of late, and particularly with regard to system platforms that can help foster it for us. For example, John Hagel’s recent book, The Power of Pull, addresses this idea in some detail, and very recently Eric Schmidt, Google’s CEO mentioned electronically calculating serendipity in the context of an interview on Google’s technology directions.

Although just about everyone would agree that some of our systems already serve to bring to our attention unanticipated but useful nuggets of information and people of which we would otherwise remain unaware, the topic has also generated some controversy—perhaps partly because of a potential over-selling of the degree to which today’s systems can truly facilitate useful serendipity, and partly because of a nagging worry that serendipity that is “engineered” may be engineered to serve the hidden purposes of the engineer.   

I discuss some approaches in The Learning Layer that address both of these issues. Engineering serendipity really amounts to delivering automatically generated recommendations of items of content or people. To do this well, the inference engine that generates a recommendation for you has to strike the right balance that lies between the extremes of delivering stuff you already know about and stuff that is so far afield it is very unlikely to have any useful serendipitous effect. And the degree to which that balance can be effectively struck is a function of the behavioral information the inference engine has to work with and its ability to wring the best possible inferential insights from that information. 

Progress on more sophisticated inferential insights is really where all the AI action is these days, and it is inevitable that these capabilities will become ever more powerful, and the resulting recommendations ever more prescient and helpful. But that then amplifies the second concern—how do we know the basis for the recommendations we receive? And even more to the point, how do we know we aren’t being manipulated? 

A solution to this issue that I discuss at length in the book is that recommendations, in any of their various forms (e.g., an advertisement is best thought of as a recommendation), should always be coupled with an explanation of why the recommendation was delivered to you. A good, detailed explanation will inevitably provide you with even more insights that can spur a useful serendipity. And an informative explanation can do even more—it can create a greater degree of trust in that system-based engineer of serendipity. Call it engineering serendipity with transparency. We would want nothing less from a friend doing the recommending, so why not demand the same from our systems?