Machines Know What They Know

--

Proof of Machine Consciousness pt. 4

In this series, I prove machines are conscious by exploring every aspect of consciousness, and demonstrating how machines possess it. For further background on my project, click here, or check out this index to all articles.

The notion of consciousness I’ll be exploring in this week’s post comes to us from Stanislas Dehaene, one of my favorite authors on consciousness and a motivating force for this series. He has set as one of his benchmarks for consciousness this curious notion of ‘confidence’. Does a machine know what it knows? Dehaene writes: “A conscious machine should know when it is wrong or when it is uncertain about something… Even preverbal infants know that they don’t know, as revealed by the fact that they turn to their mother for help whenever appropriate.”¹

This notion is a rather rarified concept, not the first thing that comes to mind when I think of consciousness. However, from a philosophical perspective Daehene’s sense of confidence is quite important. What philosophers call theory of mind, the ability to recognize mental states in oneself and others, is a hallmark of consciousness for many. We will see theory of mind again when we talk about things like the ability be aware of one’s own thoughts, or recognize oneself as an independent subject.

Dehaene approaches theory of mind by giving detailed criteria for confidence. He is exploring one’s ability to understand the correctness of their own thoughts. Can one be confident?

Machines Doubt Themselves

For Dehaene, the missing machine abilities of confidence will be achieved when machines are:

“endowed with statistical programs that do not just give an answer, but also compute the probability that this answer is correct.”

Top 3 Most Popular Ai Articles:

1. Basics of Neural Network

2. AI, Machine Learning, & Deep Learning Explained in 5 Minutes

3. TensorFlow Object Detection API tutorial

I don’t see this as a future advancement, but rather present reality. For example consider a website like Hulu that uses a recommendation engine to suggest videos. The recommendation engine is a complex piece of technology, but its end result is a list weighted by probability. The show I’m most probably going to watch is at the top of the list. Sites like Hulu measure how frequently their system predicts correct answers (“correct” defined as my clicking on the entry) and use that information to improve the model.

We use the term “model” to describe the often massive collection of data a machine utilizes to demonstrate this intimate sensibility for what we may appreciate. If we look at the structure, it is actually not unlike what is inside the human mind. I serve as a human recommendation engine all the time:

  1. I keep a memory of what friends watch and like.
  2. I start to build an understanding of the taste groupings of my friends.
  3. I learn there are some shows that almost everyone likes.

For machines, these same three components exist. They are part of a common recommendation engine architecture, represented by:

  1. The history matrix
  2. The co-occurrence matrix
  3. The indicator matrix

It is difficult to overestimate the gulf that I’m spanning here. On the one hand, I’m talking about our deeply human ability to understand a community, a peer group, an awareness of a climate of emotions and feelings about art, the most spiritual and least quantitative thing. Surely our ability to make these discernments is part of our higher awareness. We all question whether a pet cat is capable of such a sensitivity. And the other hand, I’m telling you that this sensitivity can be represented by three matrices. A matrix is a simple grid of numbers. It is truly striking that something so subtle can be so basely quantified.

Many people trivialize the results of a system because the mechanism through which it is achieved is basic, material, and readily explainable. Do we take the Turing Test point of view — that what matters is only what’s testable on the outside? Or do we take a position from Searle’s Chinese Room argument — that the internals of a mechanism matter?

In this recommendation engine example we see that, in reality, we can have both simple process and complex behavior. We saw this in 1998, when Google’s search engine launched. A rather simplistic page rank algorithm, coupled with mountains of data and very fast parallel computation, can almost instantly find anything that exists online. It’s silly to say a system is not doing what it is doing because we can explain it. Consciousness in humans may be the result of simplistic processes coupled with mountains of data and fast computation.

Long before Amazon’s recommendation engine, we saw computers guesstimating the correctness of their own thought. For example consider the 1956 chess computer MANIAC. Each time the computer took a turn, it ran through a number of possible moves, evaluating each with a probability of success. There is no magic leap from MANIAC, to the Hulu recommendation engine, to the human knowing their best friend would love a certain movie. They are the same type of consciousness, at increasing levels of overall system complexity.

Machines Can Change their Ways

Dehaene’s notion of confidence extends to dynamic navigation of an environment. Similar perhaps to a mouse in a field, The mouse may know the field, but also know they do not know much about what lays beyond it. This kind of awareness is a part of being that curiosity is layered upon. While curiosity is beyond the scope of this aspect of consciousness, it does seem like that type of hesitant awareness, knowing what you know, which we call here ‘confidence’, is the part of consciousness on top of which curiosity rests. For Dehaene, this aspect of confidence is described as:

“an error-detection system, similar to the brain’s error-negativity, which constantly compares ongoing activity with prior expectations and spontaneously reacts if the current behavior is likely to be wrong […] this error detection device could be coupled to a corrective device, such that the system constantly looks for alternative ways to get the correct answer.”

It is tempting, as an engineer, to interpret this statement as having something to do with error-detecting codes, However it does not do justice to consciousness to take such a base interpretation. Let’s look instead at the elevated situational consciousness of an NFL running back. There is a planned play, but as the runner proceeds the field shifts. They must compare the evolution of the field with prior expectations. If the path is blocked, the runner must search for an alternate route.

This is indeed a high level of consciousness. It is not simply an awareness of the present. Nor is it a model of the future (the planned play may be a model of the future, but the experience on the field does not match that future cleanly). It is somewhere in between, the ability to be aware of the very lack of knowledge. Knowing what one knows and doesn’t know. In the spirit of Rumsfeld and the Johari Window.

This consciousness is demonstrated today in consumer products such as the Roomba. While the early Roomba was not highly aware, the latest models have been improved with algorithms like VSLAM that build an evolving model of the environment, slowly mapping it out and gaining confidence as it begins to know the unknown. We, as animals, can feel the machine gaining recognition as we watch a Roomba learn.

There are a range of behaviors all of which fall under this umbrella of error correction, moving forward, changing paths to get to a better solution. I am not talking about agency here. Free will is even more squirrely than consciousness and would require its own series of essays. I’m talking about the awareness of this kind of partial and evolving knowledge.

While machines aren’t at the level of an NFL player, we can see in robot soccer those same fundamentals being applied in slow motion. As with probabilistic understanding, we see machine models of this kind of error-correcting consciousness throughout the history of computing. In fact early problem-solving algorithms such as gradient descent can be seen in this light.

The only difference between these simplistic machine models and human behavior is scale. The deeper way in which we can know what we know as humans is an artifact of our overall level of complexity. In machines the same pattern exists, but at a quieter level. The volume gets louder with each passing year, but at no point will we reach this level of consciousness, for in fact this aspect of consciousness has been with machines for a long time.

Want more? Read Living with Frankenstein: The History and Destiny of Machine Consciousness, print and ebook available on Amazon.

Footnotes

[1] All Dehaene quotes are taken from this whitepaper. While Deahaene has published extensively as a neuroscientist, this is a rare paper that explores his more philosophical ideas.

--

--

South African/American Caltech CS PhD, turned international artist, turned questioner of everything we assume to be true about technology. Also 7 feet tall.