I recently came across a 2008 lecture by Anthony Greenwald, who says he can tell you precisely how biased you are. Or more specifically, he developed a psychological test that measures unconscious associations. (You may have read this in Blink).

Greenwald uses the terms “Level 1″ and “Level 2″ to distinguish between deliberate, high-level thinking and automatic, unthinking behavior.  We’re often on Level 2 auto-pilot, even when we think we’re in Level 1. For example, if asked how you’re doing, “Fine” might be a Level 2 response, whereas “The North Koreans are on to me” suggests Level 1 engagement.  (In HTM parlance, Level 2 is the state when lower-level predictions are met.)

Next, Greenwald shows a cool optical illusion illustrating how context can create powerful, unconscious bias.  In this image, squares A and  B are exactly the same color (seriously):

Click the image above for proofs and an explanation of the illusion.

Finally, he described the Implicit Association Test (IAT).  In one test, you categorize pictures of light vs. dark skinned faces, then “good” vs. “bad” concepts (e.g., joy, love, agony, failure).  Next they pair the categories. If it’s a light face or “bad” concept, press one key. If it’s a dark face or “good” concept press another.   Then they switch the pairs.

Greenwald found significant differences in reaction time for different combinations (measured in milliseconds). He claims this represents unconscious bias:  if two concepts are associated, you answer more quickly (responding at Level 2). Taking longer means your Level 1 brain needs to be engaged to overcome Level 2. The test has its critics as well as supporters (Greenwald’s responses can be found here). But it’s fascinating to go to the IAT website and test yourself.

I found my results to be…unsettling. Greenwald said that he’s taken the test hundreds of times and his biases have barely changed.  Someone asked if one could become aware of when you were biased.  Greenwald politely replied that, well, it wouldn’t be “unconscious” then would it?

I wonder if it’s fair to say that it only takes me a few hundred milliseconds to overcome my unconscious biases if I stop to think….

If you wanted to replicate “creativity” in software, you’d need to define your goals.  According to HTM theory, all cognitive activity is inherently creative. So technically, you could fist bump if you got to pigeon-level creativity, or you could hold out for David Lynch level.

What then defines what we consider human-level creativity? Some say it’s the production of novel and useful things (which could make me reconsider David Lynch—I’m not sure if Eraserhead counts as “useful.”) Kenneth Heilman speaks of “divergent reasoning,” or abandoning known methods to generate previously unknown solutions. (This contrasts with “convergent reasoning,” which assembles stored knowledge to choose known solutions, e.g., diagnoses based on medical training.) And Rex Jung quipped to the NY Times, “Creativity is kind of like pornography—you know it when you see it.” (Spoken like a true scientist.)

Undaunted, research on creativity forges ahead. Sometimes subjects are asked questions like how many creative uses you can imagine for a brick. (1. Throw it at a pigeon. 2. Throw it at David Lynch….) One study measured brain activity at the “aha moment” when you solve a brain teaser or get a joke. (So the kamikaze instructor says, “Watch carefully because I’m only going to show you this once.”) Scientists have isolated patterns of brain activity that correlate with creative thought. So we don’t know exactly what we’re measuring, but apparently it’s measurable.

OK, so how does creativity work? Maybe I’m biased, but theories I hear sound a lot like HTM to me. For example, Semir Zeki studies the neural basis of artistic creativity (neuroesthetics). He describes two basic laws of visual perception: “constancy” and “abstraction.” The brain focuses on the aspects of object that remain constant, and forms high-level concepts to avoid being “enslaved to the particulars.” This closely mirrors HTM concepts of invariant representation and hierarchical inference.

Heilman defines creativity as discovering “novel orderly relationships.” Again, that’s a primary function of HTM.  I don’t know if HTM belief distribution algorithms can be tweaked to generate creative outputs, but it seems to be developing in the right places to try.

Postscript: A study cited in this dissertation found that “the shifts in attention and concentration measured by low frequency EEG that occur in creative thinking are similar to those experienced by unmedicated schizophrenics prior to hallucinations.” It didn’t occur to me that a buggy creative program could be, um, psychotic. Suggestion: when developing advanced AI for nuclear missile command systems, stick with the convergent reasoning algorithms.

Products would be easier to use if designers had more empathy for users. We could start by stop calling people “users.” Does the Four Seasons call you an “occupant?” “User” suggests no relationship with the designer. If you poke something with a stick, you’re a “user” of the stick. If you pay me for my product, I call you a “customer.”

And why do we invoke “user error” to dismiss customer problems? If you click “Abort” instead of “Cancel,” we blame you, not our design.

I read a forum where a designer advocated using “Cancel” instead of “Abort.” Others objected, saying this argument was “strange” and would “confuse the user more.” (That word again.) After all, the difference is clear in “everyday language.”

Maybe. Navy Seals understand that “canceled” missions mean staying home, and “aborted” missions involve racing to extraction zones carrying wounded colleagues.

To programmers “cancel” requires restoring the previous state. It’s like the difference between getting out of Dodge (abort), and returning the rustled cattle first (cancel).

But customers using your software have no idea whether it’s an important distinction or not. If it doesn’t matter, why not use the simpler (albeit inaccurate) “Cancel?” If it does matter, could we say something less intimidating (like “stop”?), or better yet eliminate the need for the distinction? Sure, rounding up cattle is a pain (especially the ones stampeded off the cliff). But delighting customers is hard work, and “intimidating” typically isn’t considered delightful.

Labels and icons create design confusion because their context differs for customers and designers. To customers, they are like inkblot tests. A “V” icon could mean anything (View? Verify?). To designers they are labels for a fixed number of well-known features, like hints for multiple-choice tests.  To the ROYGBIV app designer, V obviously means “Violet.” (And given your knowledge of the color spectrum, you’re convinced people would be confused if you combined Violet and Indigo and called it “Purple.”)

Of course some features stubbornly resist clear description (Vitamin D Video definitely has some…!).  But Glenn Beck’s opinions notwithstanding, empathy is the first step towards good design.

"User error" or "we let down a customer?"

After the Times Square bombing attempt, the Washington Post headline read, “Times Square bombing attempt reveals limits of video surveillance.”  Some claim this proves video surveillance is a waste of money that threatens our civil liberties. Others suggest that it’s necessary to double down and invest more. Still others express ambivalence–it’s creepy and reassuring at the same time.

My experiences put me in the ambivalent camp. In college, I participated in an anti-apartheid rally, and was unnerved to see a police officer recording us with a video camera. On the other hand, I’ve felt reassured seeing a CCTV camera on a deserted subway platform late at night.

So what are the facts? How effective is CCTV surveillance? Unfortunately, the evidence is mixed. A British government report summarized, “Those who expected that this evaluation would show CCTV to be either an unparalleled success, or an affront to a democratic society, will be disappointed.” The study found that crime did not decrease relative to control groups in a majority of cases it evaluated. On the other hand, they found specific situations where CCTV was effective (e.g., small and enclosed spaces), and cases where poor implementation hindered effectiveness.

The study also included public attitude surveys, finding that concerns over civil liberties actually declined 2-7 percent after cameras were installed. The percentage of people “happy…with the security cameras” declined, but overall support remained over 70 percent in all but one area.  That’s encouraging, but it’s premature to declare the problem solved.

Personally, I don’t feel like my privacy has been violated if I’m recorded in public places—providing I’m not the primary target of extensive tracking. (That might answer the columnist who asked “why doesn’t it feel as creepy” that tourists might collectively have much more video of her than the government.) I’d also argue that the ethical line could be different than the legal (expectation-of-privacy) line.  But I can see why others might argue that creating video that could be misused is in itself a privacy violation.  It’s not a simple issue, but it’s one to keep an eye on.

Previously on Vitamin D Blog, I described Hawkins’ theory of how the brain learns and recognizes what’s in the world. Data streams are fed into a hierarchical network, where each node categorizes data into patterns and passes it up the network, and higher nodes find patterns of patterns.  The next key function of the neocortex to explain is making predictions.

Predictions? I only make predictions about whether the Bills will cover the spread (spoiler alert–they usually don’t).  Why would Hawkins call prediction the “primary function of the neocortex, and the foundation of intelligence?”

Well, he doesn’t mean Nostradamus-style prediction.  HTM theory asserts that your brain continually makes small predictions of what its senses will perceive next, based on learned patterns. While sensory data flows up the network, predictions flow back down. If predicted input matches what is seen, you pay no attention and continue.

If you see unpredicted patterns, however, it raises flags. Lower-level discrepancies can be disregarded if they still fit higher-level patterns, kind of like spelling mistaks.  But unpredicted high-level patterns grab your attention. You’re surprised if a sentence makes no fuse box (like Monty Python’s Thripshaw’s Disease).

Or if you’re ice skating, you may unconsciously adjust to unexpected minor variances in pressure your feet feel. If you slip and fall, however, a metric buttload of sensory predictions are violated. As your feet fling into the air, and you see ceiling instead of wall, the highest levels of your brain switch to full WTF mode.

In other words, verifying predictions is how you understand your world.

But there’s more. If you insert a prediction instead of sensory data into the hierarchy, you can see what larger pattern emerges next. Then feed that prediction back into the network, instead of sensory data for the next time point. Lather, rinse, repeat to create a chain of predictions that constitute imagination and planning.

Prediction also directs physical behavior, the last of the four core functions of the neocortex.  I’ll get to that later, but here’s a hint: the brain treats behavior like just another pattern.

"A dolphin will rise from the sea and meet the buffalo on a great field of battle. The buffalo will suffer cruel defeat, without having covered the spread." - Nostradamus

Previously on Vitamin D Blog, I mentioned the Pennsylvania school that allegedly turned on webcams of laptops given to students, recording them in their homes without their knowledge. The school claimed that the cameras only were activated 42 times to recover lost laptops. Later it turned out that “activated” meant ongoing surveillance.  In all, 56,000 images, as well as websites visited and chat threads, were captured.  The suit alleges email from a staffer saying the surveillance was like “a little…soap opera,” and a response from the administrator of the program saying, “I know, I love it.”  (Security technology advocates to administrator: YOU’RE NOT HELPING.)

Afterward, the developer of the technology, Absolute Software, announced they were removing the feature. The company claimed “Theft Track” was a legacy feature of a product meant for “lifecycle management.” The company blog stated, “[W]ebcam pictures are not a useful tool in tracking down the location of a stolen computer.” (Apparently not, if you need 56,000.)

Distancing themselves from this feature sounds credible, however, since the company has another product that is designed for theft recovery. You have to file a police report first; then Absolute locates the machine using its IP address (no pictures), and informs the local police directly.  This seems like a much smarter approach.   As their Marketing VP said to Computerworld, “Even if you are able to locate the laptop on your own, what do you do then?”

Subsequently, in response to this event Senator Arlen Specter introduced the Surreptitious Video Surveillance Act of 2010. This bill would make it illegal to video anyone secretly in a residence who has “a reasonable expectation of privacy.”

Will this make nanny-cam recording illegal? Hidden ones might be illegal. The bill’s press release outlines important exceptions, however, for residential surveillance with consent, cameras in the workplace, undercover operations and “residential security systems which use video cameras.”

I suppose  highlighting abuses of security technology isn’t the best way to increase our sales, but it’s important to acknowledge privacy implications and educate people about them.  For more information on digital privacy rights, check out the Electronic Frontier Foundation.

Tip #1: If you see a lit LED like the green light above, your webcam is on. Tip #2: No one has developed software that can see through tape placed over the lens.

Have you ever used a product that is so difficult you asked yourself, “How did they ship this steaming pile?” Doesn’t anyone evaluate usability?

Many companies make mistakes conducting usability testing, but I want to discuss people who do it right–and still get it wrong.  In other words, when can results mislead you?

Usability testing typically involves showing people products and asking them to try several tasks. This gives you a good assessment of how easy your product is to learn, but could miss (or create) problems in other areas of usability.

For example, a designer once told me testing showed people couldn’t find a given feature. So they decided to display a dialog offering the feature every time you used the app. Problem solved. Unfortunately, other  problem created, as in, I can’t believe you show me this stupid screen every time. And follow-up testing testing might not detect such problems since people are more patient in “tests” than in real life (e.g., they may actually read wizard text rather than impatiently clicking “Next”). To cite the Heisenberg principle (albeit inaccurately), the act of measuring usability can affect the results.

Also, usability testing captures the first experience, not learning curves.  As novelty wears off, swearing may ensue.  Or conversely, some features are difficult to discover but delight customers anyway (classic example: ejecting Mac floppy disks by dragging to the trash).   A test subject once explained this paradox to me by saying, “it’s intuitive once you figure it out.” (Tip: when testing, repeat troublesome tasks to test retention.)

Most important, over time the value of intuitiveness declines, but the value of efficiency increases. You stop learning new features, but extra steps for the ones you use often become annoying. (For the theoretical extreme of “the most intuitive product ever,” see the Onion spoof of a Mac with no keyboard but a giant iPod wheel.  “Everything is just a few hundred clicks away.”)

In other words, myopic reliance on process can create a false sense of success.  Understand the limitations of usability testing, and augment it with real-world evaluation over time.

If HTM theory describes how brains work, could you build a conscious machine? Well, first you’d need a theory of consciousness. David Chalmers divides consciousness theory into “easy” and “hard” problems.  “Easy” problems involve cognitive capabilities like attention, autonomy, being awake. To explain these, you “need only specify a mechanism that can perform the function.” Easy peasy.

The “hard” problem is explaining why we have subjective experience (feelings like pain, emotions, and wonder).  Chalmers argues theories ignoring the hard problem are cop-outs.

But if I wrote a product brief for a conscious machine, first I’d ask, who needs consciousness? That’s because I don’t see HTM leading to sentient machines, but I do think  HTM could solve problems people think require consciousness.

For example, conscious machines might excel at autonomous problem solving. But you could also imagine autonomous programming without sentience. Army ants aren’t conscious, but they’re autonomous enough to ruin your safari.

Maybe sentient robots would be creative. But you can talk about creative HTM networks without introducing consciousness.

Or maybe conscious robots would seem more “human.” But if HTM leads us to creative, autonomous robots, our instinct to anthropomorphize could make clever UI tricks eerily effective. For starters, replace the vacant stare of sci-fi robots with eyes that track speakers and have simulated saccades.

In other words, if C3PO brings you vodka tonics and shoots Imperial Stormtroopers, why upgrade to the sentient model? Could you even tell the difference? (Philosophers call these “zombies”—not the flesh eating undead, but indistinguishable imitations of sentient beings.)

On the other hand, if consciousness could provide benefits beyond advanced intelligence, what are they? If you had a “conscious” Roomba with the intelligence of a pigeon, it might just be annoying.

Don’t get me wrong.  Sentient robots would be fascinating. And maybe they would be better (say some recursive observational algorithms required for self-awareness enable sentient machines to solve harder problems). I’m just suggesting HTM may not be a path to “hard-problem” consciousness, but an “easy-problem” level of intelligence would enable a pretty kick-ass robot.

P.S. Check out Jeff Hawkins’ thoughts on consciousness.

Descartes' dualist theory of consciousness: how will this lead to Cylons?

If Vitamin D has state-of-the-art human recognition, why not try recognizing faces?

We considered this, first by evaluating existing technologies.  NYU reports that in “laboratory conditions” facial recognition can generate 99% accuracy. So is this a “solved problem?”  What about the real world?  Well, mug shots and passport photos control camera angles, distances and lighting.  This lets you generate consistent metrics on geometric relationships between eyes, nose, cheekbones, and jaw. Altering your facial topology is hard, though just in case, Canadian passport photos require a “neutral facial expression.”  (Apparently, the Canadian government frowns on smiling.)

If the images you try to recognize are also consistent, like in lab tests, rules-based “template matching” works well.  Unfortunately, in the real world, criminals avoid walking up to cameras and posing. They wear sunglasses, face various angles, change appearance with age.  (Or they smile maniacally when shooting you.)  As a result, the NYU study found real-world accuracy ranging from 60% to systems that “were in fact not able to recognise any of the subjects.” Bummer. You tend to want zero tolerance when errors lead to detainment and cavity searches.

Facial recognition in photo applications is a different story. Your friends cooperate to show their faces, simplifying things significantly.  Polar Rose claims about 76%  accuracy, which it says compares favorably with Picasa and iPhoto.  76% could be distressing in the cavity-search case, but elegant photo-application interfaces make you happy with what works rather than upset with what doesn’t. (The technology does create privacy concerns, however.)

So what about HTM?  Its inherent robustness suggests HTM could recognize faces better in poor conditions. The problem, though, is that HTM is a classification algorithm. Detecting humans is a two-category problem (people or non-people). But using classification to find Osama-bin Laden is a 6,692,030,277-category problem (or however many people are in your watch list).

A more attainable short-term problem might be to tell members of your family apart.  Or we could try isolating faces in videos and licensing an existing facial recognition system. Otherwise, for now we’ll invest more in areas we feel we can offer stronger differentiation.

What causes feature creep? Conventional wisdom blames companies that don’t listen to customers, and/or customers who choose products with more features. I think it’s more complicated.

Regarding customers, if two products are otherwise identical, it’s rational to prefer the product with more features.  You need high-level innovation to keep customers out of the weeds.  (The iPhone shipped without copy and paste, for crying out loud.)

As for companies, some do listen to customers.  They ask how to improve the product, which leads to feature requests.  Unfortunately, this often means asking customers to predict the value of features they’ve never used.  You can generate valid data asking Hawaiians about snow shovels, but I wouldn’t design products around it.

Moreover, describing future features introduces positive bias. It’s easier to communicate feature benefits quickly than to articulate potential downsides (except maybe price).  Somehow surveys never ask, “what if it was slow and unreliable?” In focus groups it’s best if you can show a working prototype, but there’s also “novelty” bias. Like a first date, first impressions may not predict long-term happiness.

What’s worse, some people think focus groups are a failure if people don’t like their product.  This dangerous mistake can cost millions of dollars.  Your goal is to get the facts.  Clear consensus means a “successful” result, even if they hate your product.  Accurate news is better than good news.

So here’s a nutty suggestion:  do what doctors do. They don’t ask what treatment you’d prefer. They ask what hurts. Similarly, you can probe for pain points that your features could address, rather than ask customers to figure out the answers.  That’s your job. Henry Ford allegedly said, “If I had asked people what they wanted, they would have said faster horses.”   Today, customers requesting features may even have a different problem in mind than you do. (Like the PDA customer who asked for a spreadsheet–to keep his phone numbers.)  Find what customers need rather than ask what they want.

In other words, I’m not dismissing market research; I’m just saying don’t follow it blindly.  There is no substitute for thinking.