False positives and false negatives vs positives and negatives. 85% is meaningless as it stands. If all the errors were false positives it means the majority of us would be arrested on the say-so of a buggy computer.

Face recognition is going to have a place in law enforcement but whether it is ready is another matter.

How can an 85% success rate be considered proven, that means there is a 15% chance of either some random traveller being detained or a known security risk getting past? Not exactly the sort of chance you'd want to take in an airport, considering how bad they are now for overreating.

Lets not get hysterical about this. Realistically, no one would be arrested on the spot by SWAT teams just because the automated system singled them out. Instead it would be layered with existing security- an innocent person falsely flagged by the system would have to, at worst, show a passport and get "wanded" by security- just like what happens when a guard randomly selects someone out of the line for the extended security proceeding.

Its much like security people picking out travellers that look suspicious, just automated. Me, I always get picked out of the line at the airport. It must be the hair, or that I look just vaguely middle eastern enough to trigger a response from the paranoid.

"LogicaCMG achieved an average 85 per cent success during a pilot of facial recognition technology in an unnamed international airport rate in singling people out after matching them against a database of 1,000 wanted people."

So there's an 85% success rate when you have a database of 1,000 people. How much longer is it going to take to compare the live feed against the contents of a real database (which could easily be many thousands of people), and how will that extra time affect the success rate?

Let's just say that 85% success is what goes into production. Now lets look at JFK : over 41 million passengers per year, meaning more than 110 000 a day. If 15% of that needs questioning, it'll make for almost 17 000 people to control every day.

What kind of resources do you need to single out, take aside and securely search and interrogate 17 000 people every single day ? That's 700 people per hour, or 11 people per minute. And the interrogation process is going to last more than a minute. I would guess it takes anywhere from 5 minutes to two and a half hours. Let's just say 30 minutes mean time. You therefor have a security team occupied with one person for 30 minutes. Meanwhile, there's 350 other people being interrogated. And if you delay, you get a backlog, and that means irate customers.

No, 85% is nowhere near acceptable. It has to be brought up to 99% or the system will choke on its own inefficiencies.

I may be over-reacting but that did raise a smile...

...is a worry. Not that false positives will be incarcerated on the say-so of an 85% system, but that vigilance by humans will decrease, leaving all the false negatives free to roam. Though, as now, known criminals will avoid highly-surveilled locations anyway, and it's the "first timers" who are actually dangerous.

I suppose even being able to claim cameras have picked someone as worth following up does help get people off the hook when it comes to claims of discrimination or racial profiling.

"Sorry sir, but it's not *me* choosing you, it's the *camera*. You must look like someone on our database".

Never mind face recognition, how about other applications of this technology?

There ought to exist an automated method for taking a binary program and generating some Source Code which would compile on a machine with the same architecture to give that same binary. In other words, a decompiler. While the output might not match the original Source Code exactly (according to the Church-Turing hypothesis, it need not even be in the same programming language!), the resulting binary *will*, by definition, be identical. This output could even be compiled on a machine with a different architecture; and ought to run faster and more smoothly than the original through an emulation layer.

The mathematics underlying this problem are fundamentally the same as the mathematics underlying face recognition. Machine-language instructions are analogous to vertices, and high-level control structures such as loops and subroutines are like features. Now, which vertex belongs to which feature? And which machine-language instruction belongs to which high-level control structure?

It's no exaggeration to say that the existence of a working decompiler would be one of the greatest breakthroughs in computing to date.

If a program written in one language could be decompiled into another language, this would give software developers the ability to co-operate with one another without the need to have a programming language in common! A developer using C and a developer using BASIC could work together on the same project.

Furthermore, the era when a few cowardly software vendors could hide their Source Code behind the workings of a compiler would at last be over. (Variable and function names, if not preserved for debugging purposes during compilation, would have to be guessed from context ..... so there's still some rôle for human hackers yet). We users would finally have the ability to take our Freedoms One and Three by force, if necessary!

Cuba here I go.....wayhay!

So the recognition drops to 65% for people looking downwards. Like for shifty criminal types, not wanting to be spotted / recognised ?

I was lucky enough to witness a demonstration of this technology by another company in 2000. At the time I was very impressed and the success rate was similar to the system discussed here. However I do specifically remember that it was tested on an airport escalator! Why hasn't this improved in the last few years? My guess is processing power and the high cost of cameras accurate enough to take accurate facial readings.