Three strikes

The idea that Peter Mandelson should even have the title of "Lord" just shows how corrupt the British honours system really is. He's been forced to resign on a couple of occasions previously due to various corruption scandals, but managed to claw his way back into high office in some dubious way or other. Mandelson represents the worst of British politics, and always reminds me of the fictional character Grima Wormtongue.

If Mandelson's "three strikes" rule gets implemented the logical outcome is that a much larger fraction of traffic on the internet will be encrypted. The biggest problem with forcing ISPs to police copyrights is that it's going to be hard for them to establish exactly what content is moving through their system at any point in time. Once encryption comes into play even if the "deep packets" are inspected they may be meaningless, and the only policing option is then to become a user of the P2P networks and try to actively infiltrate social networks, which can be countered by peer ratings systems and other trust based mechanisms. The whole thing becomes a needless quagmire of spying.

Viewed from a scheming Mandelsonian perspective maybe this is just part of a larger plan. The ISPs have made clear that trying to enforce copyrights in the manner proposed isn't going to work - at least not using the internet in its current form. But maybe eliminating net neutrality can help. If you can separate internet provision via ISPs into two channels - an "official" channel providing government services and officially approved news, and then an "unofficial" channel for everything else. That way you can arbitrarily cut off people's access to the wider internet, but still leverage its capabilities to for state control and propaganda purposes, whilst still remaining in line with EU regulations on internet access. ISP based copyright policing also provides a convenient mechanism by which political blogs can be suspended, if you don't like what they're saying. All you need do is accuse them of repeatedly linking to illegally copied or leaked information.

BigDog on holiday

When off duty, BigDog takes a relaxing stroll on the beach.

Minoru stereo

Stereo disparities obtained using a combined binary descriptor and horizontal gradient matching strategy using the Minoru 3D webcam. This system also uses anticorrelation to help eliminate false matches.

Alt Party

Sounds like fun, and Cray supercomputer demos would be awesome.

Loebner

The Loebner prize for 2009.



The mainstream view on this is that it's not regarded as being a very useful test of AI, although it does stick reasonably closely to the one described by Alan Turing in Computing Machinery and Intelligence. Natural language understanding and commonsense reasoning are amongst the hardest unsolved problems in AI, and as the man himself admits in the video even after a couple of decades of such contests not a great deal of progress seems to have been made.

Probably the biggest criticism is that this type of Turing Test, performed over a five minute period, encourages entrants to produce programs which try to fake the superficial appearance of intelligence rather than actually producing something which has some intelligence. So over a short period of time using some clever statistical tricks based upon analysis of real human diologues it may be possible to fool someone that the computer is a human, but even this seems hard to achieve.

I think Ray Kurzweil's prediction is that by 2019 there will be some question mark over whether this sort of Turing Test has been passed or not, with a more decisive pass by 2029. However, making progress on this problem involves more than simply the assumption of the continuation of Moore's law. I expect that the reason why the test remains difficult isn't just because the conversational databases aren't big enough or fast enough, and that there are conceptual breakthroughs which will need to be made. It has been claimed by some that improvements in algorithms are also following Moore's law, but given the incredibly slow progress in some of the algorithms with which I'm familiar I'm not terribly convinced by that claim.

Scanning laser rangefinders

I've been keeping an eye on the prices of scanning laser rangefinders for the last five years or so. Many people regard these devices as the key enabling technology, whose cost will decide when most consumer level utility robots with sensible navigation performance become possible. Reviewing the prices today it looks as if very little has changed in the last few years. The least expensive laser system still costs in the region of £1K including taxes, with prices ranging between £1-4K. Taking inflation into account perhaps prices have fallen slightly, but on the current trajectory it will still be many years before mass market applications are possible. The price/performance trajectory doesn't appear to be following Moore's law.

Economies of scale may reduce the costs, and probably the biggest potential market for these sensors is in the automotive industry.

But lasers also have their limitations and safety risks, and in the mean time the prospects for camera based navigation seem to be far better. The amount of computation needed to turn images into maps is much greater than the same operation using lasers, but Moore's law has been acting aggressively to make this a near term prospect and processors with tens or hundreds of cores will only make the task easier.

Unusual robotics petition

The recent Alan Turing result has given me a little more hope that petitioning as an activity can, on occasion, result in some kind of action being taken, so I was browsing through the list of technology related petitions and found this one.

"We the undersigned petition the Prime Minister to Increase funding for Robotics with money from companies that do not provide employees with diverse and enjoyable work."

The full description is:

"Places of work that do not provide employees with diverse and enjoyable work such as some wherehouses, post offices, cleaning firms, olds peoples homes, carpet stores and many, many others... should have to pay a tax if its decided that the work could be done by robots. The tax would then help fund government scientists to produce robots to do the work. The television program Tomorrows World promissed that all repetative labourious tasks would be done by robots. But while companies can still get cheap human labour to carry on with these dull taskings without being taxed for it the robots will never come."

I didn't sign the petition, primarily because I think market forces will do this particular job once the robotics becomes cheap enough and sufficiently capable. There might be a good case for increased funding in the almost non-existant UK robotics research establishment, but as I've blogged about previously I think that any additional research should be focussed upon tasks or key enabling technologies which are likely to lead to useful applications, and not simply the production of indulgences such as the robot bees or other electromechanical esoterica. Also, with the economy in its current condition it seems unlikely that placing an extra tax burden upon companies is going to have a very positive effect, and might even end up with the public developing a resentful attitude towards the targeted research.

Still, it's nice to see that at least someone out there is thinking about the possibility of using robots within sectors of the economy from which they are presently entirely absent, and that some jobs may not be too far away from the expanding scope of automation.

Hofstadter's lists

Douglas Hofstadter has accumulated some amusing lists. You can skip over the initial 15 minutes of dreary introduction.



What I think this shows is something similar to the society of mind idea, where multiple concepts having similar meanings are competing with each other to become the eventual output, and in a few circumstances you end up with some peculiar combination of the competitors. In neural terms you can think of it as networks having excitatory and inhibitary connections.

Robot risks

A study of the possible security and privacy risks posed by current telerobots can be found here. This is the first article which I've read which specifically deals with risks of the use of telerobots in a domestic environment, although I have previously read papers which refered to these kinds of risks as an afterthought within the conclusions.

Singularitarians like to talk about risks such as how AIs might somehow magically turn the universe into paperclips, but this is not a sensible discussion of technological risks and instead I think has more to do with entertaining the audience or rallying concerned supporters around a particular institution. There certainly will be risks associated with robotics and AI though, although they're of a far less glamourous nature than the grandiose visions promulgated by a certain contingency of futurist thinkers.

I doubt that the telerobots reviewed in the article pose any serious risk at present. At worst possibly such robots might be used to survey a house for any possessions of value prior to a burglary. They could also be used to trip people up, but with the current generation of telerobots achieving this feat would require not inconsiderable skill and timing on the part of the teleoperator. Bigger problems may arrise in the not too distant future though. The logical evolution of these small telerobots is to make them somewhat larger and to add one or more manipulators which can be used to carry or move objects from one location to another (the Robosapien could be said to have manipulators, but its ability to carry objects is very limited). Even if the level of AI ability remains at something close to zero a large robot with a manipulator could be used to cause damage if its internet security can be compromised - and I think it will be fair to assume that most mobile robots will be in some way connected to the internet.

How can such risks be mitigated? There are a few possibilities.

1. Ensure that at the very minimum the core communications functions of the robot are based upon open source code, so that security flaws can be found and fixed as quickly as possible.

2. Have a method of delivering software updates which is hard to infiltrate. Preferably the delivery mechanism should be automatic, although this itself would increase risks of interception or masquerading.

3. Limit the properties of the manipulator/s. The robot arm should be able to move no faster than a certain rate, and be restricted to moving in ways which are specific to tasks within the design specification. Limitations could be built in at the electrical or mechanical levels of design, and preferably should be separate from the software control signals (kinematic checksums?). Restricting the rate of movement would for example prevent the robot from throwing an object, and would also help to prevent any unsafe behavior arrising from inadvertant software bugs.

4. Fit the robot with an industry standard emergency stop mechanism. This is routine in industrial robotics, but seems to be lacking on the current generation of telerobots - probably because they're small and unlikely to be capable of causing much damage to anything.

5. Devise safety guidelines/rules which manufacturers can follow. Perhaps some regulatory body should be formed to oversee this, or an existing organisation designated to it.

6. Use standard, openly published protocols for internet based communication with the robot. Although I expect there will be a lot of innovation and diversity within the onboard sensors and control systems the software which acts as a communications interface between the robot and the internet should be as standard as possible.

Ubuntu Software Center

Listening to the Ubuntu UK podcast one thing really caught my attention 15 minutes into it. They're talking about the Software Center, which is a new feature in the Karmic release. I've tried this myself in the Karmic beta, and it's a good idea to try to unify what was previously a couple of different programs in order to simplify the user interface and add extra abilities, such as screen shots and user reviews. According to the Canonical guy after version 3 they will start selling "commercial software", which is confirmed on the Wiki.

It's unclear what "commercial software" means in this context, other than you'll pay for it. Usually this means closed source, proprietary software where the source code is not available for independent scrutiny. If Canonical are planning to introduce blatantly proprietary applications into what is (very) arguably the flagship free operating system this will be quite a turnaround for them, and it would leave a doorway clearly wide open to all the worst aspects of software which runs on Windows, such as:

• Spyware
• Disableware
• Adware
• Licence serial numbers (a real pet hate of mine)
  
• Freeware containing an undisclosed malware payload
• Ransomware
• Digital Restrictions Management (DRM)
  
• EULAs which give the end user no rights whatever
  
• Forced software updates
• Nag screens

In short most of the reasons why I really wanted to get away from using Windows. Also increased use of closed source software inevitably means computer viruses and the abject horror of the "security circus".

Of course there is already a certain amount of proprietary software within Ubuntu, in the form of drivers and codecs (the "bad" and "ugly" packages), but this had so far been considered a necessary evil and the need for these has been slowly reducing over time.

So we'll just have to wait and see what happens with the Ubuntu Software Center. In the worst case it could spell the end for the popularity of that particular distro, but it doesn't necessarily have to be this way. One thing which would be a good addition, facilitating commerce but avoiding the negative aspects of proprietaryness, would be to add some feature to the user interface which allows you to make a donation to your favourite FOSS projects. If donating was very easy and simple I expect that more people would do it, and Canonical could take some percentage to sustain their own operations. A smart way to do it might be to monitor how many times an application is used, and then suggest donations to those projects, or have a generic donation value which is then automatically split up depending upon application usage history.

Embedded stereo correspondence

A video showing the results of the stereo correspondence algorithm embedded on the Surveyor SVS. This is running on two Blackfin DSPs with synchronised frame capture. Larger red spots correspond to objects closer to the cameras. The big metal object is the left shoulder of the GROK2 robot.



There is some "pollution" where matching of horizontally oriented features causes bad matches in the sorrounding region of the image, but most of the time the spot sizes are inversely proportional to range. There is always a certain amount of noise in the matching, but if the features are converted into sensor models and used to update an occupancy grid then this tends to cancel out with repeated overlapping observations.

Ordinarily horizontally oriented features can't be matched, at least by sparse feature based correspondence methods, but it does seem to be a useful heuristic to apply similar disparities to horizontally oriented features which are in the neighbourhood of matched vertically oriented ones. Even though it sometimes produces errors, this helps to provide more information about the structure of the environment which can be useful when creating maps or avoiding obstacles.

For anyone interested in the algorithm the code, written in C, can be seen here.

Will the conservatives scrap ID cards?

The conservative party claim that they're not in favour of the proposed ID card scheme, and wish to scrap it according to a report on their web site. Perhaps this is a genuine claim, but politicians do have a habit of reversing their pre-election ambitions.

Stepping into the time machine and going back to the news of 1997 shows just what a merry-go-round political system we have in Britain.

"Tony Blair has unceremoniously junked all Conservative plans to introduce an identity card system, proposed as a measure to help the police, reduce crime and bolster law and order.

The Independent has been told that there is no question whatsoever of any move being made to take forward the Tory plans. 'They are stone dead,' one source said."

There's another reference to the conservatives enthusiasm for ID here. Notice how interchangable the terms Conservative and Labour are? In the olden days - meaning more than ten years ago - it would have been difficult for the average reader to find similarities between political policies like this, unless you were a political historian, had a long memory (which most people don't) or if you were prepared to take time out to visit a library meticulously scan through old newspapers. The web and modern search engines make looking for historical precedents and quotations or slogans much simpler, in spite of their ephemeral nature. It does however appear that the 1994 green paper on ID cards, produced by the John Major administration is lost in the mists of time, or at least as far as the internet is concerned.

Having lived through administrations of both persuasions it seems to me that politicians generally, regardless of their party affiliations, would like to have an ID card system, and have attempted repeatedly to introduce one but without much success thus far. So it wouldn't surprise me at all if a year or two into the next administration Mr Cameron experiences a sudden change of "vision", probably shortly after another moral panic or terror alert.

Stairs and Bees

IEEE Spectrum reports on some novel approaches to the problem of having a robot climb stairs. I've not seen robots going up stairs in quite the same way as this previously, and probably the simplest method would be the two section tracked approach used by the Pacbots.

The only snag with these novel concepts is that if you really want the robot to do anything useful within an environment which contains stairs, to the degree that people go from saying "that's kinda cool" to "where can I buy one of those?", it's probably going to need to be able to carry objects at least as large as a cup or a bottle. With the designs shown here it seems very unclear how that could be achieved, and locomotion by hopping is going to introduce a lot more problems from a materials transport perspective.

In my opinion a lot of the academic robotics research suffers from lack of focus, such as the recent bee robots. Ideally we should be looking to develop systems which are pushing the state of the art but also which might have commercial spinoff applications which are useful or beneficial to the world at large. The range of possible applications for bee robots seems to be incredibly limited, and if I were in charge of allocating a $10 million grant I could think of much better ways to spend the money which might really provide some payback in terms of spinoffs or startup technology companies. Perhaps bee robots could be used to monitor traffic, but this just seems like an absurdly inefficient way to achieve that goal, given the alternatives.

Are we headed for a lone twig monoculture?

Since Homo sapiens have not been on the scene for very long, and our last hominid cousins only died out very recently, maybe we are the last lone twig of our particular lineage.

However, I think that the prospects for future speciation of humans (or "forking the project") are excellent. It seems likely that in the very near future we will be able to effectively select and modify our own genetic makeup at will. How the genome is altered will depend upon what people find desirable or undesirable, and desires vary between cultures and sub-cultures. Even in the unlikely event that all cultures select their genetic attributes in the same way, if humans are eventually able to construct sustainable colonies beyond the home planet then classic speciation seems certain.

Calling all brain hackers

An interesting software challenge to reverse engineer "the cables of the brain", grouping them into a few major bundles.

"The brain has an estimated 500 regions with over 150,000 km of fibers. These fibers are organized into fiber tracts which appear in atlases used by surgeons and neuroscientists. In this competition we use Magnetic Resonance imaging (MRI) based diffusion weighted techniques to image the diffusion 'pipes' representing bundles of cortical tissue non-invasively on live humans. Using advanced MRI based diffusion imaging (Diffusion Spectrum Imaging) and image reconstruction techniques (generalized Q-Ball) we created High Definition Fiber Tracking (HDFT) data sets for three individuals. The data sets include about 300,000 fiber streamlines per person."

Opening doors

Using Twitter: Ok in Iran. Not ok in the US.

An amusing show of double standards. The guy may have unconventional beliefs, and I'm not exactly sure what the G20 protesters were protesting about, but providing that he was protesting peacefully and not pushing wheely bins randomly around in the street that should be ok. Arresting someone on the grounds that they used Twitter to communicate about something they saw happening during a protest seems silly.

Pipelining human intelligence

The internet provides new opportunities for remote observation and remote actions. On tasks where automatic systems currently perform poorly raw human intelligence when combined with the large scale of the internet can be packaged and sold as a service. The most obvious example of this would be telerobotics, which seems likely to be a growth industry in the next decade. But even legacy systems can benefit from pipelined human intelligence.

A web site called Internet Eyes allows users (or players, depending upon how you view it) to log into CCTV cameras and spot then report crimes being committed. In principle this kind of system is a good thing, permitting local communities to police their own neighbourhoods. But the way that this service is set up, using a combination of anyonymity and direct financial incentives rather than encouraging transparency and a sense of civic responsibility, will produce an unvirtuous cycle which leads to increasingly trivial activities being classified as crimes. Often minor offences are dealt with using fines, which are typically paid either to local police authorites or local government organisations, so there's a financial incentive to catch certain kinds of criminal and to increase the scope of misdeeds which can incur fines where the amount of the fine exceeds the cost of administering or enforcing the law. Combine this with free, or nearly free, human intelligence delivered via the internet and the administration costs fall dramatically, making the levying of fines for minor offenses far more profitable than was previously the case.

It is by now well established that when people believe that thay're anonymous they may do things which they wouldn't otherwise do if their activities were exposed to greater scrutiny by their peers. The anonymity provided to users of services like this means that there is no degree of accountability in the system. If some anonymous person accuses me of commiting some crime on a CCTV camera I have no idea who they are and can't reply to them, perhaps providing more information about the larger context within which my actions took place. There's a transparency/democratic defecit, or an equiveillance imbalance.

Anyone who has done any amount of AI, particularly with regard to interpreting video or static images, knows how fundamentally ambiguous these can be - especially when the field of view is narrow. If you see someone knocking someone else to the ground are they committing an assault, or carrying out a citizen's arrest on a handbag thief? From the video it may be very unclear, and you might initially jump to the wrong conclusion.

I see no reason why services such as Internet Eyes wouldn't become popular, especially in recessionary times when there may be many people who are time rich and cash poor, and who don't mind having a video feed in the corner of their computer screen. But I think the financial incentives are going to bring out the worst aspects of human nature, with the logical outcome being more and more commonplace human activities being classified as "crimes" for which fines are levyable - i.e. society becomes increasingly controlled all the way down to the most trivial activities. Online systems such as Second Life demonstrate admirably, at least as far as I've observed, that people will do almost anything for money, no matter how ridiculous or tedious.

There's a much better way to set up a system like this. There should be no financial incentives involved, but instead the site should encourage individuals to "police your own local community". Individuals have a natural vested interest in keeping their own back yards in an orderly fashion, and not much interest in what goes on further afield. Also the people watching the cameras should not be anonymous. There should be some part of the web site where you can see who is watching which camera. This means that there would be a natural balance of power between watchers and watched - equiveillance - and that if an accusation is made you know who is doing the accusing. Effectively it's just an electronic version of the ancient forms of personal accountability occuring within small tribes or villages, where everyone knows what everyone else is doing.

Just as with other kinds of technology, I think that working towards balance of power type situations will lead to healthier, more productive and more democratic societies. Where power imbalances arrise - and remember that it has never been more true that "information is power" - there is the opportunity for authoritarianism, despotism, and very unhealthy kinds of society.

Alan Turing's neural nets

I was pleased to see that Turing did get his apology after all. Another less known and far less newsworthy discovery is that Turing's interest in simulating neural networks dates back to at least 1946. In a letter to William Ross Ashby he proposes using the Automatic Computing Engine (ACE) to simulate a neural network.

"I am more interested in the possibility of producing models of the action of the brain than in the applications to practical computing...

...although the brain may in fact operate by changing its neuron circuits by the growth of axons and dendrites, we could nevertheless make a model, within the ACE, in which this possibility was allowed for, but in which the actual construction of the ACE did not alter, but only the remembered data"

Unfortunately, Turing's 1948 paper on Unorganised Machines was apparently dismissed by his supervisor, Charles Galton Darwin (grandson of the more famous Charles Darwin), as a smudgy "schoolboy essay".

You can find out more about Turing's neural nets here. Whether he ever did get to try them out on the ACE as he proposed seems doubtful, since the very early digital computers at that time would have been used for far more important jobs such as producing trigonometric tables or calculating the trajectories of artillery shells.

Skeptoid on the moon

Presumably alongside the Lancaster bomber and the secret Nazi moon bases.

The almighty cloud computer

...whose merest operational parameters we are not worthy to calculate.

Is Google a god? Perhaps not, but it's probably the closest thing to a divine entity which we've thus far created. Here are a few reasons why.

Size matters

Processing the GAC-80K corpus to extract scalar terms used to construct semantic axes, it appears that size really does matter. Here are the scalar words extracted, ranked in order of frequency.

1.00 small
1.00 large
0.81 good
0.72 cold
0.72 big
0.63 hot
0.63 fast
0.54 old
0.45 similar
0.45 popular
0.45 nice
0.45 little
0.45 hard
0.36 smart
0.36 long
0.36 far
0.36 common
0.27 young
0.27 well
0.27 useful
...

So the things which people are most commonly concerned about are related to size, temperature, speed, age and popularity.

Plotting concepts along various axes, known as semantic differentials, can be useful in order to help compare different types of things, and also to make a guess about what properties new concepts might posess. So if I hear a new word which I don't know the meaning of I can compare it to similar sounding words which I've heard previously (or an entirely novel sounding word used within a familiar sounding context) and make an initial guess about what properties it might have - i.e. assign it an estimated position within semantic space. In this way previous knowledge can be used to help acquire new knowledge more efficiently than would otherwise be the case. Intuition would suggest that the mechanism for acquiring new linguistic concepts is fairly simple and not very far from the level of the fundamental operations of speech and speech sounds - hence many jokes rely upon confusions between similar sounding words, freudian slips, or familiar concepts placed unexpectedly within unfamiliar contexts (similar to surrealism). Also the classical proximal zone of development would make some sense within a semantic space kind of representation, as a region of maximum possible interpolation around the current state.