General Interests

Updated 2010-01-31

Evolution, non-linear dynamics and biological systems. As an engineer one is often forced to work in the realm of closed (and linear) systems with fixed structures, whereas most adaptive systems are open, non-linear, and change their structure according to context (nature) and nurture (genetics). One of the most exciting developments in biology and genetics (I think) is learning more about the role of "junk DNA" (introns), by some researchers called the software of DNA. It probably has many roles, one of them being a regulation machinery for gene expression. Others may be that of hosting genetic leftovers of viral infections (or other genetic parasites) that might speed up evolution if triggered by changes in the environment (as proposed by Greg Bear). The latter is still pure science fiction but leaves a lot of room for interesting research agendas such as explaining why evolution seems to makes "jumps" (punctuated equlibria, PE). The jumps do not seem occur between species but higher up in the phylogenetic hierarchy and are manifested in "transitional fossils". Another fascinating development is new evidence towards epigenetic evolution, where - paradoxically - heritage can be passed on to a new generation partly influenced by the environment and not primarily by Darwinian selection. Pseudogenes are remnants of genes that for some reason have been deactivated and lie formant in the DNA. It has turned out that even if they don't code for proteins, they still have a regulatory function, which proves again we just have begun to scratch the surface of how DNA is actually used and expressed in the cell.

The biologist Carl Woese became famous for discovering an unknown branch of the tree of life, the Archaea, non-bacterial procaryotes. Recently, he and Nigel Goldenfield has suggested so called "horisontal" exchange of genetic material in early micro-organisms as necessary to explain why all (known) organism use the same extremely robust (polymorphic) genetic alphabet AUCG to express proteins. This mechanism of sharing genetical material has been known, but never seriously been investigated as a mechanism to explain the origin of DNA. Backed by simulations, they show that life would have been unable to develop the using the universal principle of DNA, genes and mutations, unless it had started with phase of horizontal genetic sharing between organisms. This is one of the most radical ideas in biology, and might an important step to explaining the origin of life.

Creationsists have used the lack of evidence for intra-species evolution as an argument to disprove evolution, but as usual the reasoning has nothing to do with science. The fact that some people would like "intelligent design" to be taught in schools in the US today is a sad reminder of modern religious ignorance and fundamentalism. Luckily, the Kitzmiller vs. Dover verdict utterly stomps the attempt of a local school board to teach ID on the same footing as evolution. Unlike Sweden, religion is not taught in any form in most American schools, so students are never exposed to different belief systems in a natural "neutral" environment. If science classes would have been kept separate from any influence from religion and ID would be taught as a part of the curriculum on religion, the debate would have been a lot more healthy (and even fun).

After having read Richard Dawkins wonderful book "The Ancestor's Tale" and having read parts of Luca Cavalli-Forza's pioneering work on the origins of human genes, I found out about the joint National Geographic and IBM Genographic Project" headed by Spencer Wells. Not expecting that my genes will provide any scientific milestones, I nevertheless decided to participate in the genetic mapping project by providing samples of my DNA. I found it strange that the mapping of genes has been condemned by some people as being "racist". In fact the very definition of race is not well-defined and is not part of the standard evolutionary phylogenetic tree and most evolutionary biologists agree that it's mostly useful nomination of some genetic traits.

One of the most interesting trends is the realization that mathematics and technology must go hand in hand with biology to tackle interesting problems. In 2000 the Institute for Systems Biology was cofounded by Leroy Hood and has set out attack problems like complex interactions (in a broad sense) with statistical and topological methods. Similar work is also underway in Japan at the Kitano Symbiotic Systems Project where the focus is both on the design of actual artefacts such as humanoid robots as well as a system-theoretical understanding of biology. It will probably be a long time before we understand the principles of lower-order (animals) and higher-order (animals and humans) cognition and the more elusive property called consciousness. Current work still involves a good deal of speculation and philosophy and personally I am attracted to the views of people like Patricia and Paul Churchland and Daniel Dennett. When it comes to building models and artefacts of real systems the ideas of Gerald Edelman are interesting: We should make model "brains" and study how the interact and change their structure with the environment. The Neuroscience Institute in San Diego is doing a lot of interesting research in this direction, as is also the Koch laboratory where the focus is on both vison, consciousness and biophysics.

Computer vision as a research field has made remarkable progress in fields like tomography and 3D reconstruction and can provide a set of powerful tools for low-level pattern recognition and basic object detection and recognition. It still, however, far from the goals set out by David Marr and others and has made very modest progress in the sense of combining cognition and vision. I believe that breakthroughs will only only when the internal complexity of systems have reached a certain level. Another reason for the slow progress in the field is that researchers are still focused on algorithms and methods as opposed to building systems with sufficiently complex interactions with the environment and internal learning structures. The book "Ways of seeing" by Pierre Jacob and Marc Jeannerod gives some perspective on computer vision from the standpoint of cognitive neuroscience and philosophy.

A dream I have is to work with "closing the loop" in developing an autonomous robotics application complete with transducers (active and passive), learning structures and locomotion. Robotics is the ultimate challenge for computer vision and a major challenge is to try use learning to adapt robotic behaviour from past mistakes and sensor input, as well as try to figure out how to prewire behaviours and learning structures. The first step is probably instinctive reactive behaviour, and the next challenge is to try to understand how to accumulate knowledge from experience for long term planning. It took evolution millions of years to achieve this in humans, so we should not expect rapid progress. It finally seems that the vision community is slowly learning from the lack of progress in robotics and autonomous systems. The old view that complex representations begets cognition and action has now been replaced with a more flexible one where percepts are directly linked to action and where symbolic representations is a separate faculty, more related to language and high-order consciousness. This is certainly closer to the biological "truth" and it will be interesting to see if this will speed up the progress in designing robots and even software systems based on sensory-visual input.

Equally challenging is the problem of locomotion, in Japan there is a lot of interest in humanoid bi-legged robots. Making a robot walk and avoiding obstacles is very challenging problem that will require decades of research before we see robots like in the film "The Bicentiennal man" or "I robot". More interesting perhaps are smaller insect lite robotos with legs (or rather "whegs", wheel-legs") from a a group at Case Western University. This film shows a wheg robot in action. Some companies such as Irobot are developing similar robots, and Swedish company Rotundus has developed a unique spherical robot that is propelled by a swinging pendelum. Interestingly, in agriculture there is a lot of interest in using computer vision for precision farming, one example is the research mechanical weeding at Halmstad University which can be used to eliminate the use of dangerous pesticides in eco-farming. One of the most innovative attempts to construct robots that are energy self sufficient are the Ecobot and Slugbots from the university of Bristol, work done at the IAS laboratory. The EcobotI and its successor EcobotII are robots that extracs its energy directly from the enviroment from so called Microbial Fuel Cells (MFC:s). In principle this gives the robot energy autonomy and thus needs no batteries. The slugbot was designed to collect slugs and use these to generate its own power. The slugs were fermented in a separate off-vehicle unit to generate electricity in a methane fuel cell. The EcobotII, in contrast, can directly "digest" flies and rotten fruit using a MFC fuel cell: at the anode bacteria act as a catalyst to break down the fruit and flies and at the cathode free oxygen closes the circuit. Robots like these might be perfectly adapted for agriculture for opto-mechanical weeding or even tasks like grape-picking. For those of us who like Sci-Fi, in the film Runaway robots like these played an integral part of the story, where a policeman played by Tom Selleck is assigned to a unit that specializes in hunting down and fixing robots that are malfunctioning.

The autonomous vacuum-cleaner Trilobite is a great example of a consumer product where robotics techniques have been put to use. It was largely thanks to algorithms developed at Communicator and later at KTH that Electrolux got the robot to work. Some companies like SAAB are interested in unmanned aerial vehicles (UAV:s) for which there are a number of interesting problems to be solved. One is that of autonomous path planning and another is Sense and Avoid or how to avoid colliding with other aircraft using data from active and passive sensors when TCAS or transponder based systems are not available. I'm currently working with bearings-only tracking (also called passive ranging), a challenging problem that requires flexible modeling approaches (such as Extended Kalman filters, Unscented Kalman filters or Particle Filters) and suitable coordinate systems such as Modified Spherical Coordinates {φ,φ´cos(θ),θ,θ´,r´/r,1/r}. Although advanced UAV:s have been used primarily in military operations, they are also beginning to be used in civilian applications like agriculture, fire fighting, disaster monitoring, power line monitoring (pdf), chemical sensing and Oil spill detection.

Autonomous systems are also being used for underwater explorations, the Puma and Jaguar robots from the University of Maryland are being used to detect (Jaguar) and explore (Jaguar) of the Gakkel ridge in the North Atlantic.

Work related stuff

updated 8-January-2006

Image understanding, pattern recognition, classification, signal processing. While I worked with remote sensing I implemented a region growing algorithm based on iterative merging of smaller regions using their statistical properties. Here is a brief description of the algorithm, and a visual caption of the results for optical and radar imagery. The source code (C++) is available available here, together with application examples, test images and scripts. The region growing implemented here should be compared to the "Mean Shift" algorithm now widely used in both real-time and off-line systems. A good background description and source code can be found here. The imlib command line programs are intended as a help for people who want to develop their own image processing applications, without having to bother about image formats. It handles 8-bit binary PGM and PPM formats, and 8,16 bit and 32 bit floating point BIL (Bits Interleaved By Line, the latter is frequently used in remote sensing), and strip-wise I/O for large images. You can use the netpbm or ImageQuick" packages to convert to and from other commonly used image formats. I've tried to design it as a toolbox with the intent to put most stuff in libraries and make the actual application code as small and easy to understand as possible. Example applications implemented so far are wavelet-regularised iterative Lucy-Richardson/successive approximations and Wiener deconvolution algorithms, region-based region segmentation/merging, inertia tensor local orientation,Lucas-Kanade optical flow. Look here for more toolbox features. In the imlib code I use a nifty matrix library in C++ by Robert Davies. The package is included in the imlib package; should you have any problems compiling or configuring it, let me know. You find more on the matrix library and online documentation on his website. Most of the toolbox applications also work for windows platforms, and project files are supplied for Visual Studio/C++ (v. 6.0, with service pack 4 or better). And please note that it is a _strictly_ script/command-line based toolkit, but with modern graphic toolkits such as Qt or Tcl/Tk it's possible to make a graphical user interface without too much effort. I prefer to avoid GUI:s for three reasons: (1) they are not portable, (2) the few alternatives around are clumsy to use, and (3) they do not allow for automation through scripting. Remember: "In the beginning was the command line" :-).

An interesting software package that uses the generic template programming paradigm in C++ is the VIGRA package (Vision with Generic Algorithms) available from the university of Hamburg. Finally (but most likely not the final word!) there is the VXL system, a spin-off from the Image Understanding Environment and TargetJr projects. It contains a wealth of computer vision source code, and now has a template-based C++ interface to many existing numerical analysis and CV algoritms. The Boost framework has been suggested as a part of the updated C++ standard, and has matrix container classes. Whether there will be an interface to LAPACK or some other linear algebra package seems unclear at the moment. A wealth of information on computer vision can be found on the CVonline site, and at CMU there is a nice portal for many computer vision related research groups the world, the computer vision home page.

Other interests:

Cosmology, gravitation, astronomy and neural information processing.. The latest search for the holy grail of quantum mechanics and classical gravity is called Loop Quantum Gravity (LQG). A good review can be found here. Renate Loll and her group in Utrech has been investigating how to derive quantum foam "from scratch" using a causal triangulation procedure. Amazingly the numerical results indicate that with very few assumptions a 4D structure emerges at large scales.
The currently established and working descriptions of gravity give very accurate descriptions of the physical law, but say nothing of the origin. In a recent beautiful paper, Erik Verlinde at the University of Amsterdam has used straightforward concepts of Entropy and the thermodynamics of black holes (the Holographic principle that goes back to Bekenstein, Hawking and others) to yield a straightforward explanation of Newtons law. It shows that gravity is a system property and seems to work in both Relativistic and classical domains. The result is brand new but has been met with a lot of enthusisam in the physics community, and was recently featured in New Scientist.
There are a number of cracks in standard Big Bang cosmology, one of them is the notorius issue of CDM (Cold Dark Matter). Modified Newtonian dynamics MOND maybe a way to get around this. Some authors even believe that the big bang must be replaced with something else. Maybe general relativity must be modified to take into consideration the effects of plasma physics, after all plasma accounts for 99 % of all matter in the universe. An early attempt of a cosmological model based on plasma physics was made by Hannes Alfven. This model is today of historical interest only, but shows that many important phenomena have been left out in standard GR cosmology, which is based on a simplified model of a perfect newtonian fluid. The situation is the same as in plasma physics itself: since the full plasma equations are highly non-linear, a lot of plasma physics is done using the MHD approximation, which is known to model only parts of the physics involved.