BmiLinks

not sure where to best put this but looks sorta interesting: eeg.pl

• dbacl - a digramic Bayesian classifier -- The dbacl project consist of a set of lightweight UNIX/POSIX utilities which can be used, either directly or in shell scripts, to classify text documents automatically, according to Bayesian statistical principles. dbacl(1) is also the name of the core utility.
• Reinforcement Learning - Sutton and Barto -- Reinforcement learning are solutions devised to learn by exploation - without a teacher signal.
• Jurgen's Rnn the interesting thing about these networks is that they have infinite memory. (or long short term memory - they do not easily forget due to integrator in cell). There is a lot of stuff on this page!
• 022105 - Tomaso Poggio came to Duke. His talk was not very informative, despite his status in the field. He has modeled a section of the mamallian visual system & claims that it performs well in object classification and recognition tasks. might want to look here: [1]
• Complexity measures of a time-series by Bruce Land. Excellent and well-cited summary of many techniques... *also* see Representation of High-dimensional systems with parallel projection The idea is really simple. each dimension is indexed along the x-axis (abcissa), and each value is a point along the y-axis (ordinate), and each point hence becomes an jagged, (colored?) line.

• Densities, Bayes and MCMC from oxford uk; mainly concerns diffusion MRI (i think). want to estimate the probability of observing a set of parameters given the model (M) and data (Y): The monte-carlo idea is to decide on a M sample over (w). the markov chain comes in by placing more samples in regions of parameter space where the parpameters are more probable.
• OpenBUGS implements a solver for MCMC models & has a nice interface, scripting language etc. there is a movie showing how to use it here: http://www.statslab.cam.ac.uk/~krice/winbugsthemovie.html

KYJY7n <a href="http://hsvhirwjruyv.com/">hsvhirwjruyv</a>, [url=http://gfepuqmjfpoz.com/]gfepuqmjfpoz[/url], [link=http://dhwqivewaapd.com/]dhwqivewaapd[/link], http://jrpfqumetlqo.com/

• comments on the use of NLPCA - in the course of discussing why another author's denouncement of NLPCA in climate models is neccesarily flawed, this article gives a good, terse description of how to properly employ NLPCA and make sure the fits are proper (the test error/variance explained should be lower than the fit error!) they also discuss the "the arc-length parameterisation is a convenient parameterisation because it is designed explicitly to measure the density of points along the approximation curve in the state space" sounds good, would have to look this algorithm up.
• Description of Kohonen's Self-Organizing Map - a good introduction to the nature and use of self-organizing maps. Other, more recent and advanced algorithms exist Nonparametric topographic density estimation, but the initial conception is simple and workable - the SOM is a ANN where each unit is charaterized by a vector (with the same dimension as the input space). During training, the unit with the best match to the data (in terms of a distance metric, i.e. euclidean) & this units nearest neighbors are simultaneously updated (or vectorally moved toward the weight / addition i guess) based on a smoothing/generalization function that starts very large and becomes smaller as the training proceeds. A SOM can be used to compress high-dimensional data into a smaller dimension while preserving topology.
• Projection pursuit is used to estimate high-dimensional probability density functions by projecting onto independent axes, smoothing the projections, and summing the one-dimensional fits (so far as I can tell).
• Stanford Linear Acceleartor 1981 publication - good paper on projection pursuit regression.
• Incremental Online Learning in high dimensions -
  • details the locally weighted projection regression algorithm (LPWR). Paper is great, very well written! They use the LPWR algorithm to model the inverse dynamics of their 7DOF hydraulically-actuated gripper arm. That is, they applied random torques while recording the resulting accelerations, velocities, and angles, then fit a function to predict torques from these variables. The robot was compliant and not very well modeled with a rigid body model, though they tried this. The resulting LPWR generated model was 27 to 7, predicted torques. The control system uses this functional approximation to compute torques from desired trajectories, i think. The desired trajectories are generated using spline-smoothing ?? and the control system is adaptive in addition to the LPWR approximation being adaptive.
  • The core of the LPWR is partial-least squares regression / progression pursuit, coupled with gaussian kernels and a distance metric (just a matrix) learned via constrained gradient descent with cross-validation. The partial least squares (PLS) appears to be very popular in many fields, and there are an number of ways of computing it. Distance metric can expand without limit, and overlap freely. Local models are added based on MSE, i think, and model adding stops when the space is well covered.
  • I think this technique is very powerful - you separate the the function evaluation from the error minimization, to avoid the problem of ambiguous causes. Instead, when applying the LPWR to the robot, the torques cause the angles and accelerations -> but you invert this relationship: want to control the torques given trajectory. Of course, the whole function approximation is stationary in time - the p/v/a is sufficient to describe the state and the required torques. Does the brain work in the same way? do random things, observe consequences, work in consequence space and invert ?? e.g. i contracted my bicep and it caused my hand to move to the face; now I want my hand to move to my face again, what caused that? Need reverse memory... or something. Hmm. let's go back to conditional learning: if any animal does an action, and subsequently it is rewarded, it will do that action again. if this is conditional on a need, then that action will be performed only when needed.. when habitual, the action will be performed no matter what.. this is the nature of all animals, i think, and corresponds to rienforcement learning? but how? I suppose it's all about memory, and assigning credit where credit is due. the same problem is dealt with rienforcement learning. and yet things like motor learning seem so far out of this paradigm - they are goal-directed and minimize some sort of error. eh, not really. Clementine is operating on the conditioned response now - has little in the way of error. but gradually this will be built; with humans, it is built very quickly by reuse of existing modes. or conciousness.

ok, how about this: nothing has to go back in time. when you want something, you mess around thinking of what might have caused that in the past - e.g. test hypothesis - and select the best via some sort competition.

somehow, in here, a hierarchy is set up. hierarchy of causes. and habits. in humans it is quite dramatically large!

other shit to think about: STDP, precise spiking in the cortex. where statistical learning is heading.(??!!) the difference between error-based learning (feedback, human) and reinforcement learning

• A much more recent version of the paper (though significantly different in the examples) been published in the Journal of Neural Computation.
• There is also a tutorial on the LPWR software. I really want to understand how this can be applied to the impressive motor learning they demonstrate on their site - http://www-clmc.usc.edu/ (@ Univerity of Southern California (directed by Stefan Schaal)).
• The Kernel Trick excellent description (and style!). also see the book on probability on his page Oph2MY <a href="http://tbolegtdmgvk.com/">tbolegtdmgvk</a>, [url=http://andaoutgaoen.com/]andaoutgaoen[/url], [link=http://bragcvyrtmuj.com/]bragcvyrtmuj[/link], http://hetswkaffwlm.com/

• Mitsuo Kawato's homepage
• Statistical Learning for humaniod robots the dynamic brain project - see Kawato.

Thoughts about the task: there are two modes of operation: forward-dynamics mode, where the simulated muscle activations are used to control the monkey arm, and inverse dynamics, where the joint positions and velocities are used to determine muscle activations, which the neuronal activations are then fit to. The first task, of course, is to do the forward dynamics mode. In order for this to happen, we need to formulate the forward model, including lagrangian mechanics (?) and muscle activations based on length and velocity. length is simple. velocity should not be so hard.

UUdUCk? <a href="http://rtkueqhoyszb.com/">rtkueqhoyszb</a>, [url=http://wueoavmkhkoz.com/]wueoavmkhkoz[/url], [link=http://hbrbdhpideoi.com/]hbrbdhpideoi[/link], http://quucrzldeyej.com/

• Richard Anderson, Caltech
• Donoghue Brown University.
  • chapter in book has many refs.
  • publications. "neural data from 100 electrodes at a sampling rate of only 10khz with a resolution of 10 bits requires a baud rate of 2gbits/sec. 1e2*1e4*1e1 = 10Mbits/sec
  • wired article May 2005 quote: Donoghue remains convinced that the only way to give people with immobile bodies full interaction with their environment is through embedded electrodes. "No other method gives you the power and clarity you need to transform this noisy signal into something that a patient can use," he says. "The people who question whether this will really work, I don't think they realize how much has already been done. We've got a 1,098-day monkey, who had a working BCI for almost three years. The question is, how long do you want to keep doing this in monkeys?" I cannot find any published data from this experiment,wtf?
• Neural signals inc - Phillip Kennedy, Emory University. not sure what he has been doing lately; the news on his page is old.
• Andrew Schwartz U. Pittsburgh has not published results from his self-feeding monkey experiment? n.e.wy. Dawn Taylor's 2002 science paper
• John Chapin - really needs to update his site. Linda & Ray have left the lab but this page still lists them as belonging. I think Lee was going to move over to acenet (my hosting provider)? Lee's site / CV
• us flashy site!