Machine Intelligence - Science topic

Our paper is related to Localization in underwater Visible light communications using neural networks, I believe Localization using neural networks or deep learning is a hot topic in UVLC, There are many papers on the indoor environment, but more researchers are attracted to underwater communications as the optical domain is far better than RF inside the water for different types of water (Sea or Water or Turbid water). Introducing the blockage and dynamic motion is challenge in localization and transmission in UVLC.

Two points. The approach is pretty similar to what Boscovich proposed in the 1700s, so if dating this type of procedure you can go further back (Farebrother, R. W. (1999). Fitting linear relationships: A history of the calculus of observations

1750--1900. New York, NY: Springer.) Second point. A disadvantage is it will be slow for even medium sized n. Here is a quick coding of T-S compared with the lm function in R (which is slower for n=10 because of all the checks it does before estimating the model) and what I assume is similar to the main computation bit in it. At n=10 they are similiar-ish, but n=100 T-S is much slower (note different units).

> theilsen <- function(x,y){

+ n <- length(x) # assuming no missing

+ slopes <- {}

+ for (i in 1:(n-1))

+ for (j in (i+1):n)

+ slopes <- c(slopes, (y[i]-y[j])/(x[i]-x[j]))

+ beta1 <- median(slopes[is.finite(slopes)])

+ beta0 <- median(y - beta1*x)

+ return(list(beta0=beta0,beta1=beta1))

+ }

> lmb <- function(x,y) solve(t(x) %*% x) %*% t(x) %*% y

> library(microbenchmark)

> x <- rnorm(10); y <- rnorm(10)

> microbenchmark(theilsen(x,y),lm(x~y),lmb(x,y))

Unit: microseconds

expr min lq mean median uq max

theilsen(x, y) 249.101 271.3515 764.2949 303.4510 373.6505 42047.800

lm(x ~ y) 1222.101 1293.1510 1496.4859 1419.3010 1594.7010 5597.801

lmb(x, y) 100.001 103.3010 271.7730 120.2015 186.0010 7302.101

neval cld

100 ab

100 b

100 a

> x <- rnorm(100); y <- rnorm(100)

> microbenchmark(theilsen(x,y),lm(x~y),lmb(x,y))

Unit: microseconds

expr min lq mean median uq

theilsen(x, y) 60628.902 75446.151 91017.986 76951.5510 80715.9015

lm(x ~ y) 1187.001 1377.951 1619.025 1659.1510 1807.2015

lmb(x, y) 100.600 111.702 176.185 192.4505 215.2015

max neval cld

543952.501 100 b

2262.701 100 a

303.602 100 a

>

combining MARL and safety would be an interesting area

Okay Stay but I think in order for a pattern to exist a theme must recur, therefor it has a characteristic by which it abides, so that is a restriction or physics in a straight jacket as Feynmann calls it. I think it is much like improvising on a musical intsrument, you have to develop your ideas according to a rhythmic cycle. The proton-second is an abstract idea that can be applied to large amounts of particles over macroscopic time periods. Alas the timescale of a second is characteristic of the proton, in this paper it determines its radius.

Dear Kamal Berahmand ,

In machine learning projects we often run into curse of dimensionality problem where the number of records of data are not a substantial factor of the number of features. This often leads to a problems since it means training a lot of parameters using a scarce data set, which can easily lead to overfitting and poor generalization. High dimensionality also means very large training times. So, dimensionality reduction techniques are commonly used to address these issues. It is often true that despite residing in high dimensional space, feature space has a low dimensional structure.

Regards,

Shafagat

Spurious correlation, or spuriousness, occurs when two factors appear casually related to one another but are not. The appearance of a causal relationship is often due to similar movement on a chart that turns out to be coincidental or caused by a third "confounding" factor.

A more data-driven approach to diagnosing spurious correlation is to use statistical techniques to examine the residuals. If the residuals exhibit autocorrelation, this suggests that some key variable may be missing from the analysis.

Destroy Image Classification by Ensemble of Pre-trained models | by Vardan Agarwal | Towards Data Science

Dear @Kabbilawsh Peruvazhuthi,

In R, you can use several packages:

-climtrends package: VonNeumann test, Cumulative Deviations Test, and other homogeneity tests such as SNHT, Buishand, and Pettitt

-DescTools package: Dunnett Test

-multcompView package : TukeyHSD()

-Function bartlett.test(x, …)

like Mr . Joel Chacón, I think also that question is a dependent problem, there is no exact answer.

You are most welcome dear Pingbo Tang .

Wish you the best always.

Thanks for sharing.

Dear Shima,

This is a very good question, which conferences are valuable to submit and participate. In Machine Leaning field the top conferences can be :

Neural Information Processing Systems (NIPS)

CVPR : IEEE/CVF Conference on Computer Vision and Pattern Recognition

AAAI Conference on Artificial Intelligence

If you need more technical information about the rank and quality of the ML and AI conference please let me know by email. (neshat.mehdi@gmail.com)

Dear Milad Eyvazi Hesar

There are numerous benefits to utilizing the Transformer architecture over LSTM RNN. The two chief differences between the Transformer Architecture and the LSTM architecture are in the elimination of recurrence, thus decreasing complexity, and the enabling of parallelization, thus improving efficiency in computation.

Kind Regards

In order to analyze sparse data, the best way is to perform dimension reduction, principal component analysis or singular values decomposition for data shrinking or removing the unwanted features exists in the given data base.

You can check the following bearing datasets (vibration signals Vs time):

PS: both of the dataset are treated in this paper :

IMS bearing : https://ti.arc.nasa.gov/tech/dash/groups/pcoe/

prognostic-data-repository/

Wind turbine high speed shaft data: https://www.mathworks.com/help/predmaint/ug/ wind-turbine-high-speed-bearing-prognosis.html

Palak Nath

Cloud based tools like Figma for great quick collaborative real time teamwork. Perfect for UX UI designers, writers and developers and testing.

Yan-Ning Sun ,

While it is good to take a broad view to understand the area that you are working to gain understanding of the field, your objetive for research should be focused on a particular area. I will agree with Karim Ibrahim in breaking the two categories and choosing one for research purposes as you have done in your previous publications.

To continue your search, I would focus on what are the areas of your potential advisors and look for common ground between your interest and theirs.

Regards

Aparna Sathya Murthy Thank you and sorry for long reply. I will use two alphas to create anistrophic filter!

You said that Deriche filter is a high pass filter, but Hale in [2] confirm it's a low pass filter.

[2] https://inside.mines.edu/~dhale/papers/Hale06RecursiveGaussianFilters.pdf

I agree with Mohammad Deeb. This link is useful: https://stackoverflow.com/questions/52356068/validation-accuracy-constant-in-keras-cnn-for-multiclass-image-classification

There are actually quite a lot of nice application of machine learning techniques in the context of business process variant analysis, which is a fairly large subset of the process mining literature.

For example, Folino, Cuzzocrea et al. have done a series of studies on variant analysis (or deviance mining) using various machine learning methods, including ensemble learning and clustering:

We recently conducted a literature survey of methods in the field of variant analysis, many of them based on machine learning techniques:

Related to the above, there is work on bayesian networks for delay analysis (explanatory rather than predictive):

The above is related to variant analysis and performance mining. But there is also work on anomaly detection in event logs using bayesian networks:

And using deep learning architectures:

As well as using deep learning models to compute alignments in order to correct anomalies:

And a bit related to the above, there was quite a bit of research on using trace clustering in the context of automated process discovery (e.g. Jochen De Weerdt)

So we can say that process mining and machine learning go well together. One should not forget though that BPM and process mining are application-oriented disciplines - their objective is to design approaches to improve business processes. Whereas machine learning is a horizontal discipline, it seeks to develop methods that can be adapted to a broad range of problems/settings. Process mining has tapped a lot into machine learning, but sure it has a lot more to exploit from it.

Topics

Ashutosh Karna Thank you for your response.

my main objective is in oil and gase (industry 4.0) equipment maintenance use case and i need temperature,pressure ,humidity or volume flow sensors to predict supervised failures to train a model.

Artificial Intelligence AI is the scientific broader of intelligent machines (smart machines), while Machine Intelligence or Machine Learning is one of the main applications of AI.

You can see all the applications of AI covered by the International Journal of Distributed Artificial Intelligence (IJDAI) , a specialized AI journal

See this publication

Karayaneva, Y., & Hintea, D. (2018, February). Object recognition algorithms implemented on NAO robot for children's visual learning enhancement. In Proceedings of the 2018 2nd International Conference on Mechatronics Systems and Control Engineering (pp. 86-92). ACM.

Upadhyayaa, P., Farooqa, O., & Abidia, M. R. (2018). Block Energy Based Visual Features Using Histogram Of Oriented Gradient For Bimodal Hindi Speech Recognition. Procedia Computer Science, 132, 1385-1393.

Kaur, H., & Kumar, M. (2018). A comprehensive survey on word recognition for non-Indic and Indic scripts. Pattern Analysis and Applications, 1-33.

Benmoussa, M., & Mahmoudi, A. (2018, April). Machine learning for hand gesture recognition using bag-of-words. In Intelligent Systems and Computer Vision (ISCV), 2018 International Conference on (pp. 1-7). IEEE.

Junejo, I., Dexter, E., Laptev, I., & Perez, P. (2011). View-independent action recognition from temporal self-similarities. IEEE transactions on pattern analysis and machine intelligence.

Developing a true test as to the consciousness of Artificial Intelligence would be difficult, but the truest measure is that fully self programming AI's are virtually unknown. The self-modeling, self-generation capacities of humanity are of definite interest in generating an AI with greater adaptability in a broad set of circumstances. That, in my opinion, is the next threshold to cross in Artificial Intelligence. Another difficulty with AI is  how do you define an Artificial Intelligence without having a fully accurate model of human intelligence? Our adaptability is our best defined feature as humans, so adaptive artificial intelligence may be a better goal than strong artificial intelligence, at least until a clear picture emerges of more general features such as consciousness.

You can take the average error and select the model whose error is near about the average error. You can also use ensemble techniques. Please try with the link:

Thanks Surender! In Section 13 of Robinson's CL 2000 paper, he confirms that he was the introductor of the term in the 1970 paper I found.

You can used clustering algorithm like

1. K-means/medoid

2. Fuzzy c-means

3. partition-based clustering etc. for the classification issue.

Artificial Intelligence has been around for a long time – the Greek myths contain stories of mechanical men designed to mimic our own behavior. Very early European computers were conceived as “logical machines” and by reproducing capabilities such as basic arithmetic and memory, engineers saw their job, fundamentally, as attempting to create mechanical brains.

As technology, and, importantly, our understanding of how our minds work, has progressed, our concept of what constitutes AI has changed. Rather than increasingly complex calculations, work in the field of AI concentrated on mimicking human decision making processes and carrying out tasks in ever more human ways.

Artificial Intelligences – devices designed to act intelligently – are often classified into one of two fundamental groups – applied or general. Applied AI is far more common – systems designed to intelligently trade stocks and shares, or manoeuvre an autonomous vehicle would fall into this category.

Neural Networks – Artificial Intelligence And Machine [+]

Generalized AIs – systems or devices which can in theory handle any task – are less common, but this is where some of the most exciting advancements are happening today. It is also the area that has led to the development of Machine Learning. Often referred to as a subset of AI, it’s really more accurate to think of it as the current state-of-the-art.

The Rise of Machine Learning

Two important breakthroughs led to the emergence of Machine Learning as the vehicle which is driving AI development forward with the speed it currently has.

One of these was the realization – credited to Arthur Samuel in 1959 – that rather than teaching computers everything they need to know about the world and how to carry out tasks, it might be possible to teach them to learn for themselves.

The second, more recently, was the emergence of the internet, and the huge increase in the amount of digital information being generated, stored, and made available for analysis.

Once these innovations were in place, engineers realized that rather than teaching computers and machines how to do everything, it would be far more efficient to code them to think like human beings, and then plug them into the internet to give them access to all of the information in the world.

Neural Networks

The development of neural networks has been key to teaching computers to think and understand the world in the way we do, while retaining the innate advantages they hold over us such as speed, accuracy and lack of bias.

A Neural Network is a computer system designed to work by classifying information in the same way a human brain does. It can be taught to recognize, for example, images, and classify them according to elements they contain.

Essentially it works on a system of probability – based on data fed to it, it is able to make statements, decisions or predictions with a degree of certainty. The addition of a feedback loop enables “learning” – by sensing or being told whether its decisions are right or wrong, it modifies the approach it takes in the future.

Machine Learning applications can read text and work out whether the person who wrote it is making a complaint or offering congratulations. They can also listen to a piece of music, decide whether it is likely to make someone happy or sad, and find other pieces of music to match the mood. In some cases, they can even compose their own music expressing the same themes, or which they know is likely to be appreciated by the admirers of the original piece.

These are all possibilities offered by systems based around ML and neural networks. Thanks in no small part to science fiction, the idea has also emerged that we should be able to communicate and interact with electronic devices and digital information, as naturally as we would with another human being. To this end, another field of AI – Natural Language Processing (NLP) – has become a source of hugely exciting innovation in recent years, and one which is heavily reliant on ML.

NLP applications attempt to understand natural human communication, either written or spoken, and communicate in return with us using similar, natural language. ML is used here to help machines understand the vast nuances in human language, and to learn to respond in a way that a particular audience is likely to comprehend.

A Case Of Branding?

Artificial Intelligence – and in particular today ML certainly has a lot to offer. With its promise of automating mundane tasks as well as offering creative insight, industries in every sector from banking to healthcare and manufacturing are reaping the benefits. So, it’s important to bear in mind that AI and ML are something else … they are products which are being sold – consistently, and lucratively.

Machine Learning has certainly been seized as an opportunity by marketers. After AI has been around for so long, it’s possible that it started to be seen as something that’s in some way “old hat” even before its potential has ever truly been achieved. There have been a few false starts along the road to the “AI revolution”, and the term Machine Learning certainly gives marketers something new, shiny and, importantly, firmly grounded in the here-and-now, to offer.

The fact that we will eventually develop human-like AI has often been treated as something of an inevitability by technologists. Certainly, today we are closer than ever and we are moving towards that goal with increasing speed. Much of the exciting progress that we have seen in recent years is thanks to the fundamental changes in how we envisage AI working, which have been brought about by ML. I hope this piece has helped a few people understand the distinction between AI and ML. In my next piece on this subject I go deeper – literally – as I explain the theories behind another trending buzzword – Deep Learning.

Bernard Marr is a best-selling author & keynote speaker on business, technology and bigdata. His new book is Data Strategy. Toread hisfuture posts simply join his network here.

Dear Zhiliu

All classifications algorithm design for binary class classification problem then they extended to deal with multi-class classification problem. so, if your problem is binary class use binary other wise use multi class.

Regards,

 how can we initialise weights? Can you explain with small example.

I also agree with Sameer. Speed is not everything. You also need to focus how accurate it can detect or predict spyware stealing data from the system. Today various extended part of data mining(Deep Learning) is also used for this purpose. You may follow these below links for your help.....

Encoding robot ethics through sets of human-designed rules may not be computationally tractable, and as Asimov shows, it's incredibly difficult to create a small set of ethical rules to follow without encountering situations in which those rules do not play out as intended. 

You may be interested in this recent paper from our laboratory in which recent approaches to automatically learning ethical principles through Inverse Reinforcement Learning are critiqued: https://hrilab.tufts.edu/publications/aaai17-alignment.pdf

A quote from that paper seems directly relevant to your question: "The idea of training a system on data (either supervised or unsupervised) has captured more

and more attention as a way to understand how ethics and AI might best function in concert. Coding ethical values “by hand,” in the manner of many other traditional forms of coding – seems destined for the lesser task of lending basic “scaffolding” within which machine learning can operate (Tanz 2016)."

I also highly recommend the book "Robot Ethics" (https://www.amazon.com/Robot-Ethics-Implications-Intelligent-Autonomous/dp/026252600X) which presents several approaches to robot ethics, and discusses some of the ethical challenges faced by robots. 

Hi Bejoy,

If you are able to post an example image or dataset, we may be able to put together a recipe file for you in our new extended free trial software MIPAR (http://MIPAR.us) that should at the very least make you aware of the processing techniques involved in prostate segmentation if you need to track down open source codes yourself.

Cheers,

John

Surely they learn. In a technical sense

learning is nothing more than the changing

of a system's state in order to do

something better.

What 'better' means is a matter

of philosophy.

Regards,

Joachim

Also, these files may help you. Good luck

 Hi

I attached you a paper that explained algorithm of extraction  triple features by trace transform and also you can email the author to obtain the code

Regards

 Dear friends, Dorian, Dick, Roman, Mario, et al.,

            There is hope (Haykin and Fuster, Proc. IEEE, 102, 608-628, 2014).  But first we have to modify our computer and some of our traditional ideas about the brain. With respect to both, here I offer humbly some of my views after half a century of working in cognitive neuroscience.  I shall be brief and cautious.  For an account of empirical evidence, read my “Cortex and Mind”  (Oxford, 2003).

            1.  Alas, the computer cannot be only digital, but also must be analog.  Most all the cognitive operation in the brain are based on analog transactions at many levels (membrane potentials, spike frequencies, firing thresholds, metabolic gradients, dendritic potentials, neurotransmitter concentrations, synaptic weights, etc., etc.). Further, the computer must be able to compute and work with probabilities, because cognition is largely probabilistic in the Bayesian sense, which means that our computer must also have a degree of plasticity.

            2.  The computer must also have distributed memory.  In the brain, especially the cortex, cognitive information is contained in a complex system of distributed, interactive and overlapping neuronal networks formed by Hebbian rules by association between temporally coincident inputs (i.e., sensory stimuli or inputs from other activated networks).  The cognitive “code” is therefore essentially relational or relativistic, and is defined by connective structure, by associations of context and temporal coincidence.  That is why, theoretically, connectionism and the connectome make some sense.

            3.  It is true that the soma of a neuron contains “memory”: in the mitochondria.  But that is genetic memory (what I call “phyletic memory,” memory of the species), some of which was acquired in evolution.  It is important for brain development and for the function of primary sensory and motor systems.  It is also important for regeneration after injury.  Further, it is the ground-base on which individual cognitive memory will be formed. But the latter consists of more or less widely distributed cortical networks or “cognits” (J. Cog. Neurosci. 21, 2047–2072, 2009).  These overlap and interact to a large extent, whereby a neuron or group of neurons practically anywhere in the cortex can be part of many networks, thus many memories or items of knowledge.  This is trouble for the connectome which, if ever comes to fruition, will be vastly more complex than the genome.

            4.  Our present tools to define the structure, let alone the dynamics, of the connectome appear rather inadequate to deal with those facts and hypotheses.   Consider DTI (diffusor tensor imaging), one of those tools presently in fashion and widely used to trace neural connections.  It is based on the analysis of the orientation of water molecules in a magnetic field.  Therefore, it can successfully visualize nerve fibers with high water content, such as myelinated fibers and some large unmyelinated ones.  But the method (I dub it “water-based paint”) is good for tractography, for visualizing large, fast conducting fibers, but not for the fine connective stuff that defines memory networks.

            5.  What’s more, those networks change all the time, even during sleep.  In sum, it is difficult to imagine a dynamic connectome that would instantiate the vicissitudes and idiosyncrasies of our thinking, remembering, perceiving, calculating and predictive brain.

            Some of this may be wrong.  But that’s the way I see it, and may be useful to model the real brain.  Cheers, Joaquín

Hi 

The short a definite answer: yes, neural networks can take weather variables (temperature, wind chill, humidity) into consideration, and also their lags (not everything is contemporaneous). Indeed neural networks are particularly well suited to forecast load demand as the relationship between weather variables and load is non-linear and the modelling is much more adequate than with linear regression.

BTW: the question of "normalising the input" is confusing´the terminology a bit: in electriciy demand forecasting you often refer to normalise the demand by deseasoanalising (i.e. removing weather effects); in Neural Networks nromalisation refers to scaling the input variables into a suitable range that the algorithm can learn from (i.e. [0; 1] or [-1, 1]).

So yes, normalise the inputs in a neural network training  sense, but do not preprocess / deseasonaliose the data as this woudl get rid of the nonlinear interactions between weather and load. 

Hope this helps, Sven

Dear Alexey,

Classification can be based on a prototype representing a class but in many cases the data used for learning the classifier are represented as a set of objects with assigned class labels.

If you run a clustering algorithm, you divide the number of input objects into a number of disjoint (typically) sets. So a cluster number (e.g. originating from k-means algorithm) can be used as a data label used to learn classifier (e.g. C4.5).

I think the criminal profiling takes into account historical data, at present collected in data bases and mined with classical techniques. It would be very hard to assume an apriori knowledge and programs. The oldest sources are probably the Bible, Greek tragedies, Mahabharata, Hamlet, etc.

Best regards

Piotr

It is NP-hard, so you won't find anything significantly better.  Johnson's rule for the case of makespan without blocking can build up a huge buffer (see reference below) so it is not going to give you a good heuristic. 

Ramudhin, A., Bartholdi, J. J. III, Calvin, J. M., Vande Vate, J. H. and Weiss, G., "A Probabilistic Analysis of 2-Machine Flowshops", Operations Research, 44 No. 6, pp. 899-908, 1996.

IQ tests don't measure "Intelligence" (whatever that is), they measure IQ (whatever that is).

The interisting thing about IQ tests in the light of Turing is that they have always tried to factor out linguistic skills to prevent cultural background and education from confounding the testing of inherent "intelligence."  Therefore these test contain many questions exercising non-verbal patern matching / prediction. So that is where I would start if I were you and it was specifically "IQ" as opposed to "intelligence" that I wished to measure.

Hi, your question is very general so an answer can only be general also.

I would add, that the needed degree of autonomy is very depending on the application you're targeting on. High autonomy would be needed for autonomous transport systems in public traffic scenario as the environment is very demanding, whereas in industrial application autonomy level could be lower as special measures or infrastructure components can be used. Second, the economic site: Depending on application you're not able to get to "full autonomy", which is expensive (sensor, data processing). Also legal reasons have to be considered.

Deriving from biological systems, I would say that intelligence isn't needed for autonomy, but it makes an autonomous system behaves in a "wiseful" way.

I disagree here. The question is not how to train a classifier in general and which are the well known courses/softwares to (learn to) do so.

The question,  as far as I understood,  is how to train a classifier when the available supervision at training time are soft class labels in the form of a class conditional probability value for each training example. In a binary classification case, this would be p(x|c=1) for each sample x in your training set conditioned to the class "c" (or its complement p(x|c=0) for the other class).

The fact that a standard logistic regression outputs a model that produces such probability values, once the model is trained and can be used on independent test examples, does not answer how to use such a soft supervision at training time.

In other words, a standard logreg package would typically require as inputs a  training  set in the form of concrete examples "x" together with their hard class labels e.g. (c=1) or (c=0).

Besides, the most standard way of fitting a logreg is by  iteratively-reweighted least squares (IRLS) and it does not look immediate that this specific type of optimization actually "minimize H(p, q), averaged over your training set."  At least, it would deserve some math to make the (doubtful) link or specify under which type of specific distributions p and q, both optimizations would happen to be equivalent.

thank you for your response. i am also basically searching on use of ANNs in HVAC. but unable to find any appropriate model using ANN for this purpose. can you please share some stuff (literature or project names) on use of ANN in HVAC systems.

Thanks

Rajeev.

TPR and FPR separately do not tell you anything. The same goes for Precision and Recall. F-score, Equal Error Rate and similar can give you some information, but it is limited and you have to be careful how you use those values.

Look at this previous question:

I'll copy part of my answer for that question here:

You have to first answer a simple question: Do You know what are the costs of the decisions?

If the costs are known, you can calibrate your classifiers (set the operating point), and compare costs of the classifiers.

If you do not know the costs, you may have some insight about a reasonable range of operating points of the target application. Lets say the target application is a surveillance system, and all positive responses will have to be shown to a human operator which is able to process one event per minute. In such case, it makes sense to compare detection rates at one false positive per minute. You can use other measures instead, but it makes sense to keep one type of error fixed and compare the other one.

If you have no knowledge about the possible operating point in the target application, you should show results in full range of possible operating points. Plot ROC, precision-recall, DET or something similar. You can also compute area under ROC or Precision-recall and get a numeric value which reflects performance over the whole range of regimes; however, such aggregation has its own problems.

From ECG or by finger plethysmography one obtains R-R intervals that is the tachogram. on such obtained time series one may perform FFT obtaining three bands , VLF,LF,HF to aobtain ANS analysis

There are several aspects to this. In the case of linear SVMs, during training you must estimate the vector w and bias b and this is usually done by solving a quadratic problem. Solving the quadratic problem is not easy (at least in the general case). For instance, testing that you have an optimal solution to the SVM problem involves something in the order of n² dot products, while solving the quadratic problem directly involves inverting the kernel matrix, which has complexity in the order of n³ (where n is the size of your training set). That being said, the time required for linear SVMs to reach a certain level of generalization error actually decreases as training set size increases, e.g.

The prediction time is linear in the number of features and constant in the size of the training data. For some more details have a look also here:

A generic answer as to what is degree of freedom is the minimum number of independent variables required to completely describe a body.

When it comes to mechanism, generally, we are always referring to a rigid body, unless you mean a compliant mechanism.

For a rigid body, it would mean the number of independent co ordinates needed to define the position of the body.

In a planar mechanism, a body (typically a link) can translate along x axis, or y axis or it can rotate about the z axis, in xy plane. Hence to completely define the position of the link, we need to specify the x co ordinate, the y co ordinate and the magnitude by which the link is inclined to either x or y axis. Hence we have three independent variables, hence 3 dof.

In a spatial mechanism, we would have 6 dof, translation along the three axes and the rotation about them.

So much about the degrees of freedom of an independent body.

Now, when it comes to a mechanism, there are more than one bodies, and they are no longer independent, since they are joined to each other in a particular fashion. This is called a constrained body, since it will no longer have as many degrees of freedom as an independent body, and is hence constrained.

The constrained are physically applied by connecting two bodies to each other through what is called a joint. Thus a joint allows certain degrees of freedom between the links, and constraints some of them. This results in formation of what is called as a kinematic pair.

The kinematic pair may allow only one degree of freedom, called a lower pair, or may allow more degrees of freedom, called a higher pair.

example of lower pairs are pin joint, which allow only planar rotation, or slider joint which allows only translation in one direction.

The degree of freedom of a mechanism is the total number of independent variables required to define completely the particular mechanism. This can be found very easily using the Grublers equation...

Gokul, I think of the 'meme' [http://en.wikipedia.org/wiki/Meme] as a container.

If the idea is worthy (weighted value) then one meme holder (machine or man) may pass the meme (high in concept, low in data) to another, for consideration. When/If a positive value-exchange occurs between two or more parties which have judged the meme container content exchange to be of interest, the meme can be populated with data by all parties and its 'worthy' value is increased.

This semantic exchange model borrows heavily from the well-proven TCP/IP header implementation: (interest) "Is it for me?", (construct) size, count of pieces, point-of-origin, etc., (content) packet.

The web services metaphor is also applicable.

Deep Learning is a way of learning everything you need for a classification or clustering problem with neural facts and models. But it requires a good amount of computational power to get away its ramification complexity. If you provide this huge clusters and GPU facilities to your algorithm it is reasonable to obtain promising results for the problem. It is what the Google does currently. They are devoting clusters of computers to learn visual concepts from their image search engine with Deep Learning algorithms. On the other side of the spectrum it is not possible to find this much power as a simple company or institution. Therefore, other methods like hand crafted features and kernel machines also need to be considered even with simple Bag of Words features. There are also some papers saying that well crafted BoW approaches possibly give overrated results in relation to deep learning methods in some of the problems in especially computer vision. Beside all those criticism it is certainly most ground breaking learning tool of the last age of ML with its neuroscientific promise as well.

Here is one example paper :http://www.stanford.edu/~acoates/papers/coatesng_nntot2012.pdf

Human and machines are quite different Human is self-regulated in all aspects of metabolism, surprising ourselves with each aspects of life, including ADN and ARN mechanisms of interconnection. But do not forget that humans spirit and mind, can reaction with other human's with empathy, capability to project more and more complex ways of life,including human communities. History of mankind and the capability to remember and understand multiple actions and effects on human ideas, is one of the main differences between human and machine. even you think in intelligent machines.

As Marco pointed out, SVM can be used in a wrapper approach for feature selection. And I fully agree that, while this is fine for small-scale problems, it becomes quickly intractable for bigger tasks. A well-known embedded feature selection method using svm is to use the L1-norm (or approximate L0-"norm") weight regularizer in linear svms, which leads to a sparse weight vector whose non-zero components are associated to the features of interest.

The trade-off parameter can be set via cross validation. This can be very fast for even large-scale problems using fast linear solvers like liblinear.