... scenario

Any correspondence to real software companies is purely coincidental.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... language

Note that recognising sign language as a whole is extremely difficult - to mention three serious difficulties: the use of spatial pronouns (a certain area of space is set to represent an entity); the use of classifiers - signs that are physically descriptive but are not rigidly defined; and finally improvised signs.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Auslan

Auslan is the name used for AUstralian Sign LANguage, and it is the language used by the Deaf in Australia.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... independent

For example, in the speech recognition community, this is known as ``coarticulation'', where the current word being pronounced depends on the prior and subsequent word. Similar problems exist in other domains.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... models

To avoid confusion, we use the term ``model'' here to represent a single hidden Markov model - that is, a single state transition diagram plus the associated probabilities - and will use the term ``HMM'' to represent the concept as a whole.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... mixture

Note that the Gaussian mixture is typically implemented under the assumption that the channels are independent, and hence if there are $c$ channels, then there are $2c$ parameters - a mean and a variance for each channel. Strictly speaking, however, $c^2 + c$ parameters are required: $c^2$ for the covariance matrix (needed if the channels are dependent) and $c$ for the means.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... evaluated

10 transition variables, and assuming independence of channels, a mean and standard variation for each of the 5 variables in the 4 different states; if we were using covariance matrices, there would be 130 parameters

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... transitions

This is not what happens in real speech recognition systems. Usually, speech recognition systems each HMM is trained on a phoneme; and these phonemes are assembled to form the starting HMMs for words. However, we do not always have this sort of information available to us.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... difficult

In many ways, dynamic Bayes nets suffer the same problems as HMMs. In fact, it can be proved that dynamic Bayes nets and hidden Markov models are isomorphic. Dynamic Bayes nets have the advantage that they do allow for far more complex state models and provide a much more intuitive mechanism than HMMs for inclusion of background knowledge.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... extent

This paper makes significant use of the work presented here, although the techniques are different in many ways.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... side

Note the extraction function given is a little simplified and does not take into account the possibility of the sequence beginning or ending in high volume. However, in practice, we would implement it to handle such cases.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... learner

The line between preprocessing and learning is a blurred one. One view of TClass is as little more than an automated preprocessor, which doesn't do any learning itself. However, preprocessing the data into a format for learning seems to be something that machines are bad at and people are good at, so perhaps it is not so little after all.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... simultaneously

There is nothing in the theory or the practice that prevents a metafeature from looking at multiple channels simultaneously. Recall that the extraction function takes an instance, and not a channel from an instance; hence it would be possible to look at multiple channels to extract events. This would be logical, for example, in the case of the three spatial features x,y,z, where some features e.g. circular motion in the xz plane, as occurs with the sign ``all'', when it is necessary to look at both planes simultaneously to determine whether in fact a circular motion was performed.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...centroids

The use of this name will become obvious in a moment.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Happy

Note that there are two points at the location (3,3), hence it actually counts as two instances. Also note that in the case of an instance lying exactly on the border, we randomly select which region it belongs to.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... membership

This is very similar to the approach taken in fuzzy logic. The fuzzy ``or'' of two fuzzy values is usually the maximum of the two.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... idea

This is one of those ideas that is kind of obvious once explained, but is not so obvious to one's own intuitions. The penny dropped for me during a lecture by Thierry van der Merckt at the University of Sydney.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... (FIR

There is another type of filter, known as the Infinite Impulse Response or IIR filter. For brevity, we will limit our discussion to FIR filters.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... applied

Within the boosting framework, each classifier may have a different weight; not all classifiers vote with a value of 1.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... readability

There has been some work on this however, also discussed in [BK99]. More recently, Freund [FM99] has introduced the concept of alternating decision trees, which incorporate boosting as part of the classifier.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...fig:tech-supp.tal

NOTE: Because the terminology has changed and to avoid confusion, we have modified the file slightly and replaced all the old terminology with the new terms. The actual file is identical structurally but uses different terms.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... value

Of course, less than or equal to and greater than or equal to are also acceptable.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... attributes

If it is a global attribute, of either the aggregate or conventional type, we leave it alone.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... learnt

The former paper is extremely interesting - it shows that in many cases, experts don't even agree with their own diagnoses after sufficient time has passed, and they disagree with their own rules. Furthermore, agreement between the expert system developed by one expert and another expert was extremely low - less than 40 per cent.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...)

This is a fairly realistic assumption. All of the implemented metafeature extraction functions are linear.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... learner

We have not found references to this as a general technique. However it seemed so obvious that it was not considered worthy of publication. Other researchers such as [Geu01] have since cited us for the algorithm.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... available

These datasets are: arrythmia, audiology, bach chorales, echocardiogram, isolet, mobile robots, waveform.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... learners

There is no result for 2 bagged learners, since the result is arbitrary: if one learner votes one way and the other votes another then it's an arbitrary choice as to which is correct, and because of the binomial theorem, in the limit you can't do any better than a single voter.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... fingers

The little finger is actually important in many sign languages. For example, in Auslan and British Sign Language, a fist with the little finger extended indicates bad. It is also used as a modifier for other signs that indicate badness - e.g. sick is a bad handshape against the body and swear is a bad handshape starting at the mouth and moving away.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... sources

In practice, it proved to be easier for synchronisation purposes to link the two Flock-of-Birds units using the Flock-of-Birds Bus (FBB) and interleave their data into a single serial port.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... degree

In fact, the main source of orientation error (and to some extent position error) was the difficulty in attaching the magnetic trackers to the gloves. It is very difficult to attach these sensors to the hands in such a way that it simultaneously provides relatively free movement and also connects firmly to some reference point on the hand. Some tradeoff is involved.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... signer)

My deepest gratitude extends to Todd Wright for his help on this project.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... times

All timing results measure the user time (i.e. those directly used for calculations) on a Pentium-III 450MHz with 512MB RAM. One has to take these measurements with a grain of salt, since the code is not particularly optimised and is written in Java).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...thank

There are two events tested against here; this is an artifact of the random segmentation process

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...schiller:ecg

Unfortunately, companies are unwilling to disclose the exact performance of their systems for direct comparison.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...downsampling

There is also the case of upsampling, but it is is of little use to us in this context.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... size

Note that the downsampling factor, strictly speaking, need not be an integer. If it is not integral, a weighted average must be taken of the points that are shared. For instance, with a downsampling factor of 1.5, each point in the output would average one whole value plus half of the next or prior value.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...apostolico:caow

Subword trees are known by many different names. These include B-trees, position trees, prefix trees, suffix trees, subword finders etc

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... it

Note that Figure 7.4 displays a lot of redundancy. For example, the whole subtree beginning with r is actually a subtree of the one beginning with ab.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.