- ...class
- It is
also possible to conceive of a more complex learning task, where
each stream has a sequence of class labels. However, for the rest
of this paper, we will only consider single class labels.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...glove
- A Nintendo PowerGlove was used to collect
the data. It is highly sensitive to noise, suffers low temporal and
value resolution, is sensitive to environmental factors, only
captures information from one hand and has no sensor on the little
finger. Newer equipment has been procured which should
significantly improve the quality of the data, and consequently
classification.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...classification
- For example, in the large Auslan domain, 30 per cent
accuracy can be obtained using only the maxima and minima of the x,
y and z channels.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...rule
- Such a rule is not
suited for classification, it is only for description.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...Localmin
- Note the
distinction between a global maximum (over the whole signal) and a
local maximum (a peak relative to the points surrounding it).
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...learner
- Laplace correction [Domingos and Pazzani, 1997]
and equal frequency discretisation into 5 bins
[Dougherty et al., 1995]
for continuous values were used.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...attributed
- In this context, attribution
occurs by finding the event from the instance that most closely
matches the synthetic event. The distance between the synthetic
event and the closest match is used as the synthetic event
attribute.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...difference
- All significance statements are
based on a paired t-test at the 99 per cent confidence level.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...domain
- It is
difficult to compare the current work with HMMs in an objective
manner, since it is not clear what states and transitions are
appropriate for the domains used in this work. Furthermore, because
of the noise levels in the data and the small number of instances
per class in the Auslan domain, we found in preliminary experiments
that in many cases a simple 3-state (54 parameter) HMM did not
converge.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.