This is a shared problem with other learning tools. However, the forms of noise are somewhat different in TC than normal ML. Noise can take the form of temporal noise (measurements being taken slightly too soon or too late - this is unique to TC), channel noise (changes in channel values that do not represent the underlying data) and classification noise (incorrect classification of a training instance). For example, in the sign language data, the gloves themselves may use sensors that are subject to Gaussian or other noise; instead of sampling at 20 millisecond intervals, it may be that sometimes the time between this and the previous sample was 18 or 23 milliseconds; and the training data may have instances where surprise is mislabelled as danger.