WORKSHOP SUB-TOPIC: What volume and type of signoff data are needed for pattern recognition/machine learning

[Addaci]

It is proposed to collect ca. 50,000 markes, initials and signatures (collectively known as signoff data) for subsequent pattern recognition and machine learning work.

• Should we collect more data?
• Do we require different amounts of data for the three proposed categories markes, initials and signatures due to different visual characteristics of these categories?
• What benefits do we anticipate from greater volumes of data? For example, doubling or tripling the volumes?
• What are the collection and processing costs of acquiring and working with greater volumes of data?
• Does the quantity of data we require vary accoring to specific machine learning techniques we may use?

[Giovanni1085]

Hi Colin,

these are good questions but hard to tackle without a closer knowledge of the data. What I would say is that:

• more data is better, if it includes something different and learnable from the original dataset and not much noise;
• I would intuitively say that probably signatures are intrinsically more complex than initials (more letters, for example), thus I would go with more data there. No idea about markes;
• cannot say, probably diminishing anyway;
• depends how you plan to acquire it, usually the processing cost do not change, the acquisition costs grow slightly sub-linearly with the amount of data;
• yes, but again it is difficult to say, and anyway it might be less of a problem than redundant, noisy or variably represented data.

Hope it helps.