What can be the examples of Rote Learning in Artificial Intelligence to understand it more easily?
Any references are appreciated.
Artificial Intelligence Stack Exchange is a question and answer site for people interested in conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment. It only takes a minute to sign up.Sign up to join this community
There's no 'Rote Learning', explicitly. Machine Learning is essentially rule based, rules are learnt over examples. But there is indeed a reference to 'Rote Learning', which is Overfitting. A good example of rote learning 'can' be said as a one class classifier, like an autoencoder.
Overfitting occurs when you try to memorize so much, that it just gets stored as-is. You (the model) stops learning features, and rather starts learning the example itself.
Steps to reproduce:
Take any ML model and a training dataset. Train it indefinitely. There's a curve associated with training and validation error. Beyond a certain point, the model starts overfitting. That essentially means that its training accuracy improves over time, but stops generalizing over new data, like the ones in test set.
Here is an example of overfitted/perfect/underfitted model:
Here is a typical loss curve. Beyond the centre dip, it is unworthy to train.
Rote learning is the repetition of linguistic expression without knowing the meaning in response to a query or other trigger that indexes what was heard or experienced in text form.
The earliest form of rote learning in machines is printing publication. This qualifies as rote learning when available in conjunction with the manuscript, a publisher, the Dewy Decimal system, and the library system with the now obsolete card catalog. However crude the indexing of the material seen through contemporary eyes, there is at least a primitive form of trigger that causes appropriate captured and saved text to be recalled. The manuscript, press, book, indexing, and librarian together perform the entire learning and recall sequence. Phonograph and film additions to libraries followed, and, if indexed properly, also qualify roughly as mechanical versions of rote learning.
All of these went digital during the twentieth century, and the web, with the ability to acquire a domain and hosting, upload files, and let web spiders index them is a form of rote learning. At this point, the web demonstrates rote learning capabilities far beyond that of humans. The algorithms surrounding the indexing and recall are proprietary, but some information about the early algorithms are available online, and these in combination with the other parts of the process listed above are certainly examples of rote learning and the recall of what was learned.
That these are not rote learning because they involve multiple components is an invalid critique. It is true that all the above examples have several components that must be leveraged in conjunction to perform both rote learning system and the later use of what was learned, but so does the human brain performing rote learning in conjunction with other system components. Vocal forms of rote learning and recall would be impossible without air. Textual forms require both light and a pen. The complexity of the senses and motor coordination involved in writing and speaking should not be dismissed either.
In this context, memorization of natural language and storage of natural language are quite similar in external function. The trigger of recall can be much the same too, so these machine capabilities are, without much we would call AI, reasonable demonstrations of rote learning. Advanced and not altogether available AI is required for the type of learning that occurs when a student studies or, even more so, a researcher performs research, but those types of learning are not rote learning at all.
The current technology is beginning to be able to apply rote learning across expression types, storing a digital representation of vocal expression and converting it to text or musical notation with text. The Dragon line of voice to text products was an early example of that. In the reverse direction, text can be stored and recalled based text and converting it to fairly real sounding vocal expression. The WaveNet-Tocotron2 approach is a good present day example of that. Once the text is indexed, a response stored can be indexed based on some textual or vocal trigger, which qualifies that as rote learning.
What is missing, is the ability to read a question in a test and pull back the answers from the text book, the text of which had been stored, or repeat what the teacher or professor said from digital audio recording in response to a voiced question, but only because there is little application for these two capabilities.