Predictions can't be wrong. The part of inference (if any other) involving statistics with probability distributions and likelihood is Intuition only - the whole is known and split into a distributed population (the probability model). Predictions are real but cast from an ordered triple "I" sequence of processes, where intuition is only the first. The yield is 100% certainty.
With the probabilistic approach, maybe spaces should be the turk's boxes when YOLO (https://pjreddie.com/media/files/papers/yolo_1.pdf) applies and punctuation should stay as noise. The CV problem of detecting whether there is a transparent screen between the single camera and the image cannot be solved with inferred output. It is solved using binary inferred unsupervised learning and matched output. In other words, when the confusion between intuition and prediction is removed so is the confusion between binary population and binary distribution, intuitive (unsupervised) learning of a large number of random samples in a binary stream should be able to predict (tell) where the word starts and ends or the word size https://reverseengineering.stackexchange.com/questions/18451/what-are-the-popular-machine-learning-unsupervised-approaches-to-learning-binary In the text example only spaces should be removed and not the punctuation. It is part of a language: https://www.constant-content.com/content-writing-service/2016/05/4-key-differences-between-american-and-british-punctuation/ On 19.07.2019 06:55, Matt Mahoney wrote:
On Thu, Jul 18, 2019, 9:40 PM Costi Dumitrescu <costi.dumitre...@gmx.com <mailto:costi.dumitre...@gmx.com>> wrote: Write input text - remove spaces in the input text - compress - send - decompress - AI - output text including spaces. In 2000 I found that you could find most of the word boundaries in text without spaces simply by finding the high entropy boundaries using n-gram statistics. https://cs.fit.edu/~mmahoney/dissertation/lex1.html <https://cs.fit.edu/%7Emmahoney/dissertation/lex1.html> So, yes you could do this and encode just the locations where the model makes errors. But I was more interested in testing language models that simulate language learning in children. In particular, babies can identify word boundaries in speech at 7-10 months, which is before they learn any words. Children also learn semantics before grammar, which is the reverse of rule based language models. I wanted to show that language is structured in a way that makes it possible to learn it completely unsupervised. Using a deep neural network, the layers are trained one at a time in the order that children learn. And now we now have neural language models that compress to one bit per character, within the uncertainty bounds of Shannon's 1950 estimate of written English according to human prediction tests. *Artificial General Intelligence List <https://agi.topicbox.com/latest>* / AGI / see discussions <https://agi.topicbox.com/groups/agi> + participants <https://agi.topicbox.com/groups/agi/members> + delivery options <https://agi.topicbox.com/groups/agi/subscription> Permalink <https://agi.topicbox.com/groups/agi/Tc1fd5fc7fae0a6a9-Mb9b990dee4c827dce3deba0e>
------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tc1fd5fc7fae0a6a9-M9a3583dd5ebe0420b7064b80 Delivery options: https://agi.topicbox.com/groups/agi/subscription
