Sensor data in Deep Learning: questions on a simple level

[nomadreid]

First of all: I am not a programmer, I cannot read code, only have a background in pure mathematics, am looking for a very simplified summary ("for Dummies") of the way Deep Learning and similar AI routines handle sensor data. I looked at a few articles such as http://papers.www2017.com.au.s3-website-ap-southeast-2.amazonaws.com/proceedings/p351.pdf, but these are written for those who already understand the basic idea and wish to go into details. So, the question is such: I imagine (in my naïve approach) the computer-without-sensors to be composed of a (first-order) logic, and then the sensor data might be attached in one of two different ways (please hold your laughter for the moment):
(a) the coded sensor data provides constant values for variables in the logic, so that the whole thing remains purely syntactical, with truth predicates either implicit or explicit, or
(b) the sensor data is the universe (in the sense of a model, i.e., an interpretation, that is, semantics) in the sense of Model Theory), and correspondence (the interpretation function in the sense of Model Theory) of the values with the syntactical symbols, guided by a given lattice of truth values (evaluations), all this then changing à la Kripke except much faster.
I emphasize that I understand that the true situation is likely not to be able to be fit into either one of these characterizations, but I present it as a start from which a knowledgeable person may start a rough explanation (without quoting lines of code).
Thanks.

 

[jedishrfu]

This is really very vague. I read universe, sensors, deep learning and give me a summary and all I can think of is this video sequence by 3blue1brown on YouTube:

 

[nomadreid]

Thank you, jedishrfu. On your recommendation, I have watched the four videos in this sequence. It addresses my question (obliquely, but in the right direction) , which, I recognize was exceedingly vague (due to my zero-background in AI), and this gives me a basis upon which to better and more precisely formulate my further questions in my attempt to understand more of AI (including deep learning) from a purely Model Theory viewpoint. So your nudge was very helpful.

 

[jedishrfu]

There is actually a Machine Learning for Dummies book that covers things in greater detail. Its a few years old and the industry is moving but looks pretty good.

Another excellent book is Hands On Machine Learning by Aurelian Geron. It covers the basics and Google Tensorflow v1 and could actually make you into an ML novice.

In general they teach the algorithm theory and how to use the algorithms but don't tell you how they work internally.

 

[nomadreid]

Thanks, jedishrfu. Upon your recommendation, I have (legally) downloaded the Dummies book, and will keep the other recommendations in mind. I will be looking more for the motivations behind the techniques than for the techniques themselves.

 

[jedishrfu]

ML is more of an art than science. Practioners select the appropriate algorithm for the task at hand. Sometimes neural nets make sense and sometimes genetic algorithms or k nearest neughbor.

What you discover though is there is nothing magic as they just divide up a search space and eventually find a reasonable though not necessarily the best answer.

 

[WWGD]

The way I understand it, you iterate your algorithm, be it regression, etc. using a threshold value , loss function, until your algorithm has succeeded meaning the loss is below the threshold value and it then stops. E ample, you conduct your regression , compute your loss. If the loss is below the threshold, you're done. If not, you tweak your parameters and iterate until your loss is below the threshold value.

 

[jedishrfu]

That’s somewhat simplistic strategy, rather you might tighten your threshold if a lot examples pass the criteria. Of course, you risk centering on a saddle point but such is life.

ML is still a flexible art where you start with schemes that work and then adapt them to your problem and try not to get lost in the data analysis woods. There is a drive to make ML more scientific but I haven’t seen much success in explainable AI or in other methods.

 

[WWGD]

Well,yes, it is too hard to give a more precise explanation without rigor and getting lost in the trees. Each aspect has its own complexity. You may also tweak or even change the algorithm, the loss function , etc. So I sacrificed rigor in order to shed some light, and hopefully it did accomplish that goal.

 

[nomadreid]

The idea of tightening the threshold, etc. is explained pretty well for text recognition in that series of videos which appeared in Post #2. For more complex tasks, such as playing chess or Go, I presume the idea of taking a program that more or less works and tweaking it according to its successes and failures follows the same lines.