Can someone explain me AI/ML/DL in traditional algorithmic way without AI jargons?

What I currently understand is that they convert the training data to numbers based on a complex black boxed mathematical algorithm and then when a new data comes in, the same conversion is done and a decision is taken based on where the the new number fits in within the geometry/graph plot of the old numbers from training. The numbers are then updated. Is this what they call AI? Nearest number/decision search?

Kindly try to avoid critic, I am having a difficult time understanding the already trending AI stuff. People say that the algo exists from long back but only now we have the compute power.

📌

📍

📍

📍(the more there are pinning this, the higher the pressure on the one answering)

🛎️

You are not far from what it really is. ML is "training" a model (which basically amounts to parameter changing) in order to have a model that correctly estimates/predicts or whatever you want to call it, a result on a given data set. One would pick a model(only after the data has been prepared) and increment/decrement numeric values until the model approximates to a given result. Trying reaaaaally hard not to use jargon here, but to get a good idea of this, try understanding the statistical concept of regression without even considering ML or AI, all of the stuff that we are doing right now is based on mathematical concepts that we've been able to do since waaay before, when we say that we just now have the computational power we mean just that, that we know have the capabilities to try different things. Consider regression, understand the concepts behind it and what it is meant to do. Pay close attention to what the margin of error inside MSE is and THEN attempt to use a computer :D

@AleCx04 once you do it manually you will be able to say "aaaaaaah ok ok it makes sense to understand why we are using tools and frameworks and bla bla now" it is far easier to visualize and change training parameters on a Python/Octave/R/Matlab based system than it is by hand. That is all Machine Learning is, a label maker really, and an overly educated guess over something.

At a very high level, all AI is searching. You have some problem domain, and for some "question" (input) given, AI searches the domain and tries to find an "answer" (output). An example is classification: given some representation ("feature set", using jargon), what is the correct label for that representation?

ML then is just narrowing the search space using data. With enough data and an appropriate mathematical model we refine the search space so that answers are more easily and accurately obtained. As @AleCx04 pointed out in his excellent answer, think of linear regression -> trying to fit a model to data. If a good fit is found, the model can be applied to previously unseen inputs and it will (hopefully) return appropriate output.

Current state of ML/DL/AI in nutshell
https://en.m.wikipedia.org/wiki/...

Some historical background is that some lawyer without mathematical diploma got bored and decided to learn some math to fuck all mathematicians deep in the ass.

Neural networks operate similar to how you're describing... But there are many techniques, so you should pick one and dig in

This video should explain it:
https://youtu.be/aircAruvnKk

tl;dr: Machine "Learning" is applied statistics. Using brute-force statistical analysis techniques, the machine tries to either classify something (eg given a picture, "this picture is a cat") or predict something (given some inputs - this is the predicted output). To help it out, it is supplied with some examples ("training data") - on which it computes some values which it can then refer to when given new inputs that don't exist in the training data.

Because of this, AI is not "Intelligent" and of course a machine cannot "Learn". It's literally just some algorithms applying statistics to a dataset and saving the results. Marketing doesn't care about technicalities like that however.

I like to think about models as black boxes through which data is pumped. And different "architectures and designs" are good for different things.

Idk, but when I read research papers I feel like it's more about "hands on trying" new things and seeing what works, and not so much about theoretisizing what could work.

AI - does not exists yet. Means a general problem solver, not a specific one. Also, self aware, and able to change it own behaivior. Think Human thought process in a computer, plus emotions.

ML - basically, a complex way to take a bunch of past observations together with thier results, feed them to a machine, let it study them, and generate an algorithem that looks at new obesrvations data, and predicts the results with some degree of accuracy, and accepted error rate. Many methods exist for ML. *Not to be confuse with AI*

DL - It is a specifc ML technique, that seems simple, but actually really, really complex. It is a neural net with hidden layers varient. It is much hyped mostly becuase existing trained models can be "fine tuned" on other datasets, and that it can do it own "feature extraction" from data.

@magicMirror

So, ML is an algorithm to generate an algorithm (say Decision tree, that is the least of the jargon I understand) upon which a decision took?

Th decision tree is based on the training data, and the data is converted to numbers/weights (Entropy, Information gain) to generate the decision tree? I know that there are way more complex algorithms (and Bagging, Boosting to ensemble those) than DT, but basically this how it works right?

Can I compare (In layman sense) this to the traditional indexing of a Database column to make search and match data faster?

@scor

So deep down just numbers/complicated structure of numbers (say model) generated from a complicated math and then fitting/searching for a new number over the model again using some complicated math/algorithm?

I know that the offensive memes comparing them to if-else are not true, but still so far am I correct? is it just a complex algorithm involving heavy math?

PS I am not aware of analogue synthesizers.

[train-of-thought]

Also, I have an intuition that human intelligence works over the physics property entropy which allows for things to happen in random unlike as in computer where only pseudo random is possible. Can I conclude that computers can never reach human intelligence unless the underlying core structure is modified to something cyborgy? That allows for randomness?

@all Thanks for all the answers. From the answers, my basic understanding on the original post seems to be correct. :D

If there is any DBA reading this, on an any rough/layman level, can I compare this to indexing of a Database column to create a model of the data (say B tree) and then search/match new data on the model?

@Amrish a little bit, if you’re referring to what gradient descent is.

@Amrish
1. The derivative algorithem is called a "Classifier". It usualy does not stand on its own, and requires some sort of datafeeder preprocessor, and result post processor. Sometimes the classifier is a bunch of data, that is used as input to the generic classifier algorithem with the data. Look at opencv for an example.

2. No - you can't say that. Decision trees are a specific category of ML tecniques. j4 and randomForest are examples of ML techniques that are based on trees. Other techniques like SVM, Bayes nets, Gradients, Neural nets and many other exists. The most basic ML tree generator uses info gain to choose the next decision node logic. This example is used in most intro level ML lectures. This gets really complicated, real fast. Ensambles, boost technique, back tracking, co learning and many many more.

3. No. ML is about using past data to generate predictions for the future. You can also use "use knowledge about past data to generate new knowledge with new data".

The comparison of ML to indexing a database probably isn't that useful. If anything that would be like feature selection - saying that there is a single feature that perfectly discriminates the datapoints, so you only need to search that one feature to get a perfect classification. However, you couldn't input a new ID into a database and expect it to "generate" a new record that matched the input! It's only a pointer to data that you already have, and the "feature" (the index) tells us nothing about the data except that they were already associated.

I see what you are getting at but it doesn't really fit as a good mental model for these concepts.