What AI book should I read to be current in AI?

I keep getting advice to read a book on AI however no one who says this actually mentions which book I should actually read...

So what AI book should I read and why is this book better than any others on the topic?

In addition in a fast moving field like AI would reading current academic papers or journals not be a better way to keep up with what is happening in this field? Which academic journals are best?

 

There is no current in AI and books are always out of date. You will never be able to keep up because you need teams and a few million to even stay near the front. All academic papers are old by default, in all fields, useful for research but not for "being up to date".

So instead you might want to specialise in a specific part of AI (it is after all, an unlimited subject).

 

I agree. I'd learn python and tensorflow. And try to make simple learning models.
Also a popular image recognition is called YoLo.
Also here are some good youtube channels to keep up to date with the latest:

https://www.youtube.com/user/keeroyz

https://www.youtube.com/channel/UCWN3xxRkmTPmbKwht9FuE5A (can be a bit annoying !)

https://www.youtube.com/playlist?list=PLRqwX-V7Uu6aCibgK1PTWWu9by6XFdCfh (Introduction to neural networks in python)

That's about all I know!

 

Start with a book like this:
https://www.amazon.com/gp/product/1493682229/

It talks about fundamental concepts that humans need to use when they want to feed data into AI systems. The same author has a couple additional books to read after that one, but definitely start there.

Don't hop directly to neural networks in a specific language, because books like that will often gloss over the very fundamental concepts about how humans need to work with data and AI.

 

That is not exactly true. There has been a lot of new applications of existing AI concepts in recent years and there have been huge computing power gains in recent years, but the AI concepts themselves have not changed significantly in recent years. Artificial Neural Networks is a decades old concept. Reading a few solid books on the fundamental concepts is really important for being able to absorb all of the latest applications of those fundamental concepts.

 

It's perfectly true though. Already, the ball moved. The question I decided to literally answer was: "would reading current academic papers or journals not be a better way to keep up with what is happening in this field"

And it's not. Because it's impossible to keep up, and the data already changed since the time between that question and your disagreement.

Now if he specified a specific branch of AI, I'd give a different reply.

...have you even seen how much stuff is being published? You... can't keep up even if you started reading now and never stopped until you died. Go have a proper look.

Also: specialise and do it early.

 

I have this bookmark of an online course in AI that seems promising, but I can't say if it's good or not as I haven't had the time. It's definitely something I want to check out though:
https://brilliant.org/courses/artificial-neural-networks/

There's also this:
https://eu.udacity.com/courses/school-of-ai

That is for learning the basics I suppose.

As for just keeping up to date with what's happening in the field of AI, I would recommend this channel on youtube:
https://www.youtube.com/user/keeroyz/videos

Here, the author goes through new papers and summarizes them together with visual examples, all usually within a few couple of minutes. It's really interesting and mind blowing what kinds of AIs people create.

 

I agree that I cannot read everything that gets published due to the sheer number of papers and the time needed to read all of them. But (and this is extremely important), I can understand all of the papers that I do read because I have taken the time to learn the fundamental concepts. I strongly encourage everybody that is interested in AI to do the same. Learn the fundamental concepts, and then build on top of those concepts.

 

Sometimes the best ways are the old ways.
https://en.wikipedia.org/wiki/Artificial_intelligence

Just for starters.

 

Reading a book or even needing to be "current" on a topic like AI isn't the really the point. Specific applications or implementations don't replace basic knowledge or understanding. To grab a couple of illustrative comments from just a single page:

These comments are more about a failure in understanding how basic technology/computers/logic works. It's not specific to AI... it's not understanding the core/foundational tech at play. You seem to want to discuss things about some random article you just on AI at high level without understanding at the most basic level how computers work. Game AI != generalized AI. Generalized AI != other generalized AI. It's not a single thing. You aren't going find an answer in a book about AI. Maybe in some fundamental CS theory books.

If you want to learn something, ask. Your methodology seems to be throw out crazy, nonsensical things and wait for everyone to explain why they are nonsense. Ask questions to learn, don't make statements to be challenged. That is why you get negative responses and people telling you to go read a book. No one is going to take the time to explain how computers work.

 

I don't think that "keeping up" is the underlying concern that lead to these questions, though. As @zombiegorilla points out, people keep recommending that @Arowx "read a book" or similar because of questions or statements that imply a misunderstanding of the fundamentals, as opposed to a failure to understand the latest advances in a particular field. Zombie's post nailed that, so I'll move on.

The fundamentals generally change slowly, so books going out of date is much less likely to be an issue for introductory stuff. The stuff I learned at university ~10 years ago is still relevant today, and I don't think it was new at the time. It's clearly not the latest stuff, but it is foundational knowledge that gives me starting points to both create my own solutions for simple stuff, and have a starting point to understand when people talk about more advanced stuff.

I agree that academic papers aren't likely to be useful, but that's because they generally assume a prior understanding of the fundamentals. You're right that you could never keep up with all of them, but you don't need to do that even to be an expert in a field (otherwise we wouldn't have experts!).

 

I tried once, vaguely. But I have no life. :sadsmiley:

 

I would also recommend siraj, who provide step by step tutorial. But yeah the state of the art is going wild and fast, but the basic barely change atm.

https://www.youtube.com/channel/UCWN3xxRkmTPmbKwht9FuE5A/playlists

Having deal with all of this myself, I would say the fundamental is really simple once you get it down to what it really is, a bunch of matrix or if matrix scares you a bunch of excel table. The math involve is mostly the same as the supermarket receipt you get, multiplication of item and addition of the result.

A neuron is basically a column, so evaluation is summing column, and learning is summing row. A layer is a group of column to form a matrix. The main stuff are really architecture design and data preparation, and by architecture it's basically how you will feed one matrix result inside another matrix column, with matrix being chained toward the final result column, sometimes matrix feed matrix a bit further down the chain, or even previous matrix if you want temporal reasoning (though now you would have to "unroll" the loop for learning sake).

The most scary looking part are the "activation function" but they are basically just curve, and can be seen as blackbox. You just need two formula per activation function, the proper activation curve that is used on the column summing to get a result, and then his derivative when summing row for learning, with the caveat that some curve are hack (relu) but in the end you just copy what's given to you.

And the last scary stuff is the error function, if you played the magic number (I have number try to guess it, I will tell you if it's too high or too low) well that's what it does. Another way to see it is like a heuristic for pathfinding, so it has the same problem with over shooting, how big a step you can make in the learning space, etc ... it's where all the juicy stuff tend to happen with how to chain the matrix together. Error can be set by setting example (pair of result vs data, the typical supervised learning) or by using an evaluator (like a q function that see if the result if good or bad), that evaluator could be another Network (gan architecture). So most stuff is typical ai concern like how to define policy, sub goal, encoding of data, decoding of the network, etc ...

And that's basically all you need to know to start, I put that here just in case, I found that most metaphor of neuron as node network, rather than matrix, tend to obfuscate learning about them rather than inform it.

 

In terms of AI, we are utilizing methods, which were already discuses in mid 50s of 20 century. Then when computing powers start accelerating from 80s onward, AI options become more viable. Only past 15 years, this really accelerated further. Saying all that, we using often principles of algorithms, which some of them are near 100 years old. But back then, wasn't possible to prove the concept with available technology, as of lack of computing power.

Hence AI basics are barely changed since many decades. Fundamentals are pretty same. However specialization is something, which tremendously vary. Hence reading latest academics papers, will lead you only into one of many corners of AI topics. You may be able to grasp some ideas. But you are likely miss the basics principles, as these tends to not cover it. Papers expect, you know something about subject already. In this days, this something, is already a lot. Hence, need look much further back, to get the idea.

Some areas, but not exclusively, vision and objects recognition, sound generation, voice recognition, analytical data processing, word processing (for example for chat bots), genetics combinatoric, animal and human behaviors, social behaviors, forensic, human - robot interaction, control systems (for example drone stabilization), path planning, and of course gaming, to name just few.

 

I love AI. I'm in an AI class this semester and loving it!

That said the book we're reading is technically dense but a fascinating read (if you can get past all the mathematical lingo and terminology). I believe it was published in 2013. Its called Artificial Intelligence a Modern Approach by Peter Norvig and Stuart Russell.

It covers many algorithms and models for approaching AI type problems.


AI is a big field, and Machine Learning, Deep Learning, etc. are not one in the same. Even the simplistic algorithms like Finite State Machine's are for a different problem. It does cover A*, Constraint Satisfaction Problems and various pathfinding algorithms. Reading the book might be challenge for those without a good higher level math background, but implementation is almost a whole different beast. I'm a damn good coder imo but the last two programming assignments we've done I carried the project and still didn't manage to complete them.

Note: ML, DL, etc. are not particularly suited to game dev, in spite of what people might believe. There are ways to apply them to produce intelligent bots however they must be handicapped or they will ruin the player experience. Look at Alpha GO by Google. Imagine that as your bossfight. Not fun.

For reference, this is my copy:

Lastly I want to mention that all this consternation about AI bias against minorities is absolute bullcrap and you ought to ignore and think for yourself.

 

Also
game dev != game ai
Ai can have application in tools, production, etc ...

 

A solid understanding of 'classic'/symbolic AI (tree search etc.) is going to be very helpful even if you're mostly dealing with probabilistic AI, machine learning and/or neural networks. For a lot of "game AI" applications, symbolic AI is (still) a great choice, especially because most of these algorithms are very efficient - A* is probably still the most widely used AI algorithm in games.

Believe it or not but neural networks and other approaches of machine learning can be very useful in game development, e. g. for procedural content generation. AI in game development is not just about adversarial game AI.

AlphaGo is a monte carlo tree search with a neural network driving the heuristic. MCTS is used in e. g. Rome II Total War's campaign adversarial game AI - and as far as I know on higher difficulties the AI is still getting massive help through cheating. So, yeah, if there were a good way to improve that process even further through e. g. neural networks as heuristics that'd be kinda cool? Besides the fact of course that Sega unfortunately does a terrible job at 'humanising' computer players to communicate intent etc. to give off the impression of an intelligent adversary, which is at least as important as having at least a semi-smart AI behind it.

Oh okay then I'll just ignore the fairly logical consensus of experts on that topic. But please do explain why you arrive at this surely well-founded recommendation?

 

A couple of posters have recommended resources for general AI, but since this is a game development community...

https://www.amazon.com/Artificial-Intelligence-Games-Ian-Millington/dp/0123747317/

 

Perhaps it's an academic perspective of "the intelligence isn't biased, it's the dataset that it's working from"? (Note that I don't agree with this as an excuse for a deployed system's behaviour more than once.)

From my limited reading on the matter there's nothing in the algorithms themselves which inherently, say, select people of colour for random tests more frequently than white people. But those algorithms use a bunch of historical data to inform their deicion making, that data was collected from a bunch of humans doing the job, and humans sure as heck carry strong biases!

From an engineering perspective it's good to realise that this is an issue so that it can be dealt with. From a practical perspective it's not a good excuse for why the problem hasn't been fixed, because dealing with problems in the input data is a part of developing an effective system.

 

Yep, looking back into nice story of racist M$ chatbot, which was suppose to be representing young girl, if I remember correctly. But it become really really naughty.

And only because, the chatbot was learning from conversations of people. Means was working on dynamically collected data sets.

And we know human nature. Which resulted, that people who knew that talking with chat bot, feed with most bizarre phrases. Often for jokes, but also seriously talking dirty. So like a child, learned a behavior, based on environment that was connected with. Yet racism is mostly coined perception of human adults. Kids don't have it naturally, other than curiosity.

 

That chatbot is pr disaster more than actually ai gone rogue, 4chan got air of the project and use that to troll m$.

But that's kinda the point, ai can be trolled and turn against its own purpose. And given I see the far right most extrem member flocking to ai, I expect nothing good for me, since there is already a history of them using the naive understanding of "normie" against people they don't like by using gaslighting to operate in impunity on broad day light. I mean Palmer lukey met Bannon, who is a big organizer of the alt right, and now is going into weapon tech ... It won't be any bias, there will be bad event and they will just say it was a mistake, got a slap on the wrist and laugh it off on their forums as usual. Of course most people don't follow these stuff because nobody want to associate them with these people, I mean what if you make a genuine mistake and then there is an investigation that found in your cache that you were on racist site, even if it's just keep up with what's happening? whoops you can get blasted, so most people would never do it.

So he is true, there is no bias lmao

 

IMO the only reason why Google and other software companies publishes "academic" papers is for vanity.
There is no reason to do so. They can make the software and describe how it works on their website.
The papers are pretty useless without the implementation details which amount to the source code. To make a paper "academic" means not writing any actual code. Sometimes the code is proprietary which means the paper is even more useless because there will be no source code to look at the implementation.

 

Go tell that to the many software engineers who implement solutions based on "white papers" and such.

The purpose of a white paper to to present what someone has learned from research so that others can apply it practically in other areas. It's a programmer's job to figure out the implementation details specific to their use case.

 

But there are papers by Deep Mind who describe their proprietary software which can't be implemented because they hold back the implementation! It's all a load of phony baloney, if you ask me. Papers are good for academic subjects like math and physics. But for a practical subject like programming, I think less so. Instead of writing a paper, write a library and an API.

 

Well duh everyone knows that. The other bits are pretty obvious.

Again, duh. That [in bold] was what I was discussing though.

Of course, defer to the ‘experts’. These experts (being human) are just as susceptible to bias as anyone else. Everyone seems to regard science a lot like a diety, it cannot be questioned nor can there be any dissent.

 

You're calling a whole field of people and their work "phoney baloney" based on a small set of papers?

Programming is just one part of a field called Computer Science.

 

You have some persecution going on or what lol? I hope you don't confuse science with identity politics, like there is an implicit undertone in your use of speech.

Science by definition is only questioning, with the major caveats that you must show proof, and if we want to go deeper, science isn't about knowledge, it's entirely about ignorance and uncovering it, turning it into knowledge, and what's left from the world when you remove everything we know? well what's left is what we don't know. So the uncovering of bias is something that as been demonstrated.

I mean just look at the work of Joy Buolamwini, she is a MIT graduate, and the AI there couldn't detect her face because she is a black woman, it's not like rocket science conspiracy level of bullshit, it's as basic as it can get, measure and see the result, if it's a face recognition but don't recognize black people, that's a bias, that's as clear as you can get. Unless you want to dissent from facts, then well you can and it's call fantasy.

 

Rubbish. The entire point of science is questioning things, including each other's results. That's why we have scientific method, peer review, experiments repeated under different conditions by different people, and so on and so forth.

Everyone is absolutely welcome to challenge science. More, it's encouraged! That's how it's advanced. But it does require that you first take the effort to understand it.

It's true that all humans are subject to bias. A true expert will at least be attempting to control for such things, though, which is a part of why we have the stuff I mentioned above.

Edit: Wait wait wait... so your argument about why we should ignore a reported bias is that we can't trust the claims because everything is biased? Doesn't that support the initial idea that the generated results are, in fact... biased?

 

@yoonitee I think you got a bit of misconception, regarding academia.

While writing papers is often not the most favor part of study, papers are tend only indicate the right path, on which given subject has been discussed. They likely only skim problem, where more described in details, is in corresponding chapter of the study. Leading to Dissertation for example.
From there, you have reference to relevant subjects.

Mind, many people writing stuff with complex algorithms and putting on the web are from academia. You got people which focuses on narrow subject. And publishing. Then people like me, or you, we can look online for good and novel implementations.

This is field, where lot of unique solutions are getting born. And many often are as MIT for example, while some proprietary.

There is plenty of course code, which is written by less scientific people, but willing to learn and cracking variaty the problems.

But without publications, message wouldn't be delivered about existing solutions.

 

I'd report that statement but I don't believe in shutting down speech, regardless of what modern society thinks.

Oh I've seen that video. I'm not convinced. That bit in bold tells me nothing other than the fact that the AI had trouble seeing people of some ethnicity, if there isn't confirmation and selection bias happening.This whole area is subjective, and interpretation ought to be left to those who can think as objectively as possible.

 

Just reporting something doesn't mean the moderators will act on it.

 

@Braineeee
I think you are definitely confusing what a bias is, bias is not moral judgement of character, trouble of seeing people of some ethnicity isn't the bias anyway, it's that the algorithm was battle tested on pale skin, not on dark skin, dark skin is the problem not ethnicity. Like I said you confuse science with identity politics. It's not about the ai purposefully selecting to ignore.

And by the way, when talking about people being bias, that's also what we mean, that the basis of their judgement has flawed that favor certain outcomes, not that they are malicious. Hence why I'm calling you persecuted, you aren't able to look at this objectively and got caught in political argument for something that definitely rather straightforward. But hey I guess modern world create insecurities when the news are full of bullshit drama.

 

Objects contours and fractions are harder to be visible on the black surface, since black color don't reflect the light.
Shades of black will of course reflect in respect to the color and mat.
Just a more technical challange, than on bright object.
Not understanding such principles, my lead to conclusions like bias.

 

yep that's true, and that's also true that the algorithm was not properly tested before deployment, it's a widely used and considered strong until then, what's your point, that's still bias anyway you slice it.

Why? because they could fix it and it has been.

 

It's famous, The training sets (AKA the pictures they used) were mostly white men. It just never occurred to think about it. A quick "facial recognition can't see black people" search got this article from the popular press, including the stats:

https://www.nytimes.com/2018/02/09/technology/facial-recognition-race-artificial-intelligence.html

The thing is, it's not specifically an AI problem. Heart attacks are the same way -- for 20+ years they researched only white men and were shocked, shocked, to find women didn't get the sharp arm pain or respond to the same treatments.

 

My mom fell into that trap, back in the 1980s.
It took them a full year and several sets of doctors to sort out a set of pills that would keep her going.

 

How about we start another Unity vs Unreal thread, with an addendum about what AAA means, so we have something contentious to talk about that's actually related to game development..?

 

Revive those threads. Do it. Make it interesting.

So there WAS an original nefarious M$ purpose! Ai hasn't gone rogue. It did exactly what it was made to do. It was m$ that revealed its expectation bias.
D*cks out for Tay!

 

Yeah. Science. Right. ha. More like Scientology.

 

Back in the day in Unity Answers every moderator was in agreement about the "too general Q probably based on an incorrect assumption". Like someone would ask how to learn "rotations." The only reply (sometimes upvoted) was "what specially do you need to do?" It always turned out they really wanted to learn animation, or merely how to save and restore a rotation, or some rando told them matrixes and cosines were vital.

The only serious replies to this thread are things like "where are you stuck in the game you're making?" Everyone knows that. But, I assume, they also know the OP is just messing around, so they are too.

 

The thing is Unity has ML AI and some blogs about it but not one linked to a Learn AI with Unity Page, which would probably be way more fun than a dry book.

https://unity3d.com/machine-learning



Dog Training ML - https://blogs.unity3d.com/2018/10/0...8.1499034433.1544206627-1708636537.1542735635

 

Sure, it would be more fun. However, it would be less useful. To master AI, you need to actually read and comprehend fundamental concepts on AI and data.

 

Arowx have you ever heard of the saying "Don't judge a book by its cover"? Yeah, sure, AI is 'boring' if you care so little as not to want to know more about it. In which case I don't understand why you continue to make board topics such as this.

I enjoy every minute of reading.

 

Yeah, it's also way less useful.

 

Details be damned the only thing I see here is two people and a number of trolls who saw something they didn't like and felt the need to "correct" the other. Such is the reason we're having this conversation. Yeah, I made a statement, you guys could have read past it and simply ignored it but that isn't what happened.

If you gotta bully someone into switching sides to save their own skin like some groups do, it proves nothing other than the fact that you're a jerk.

The fact that the moderators do essentially nothing to prevent abuse against certain people here is egregious.

 

If wanting just some ML fun, play evolution (made in Unity).

https://keiwan.itch.io/evolution

But that won't make us any particular smarter on AI, of how this actually work, without deeper study = reading papers / books / source codes.

 

This is a discussion board. If someone makes a post it's a safe bet someone will comment on it, but your statement seems to imply that you're above the rest of us and that no one has the right to comment on any of the posts you've made.

Correcting someone isn't bullying them. Otherwise school teachers, college professors, and your own parents would be bullies, but they're clearly not.

Just because you view it as bullying or trolling doesn't mean it's bullying or trolling. If moderation jumped into action every time someone thought they were being attacked regardless of whether it were true these forums would be empty.

 

AI is a bit of a magical term. The more you know about how it works
the less you'll view it as 'AI'. A TV is a pretty simple device to us, but not to a caveman.

 

You can say that pretty about anything.

 

I've dabbled in DL/ML a bit in the last few months and would like to add a few ressources.

First, i'd recommend the courses on https://www.fast.ai/
They are programming centric and very practical. Free.

Second, you could check out https://www.kaggle.com/
It's a website for data science competitions where you'll find a lot of tutorials and example code. Kaggle also lets you use a GPU with their kernels (cloud programming environment). Free.

https://www.coursera.org/specializations/deep-learning is more for a broad overview, less hands on than fast.ai. I'd say... hmm... edutainment. You can watch the videos for free, but the option is well hidden if i remember correctly.

fast.ai is a lot more work if you dont just watch the videos but work through the stuff.

You also get to use a GPU for free on https://colab.research.google.com/ for non-kaggle stuff or if the kaggle kernels make problems.

Edit: Reinforcement learning is not covered by either course if i remember correctly.