This clip is from the commentary pre-show of game 3 of AlphaGo vs Lee Sedol:
https://www.youtube.com/watch?v=qUAmTYHEyM8&t=20m52s

The developer mentioned how a neural network go bot can reach dan level without reading ahead.

Does this mean one can reach 1 dan with only good memory and extensive practice so as to be exposed to local and global stone patterns allowing for pattern recognition without reading ahead at all?

Dwyrin mentions a few times in his videos that until you're around 2d, you only need to read 2/3 moves ahead.

In my own experience, there is some truth in that. If you play according to general principles, follow josekis, do not try too hard to kill stuff and are careful not to have groups that need to live in a small area, it is possible to play a decent game with minimal reading.

That being said, there will always be some sequences en each game that need deeper reading (because your opponent did not follow the joseki or because his moyo is too big not be invaded, etc) and that is part of the fun too. Living sente in your opponent's moyo because you've carefully read that you're alive even if it may not seem so at first glance is a great pleasure

Plus, acquiring reading power is maybe the easiest part of go, so why would you deprive yourself from that?

A 1 dan player can typically see the answer to easy tsumego/tesuji problems without doing any reading, due to pattern recognition. They read it for verification. The neural net doesn't have that verification step, so if its close but doesn't quite work, the neural network will fail the problem.

It's not memorization, just practice and familiarity over thousands of problems. And that neural net is likely trained on millions of examples.

_________________
Go Books

It's hard to compare how alphago work to how the brain work
Does neural network equal to memory and practice?

Practice, yes. Memory? Not so much.

Neural nets are also used in image recognition, so let's take that as an example.

Say you give a NN 100 images of cats, and 100 images of dogs, and train it on this. When you test, if you give it one of those images of cats/dogs, then it'll correctly guess 98% of the time whether it is a cat or dog. It hasn't memorized, but it has recognized the pattern. If you give it a new image of a cat or dog, it might guess with 90-92% accuracy. Again, based on extrapolated patterns it has collected from the 200 images. Neural nets have no memory, but excel with pattern recognition.

_________________
Go Books

I would argue that memory is a necessity to pattern recognition because without memory how can you match a pattern you see to a pattern you may know (ie. in memory)?

That depends upon how you practice and what you mean by reading. If you mean without conscious calculation of variations, then yes. (But that's only part of reading. )

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

By matching upon characteristics of the given pattern.

You can memorize as many joseki as you want, but as soon as someone deviates, how do you play? Now, memory serves little purpose, and its pattern recognition and the strength of reading.

If you're arguing semantics, then I'll kindly skip the debate. I'm explaining how a neural network functions in this context.

_________________
Go Books

Neural nets have to store that "learned" pattern recognition somewhere -- in memory, or a database. It doesn't really matter what you call it or how its stored/recalled, its a form of memory.

Its not about memorizing joseki per se -- but remembering learned patterns in general. There are basic principles/patterns that should be common, so that when your opponent deviates, you'll be able to respond.

Example, training a neural network on a linear function:

f(x) = ax + b

The neural network will internally store values to represent a close approximation to a and b. They will not memorize values of x and f(x), but will know how to calculate the value for f(x) given x.

If you're talking about memorization of a and b? Certainly, those are stored in the neural net structure (in something slightly incomprehensible to the original formula). You could say it has memorized the formulas for calculating different positions. If you're talking about memorizing the values of x and f(x), then no, f(x) gets calculated each time you provide it x.

_________________
Go Books

Actually, when we learn a jôseki by heart, we don't know how the information is stored in our brain, but it is believed to look more like a neural network than a hard drive. Hence the name "neural network", by the way !

I will give this a try.

The neural net, is, itself, the memory.

It is like a slide rule. A slide rule does not remember, it just solves for the equation.

Neural nets are really analog machines, like a dial. They are not digital. A neural net on a computer, is really a digital representation, of a physical object, so to speak.

A watch does not remember the time, it just shows it.