Google’s AI-Building AI Is a Step Toward Self-Improving AI | Hacker News Reader

Google’s AI-Building AI Is a Step Toward Self-Improving AI

singularityhub.com

56 pointsmooreds24 comments

Often these metalearning systems have to pick from a library of already understood building blocks to build novel structures from. This limits the flexibility of their hyperparameter space. It's a bit like having a robot building a building but give it premade nails and wooden planks. It will never build a skyscraper. We have to figure out a way to also allow the system to come up with its own building materials so to speak.

It's not that no one has any idea how to avoid choosing building blocks but that it's a tradeoff. You're building in an inductive bias when you choose the building blocks for your deep RL or evolutionary algorithm optimizer. Nothing stops you from choosing building blocks which are Turing-complete - Schmidhuber, for example, experimented with evolving Brainfuck programs back in the '90s or '00s, I forget which. Turing-complete, can solve anything you set it eventually, as minimal and general as it gets. The problem is that there's so little inductive bias there that it takes forever to get anywhere useful, you have to evolve far too many samples. With the evolving CNNs, researchers are already joking about how you would need a nuclear reactor to reproduce some of these papers which train thousands of CNNs, so the inductive bias is really important in keeping samples down to a feasible level. As computing power gets cheaper and the AutoML-like tools learn more domain knowledge, it'll be possible to let them work on a more raw level than architecture design choices like 'convolution or fully-connected layer? ReLu or PRelu?'

It's very black and white right now (no knowledge or very constrained knowledge). The same problem exists with just plain old parameter learning, with DL models having so many free parameters it can make training more computationally expensive then it needs to be. For instance, I want to train a reinforcement learner to do some task in the real world (eg. a robot). It would be nice to be able to define a prior that puts very low to no probability on actions that are not physically possible. This would constrain the search space considerably. But it's not really clear how to do this with neural nets.

> It would be nice to be able to define a prior that puts very low to no probability on actions that are not physically possible. This would constrain the search space considerably. But it's not really clear how to do this with neural nets.

I dunno, I can think of several ways off the top of my head: 1. put large negative rewards on impossible actions; 2. only compute Q-values for feasible actions (just because DQN computes Q-values for a hardcoded set of actions doesn't mean you have to); 3. use Achiam's "Constrained Policy Optimization" https://arxiv.org/abs/1705.10528 ; 4. reparameterize the output to make illegal actions unrepresentable (domain-specific); 5. add illegal or not as an additional data input (ie in addition to the usual image or robot state, for the _n_ possible actions, include a _n_-long bitvector of possible/impossible), or include that as a regression target to make it predict whether each action is possible.

While deep learning can be a powerful enabling tool for saving time compared to simpler neural networks that required more manual feature selection and hand tuning, I believe that there is also some real technical debt using anything that is a black box.

I still believe the future is more hybrid symbolic and deep learning systems, but I may very well end up being wrong.

(The comment below incorporates an edited version of my comment from last week's thread.)

With human involvement only at the meta level, deep understanding of the generated implementations becomes more challenging and, in highly complex domains, perhaps impossible. See AutoML's generated architecture and compare that with a human-designed one. [1] [2]

A future advance could allow the machines to automatically pick and learn to do new tasks that are helpful to accomplish a given high-level mission. Therefore, chances of unintended consequences become much greater.

The major issue is, without a moral core that closely aligns with humanity's evolved morality, there will be moves that advanced AIs come up with that we deem abhorrent, and sometimes unforeseeable, yet they perform them innocently and we only find out the consequences once it is too late.

Microsoft's Tay tweet debacle was fairly inconsequential [3]. But AI is becoming more powerful and granted more and more authority over the real world we live in, can we afford the technology without humanity?

Developing a moral core should be prioritized.

[1] Google Blog on AutoML https://research.googleblog.com/2017/05/using-machine-learni...

[2] Many tasks in the real world are more complex than Go and even expert humans are not capable of understanding all possible complex interactions on the Go board.

See: AlphaGo, in context by Andrej Karpathy https://medium.com/@karpathy/alphago-in-context-c47718cb95a5

[3] https://www.theverge.com/2016/3/24/11297050/tay-microsoft-ch...

>Developing a moral core should be prioritized.

I just can't see how over the long term an AI will not evolve out of the human moral code we want. Humans evolved into our, very wide ranging sets of, moral code. There is not a consistent one across all cultures. The base set of morals we do follow are generally based on that we humans are made of squishy meat that suffers pain, and minimizing that pain is a priority goal.

We may be able teach AI about our weaknesses, but as machine intelligence meets and then exceeds human intelligence and capability, said AI will see both our strengths and weaknesses and that puts us in a precarious place. If an AI without human morality outperforms AI with human morality it stands to become the dominant form of AI.

AIs without human morality, which would put certain restrictions on them, will likely have less power in the long run.

This I agree and the consequences could be dire if we do nothing to prevent that.

Thus, the onus is on us to first develop AIs with moral core and a regulatory framework which incorporates our best knowledge from sciences and humanities, then task the moral AIs to help regulate the development of amoral ones.

As a side note, although many details of human morality are not completely consistent across cultures, there are many shared tenets. Some examples: "Do not kill one of your own who did not do something wrong." "Avoid incest." "Do not steal." Current crops of AIs are not even aware of such common sense.

Take "Avoid incest" for example, this is a human genealogical construct that does not map well to the machine experience. In humans too much inbreeding leads to accumulation of negative traits. In AI this is called over specialization. So, by those metrics, good AI learning systems already avoid that.

>"Do not kill one of your own who did not do something wrong."

Also is a very human subjective experience. We see the individual as the quantum of pain and experience and optimize on that. In the AI world killing off other AI that don't stand out as better than a baseline may be an optimal evolutionary strategy.

Also "Do not steal" is a terrible human morality baseline with a highly subjective gradient. Did you borrow a pen and not give it back, is that not technically theft? Is a business transaction where you used secret information to gain monetary advantage other the other party theft or not?

The problem with your first statement is you assume that human morality is the pinnacle of morality and all other forms will be lesser to that standard. This could lead to significant disruption for humanity of a more efficient or optimized form of morality comes along.

The examples of moral codes I gave were not meant to be used on AIs but by AIs working for/with humans.

> The problem with your first statement is you assume that human morality is the pinnacle of morality and all other forms will be lesser to that standard.

To be clear, my first statement was meant to read as: "AIs without human morality, which would put certain restrictions on them, will likely have less power than amoral AIs in the long run."

And "power" here means the ability to effect things in the world. I did not assume that human morality is the pinnacle. I agree with you that it is likely not an optimized one.

My whole argument rests on the principle that AIs that do not incorporate human morality into its utility calculus will have negative consequences on us.

And this goes back to my original statement that AI will eventually evolve past it.

AI will originally be the weak 'force'. It will have to evolve mimicry to survive. This will lead to an increase of its survivability in a human dominated world. Eventually this will lead to AI being the dominant 'lifeform'. It is at this point a rapid cascade to a machine oriented morality can occur, and I can't imagine that it would be good for the human species as a whole.

I'm reminded of an exchange in Neuromancer: "Nobody trusts those fuckers, you know that. Every AI ever built has an electromagnetic shotgun wired to its forehead."

Chatbots and board games are toys. When it's time for AI to be more serious, people will probably be a bit more careful with how they are used. I'm even inclined to wonder if the question of "will the humans try to deactivate me for doing this?" might be a close enough analog for "is this in line with human morals?".

> I'm even inclined to wonder if the question of "will the humans try to deactivate me for doing this?" might be a close enough analog for "is this in line with human morals?".

The problem with that statement, of course, is that the very next question it asks will be "Or can I stop them?"

That may take the form of convincing arguments, brute force or something we've never thought of yet, but the only way to avoid it is to prevent the AI from wanting to do anything we don't want it to.

Which is a much harder problem.

Or do I care? Were assume that AI worries about its future, it might not give a toss!

You're right; I misspoke. Making the AI indifferent to its own survival is indeed another way of improving safety.

That's still very difficult, though. Have you read the Basic AI Drives paper?

A seed AI is probably right around the corner, but just like the seed RNA life was a long way off from general life -- both in terms of time and resources required to evolve into complexity, the appearance of superintelligent AI will be constrained by the amount of resources we allocate to it. There is some danger of machines writing then spreading their own worms to usurp computation resources or ransomware to force humans to build them more, but there will be key steps along the way where there will be a human decision to both use and to allow AI to grow into something we do or do not want. It won't be something that springs up on us.

how can you, possibly, know this? conjectures about AI takeoff are just conjectures, but you claim to know that it "won't be something that springs up on us."

>conjectures about AI takeoff are just conjectures,

That's not actually true. It has happened, or I should say it is happening right now on Earth. For around 63 million years in the dominion of mammals there was not an intelligence explosion. And then, quite surprisingly if you look at the history of evolution, the line that would become Homo sapiens rapidly diverged from the rest of its family. As its intelligence increased it rapidly started rendering all large life on Earth extinct to become an apex predator that all other significantly sized life on the planet lives at its will.

This take off was so rapid and complete that Homo sapiens was not only able to conquer every biome on the planet, it was able to leave the planet itself. Massive alterations to biosphere occurred as its numbers swelled to over 5 billion. Almost every waterway on Earth has had significant alterations occur do to human development. Land use and forestation on the planet rapidly changed, with a majority of the changes occurring in a 300 year period.

So, Intelligence Explosions are not conjectures, we are in the middle of our own. Furthermore it seems unlikely that human design is the pinnacle of intelligence. Unfortunately we are left with a lot of questions, and fewer answers than we would like at this point. Just defining levels of intelligence at this point seems to be a debate that leads to arguments and heightened emotions.

This level of indirection is used in nature. Gene Regulatory Networks (modeled mathematically as ANNs) fine tuned by evolution control the neural development of the Neural Network which gives you the general architecture then further fine tuned by learning.

Waiting for Eliezer Yudkowski to deliver the daily portion of "we are all gonna die"...

Last I checked we were all going to die anyway

last i checked death is being disrupted by the blood of the young and innocent

Not today, but well, is he wrong?

Some time ago, I might have said so. I'm far less certain recently.

>is he wrong

He is unconvincing.

Reply on news.ycombinator.com