[yconst]

I found the first comment to this article quite interesting:

"Marcus has a point - even is some of what is said about deep neural networks is incorrect (for instance, they can learn and generalize from very few example, one shot learning).

However, he got it wrong with the answer. The key for machines to reach the symbolic abstraction level is the way we train them. All training algorithms, supervised, unsupervised or reinforcement learning with LSTM rely on the assumption that there is an "utility function" imposed by some external entity. Problem is, by doing so, we are taking away their capacity [of the machines] to make questions and create meaning.

The most important algorithm for learning is "meaning maximization" not utility maximization. The hard part is that we cannot define what is meaning - maybe we can't, I'm not sure. That is something I will be glad to discuss."

[thom]

You could always compare how much your perceived reality matches predictions from your mental model. If you look at the concept of 'schemas' in child development, you'll see children focusing on one thing over and over again, like 'containment'. Putting a block in a box, taking it out, putting it in, taking it out. Each time they're learning that when you put something inside something else, it's still there even if you can't see it, that you can pick them up together etc etc. They're slowly exercising the different ways they can interact with their environment, learning to model likely outcomes. Learn enough of these schemas and you can start making more complex plans to see what the world does in response.

[orthoganol]

> The hard part is that we cannot define what is meaning - maybe we can't, I'm not sure

I think you've correctly identified the million dollar question in AI, and that's why I believe the major breakthroughs in artificial general intelligence won't come from scientific researchers. They might even come from PG's dreaded literary theory types, ones who are also talented programmers (...Lit theory being maybe the only field that thinks about structuralism, deconstruction, how meaning is manufactured, etc. - things outright dismissed by the scientific community as useless, but which still hold potential to produce a 'meaning algorithm').

[emcq]

This is perhaps a bit philosophical, but assuming we had a "meaning maximization" function, what stops us from writing it as a loss function and using our current supervised machine learning frameworks?

[sawwit]

Just because we can formulate an optimization objective, it is not guaranteed that we will find an algorithm that solves it within a reasonable amount of time. In case of humans, these objectives or preferred states are possibly very simple ones, like hunger and pain avoidance, reproduction and curiosity; and it is actually easy to write down an algorithm that optimizes these objectives (if you ignore how reality actually works): Simply try out all possible ways of reacting to the environment and choose the best one.

This works in theory, but in practice you only have a limited amount of chances to try something out (because of the arrow of time). This makes learning a necessity. You need to keep a record of all trials you have performed so that you can reuse this information later when the same situation reoccurs. How to do this in an optimal way is described by Bayes' theorem.

The key to AI will be a certain set of priors, biases and fixed function units that make this computationally tractable; we'll likely need things like invariance of the learned information to various changes so that it can be reused in different settings, segmentation of data coming from the world into episodes (hippocampus), attention, control (basal ganglia), mental rotation (cortex) and path integration (hippocampus, grid cells).

[emcq]

That's true, there are certainly many optimization objectives computationally intractable, or perhaps too abstract to be useful for learning.

However, I would argue the prior of Bayesian modeling can be just as nebulous and computationally intractable as an optimization objective. Like supervised learning, Bayesian modeling is just a tool.

I'm skeptical that we will reach AI through a deep understanding or modeling of the brain. Technology and computer science advances more quickly than the biological sciences, at least in recent times. You might argue a success in robotics like [0] is a motor control system. But they built this extending mathematical frameworks not being biologically inspired, and the big wins there didn't come from fixating on a learning framework or biological mimicry; just like humans learning to fly didn't come about from flapping wings like a bird. At some point we hacked an engine (invented for other purposes) onto a wing and came up with powered flight.

As an aside, only seeing input a limited number of times would likely improve your ability to find models that generalize as your model must be able to take these one off learnings and unify them in some way to achieve high training performance. With respect to human learning, a specific individual only has one chance, but nature has had many. We are only a selection of those chances that seemed to work well enough. There are many commonalities to existence that allow for this to work well in practice.

[0] http://groups.csail.mit.edu/rrg/papers/icra12_aggressive_fli...

[doctorpangloss]

Your agent may need a way to ask for a "kind" of training instance from the world in order to maximize meaning. Like maybe I've seen mammals, and now I need to see another kind of animal to maximize my understanding of the meaning animal. A human being—perhaps instinctively, or perhaps by some other force—has the curiosity to go find / pay attention to fish and birds. A kid can tell you he wants to go to the zoo.

A supervised machine learning framework can't tell the researcher what training instances it needs to see in order to improve its meaning. A supervised learning framework can't imagine where it might find that training instance, or describe what it may look like. A supervised learning framework never asks to go to the zoo.

[saulrh]

Yes, in theory an offline supervised learner should never beat an online reinforcement learner. Adding a set of actions A that can be used to bias future examples in a predictable manner is certainly an advantage that will yield better convergence properties in almost all scenarios, simply because it lets you gain more information per observation.

[jcfrei]

I'm just curious: Can there be intelligence without some externally imposed utility function? For us humans it usually consists of access to (in decreasing order of importance) oxygen, water, nutrition, shelter, closure, etc. If an AI was free of all these constraints - meaning it had no predetermined utility function to maximize - then what is it supposed to do? It would have no motivation, no reason to learn or do anything. And if it had no goals which it tries to reach, how could we determine its performance or rather its intelligence?

[yconst]

Interesting point. However, I'd like to argue that there is a fine line between the conventional definition of a utility function and the ability of an organism to survive in an environment. The latter is truly open-ended, and not really externally imposed, i.e. different animals/people fare throughout life "optimising" completely different functions (closure?). A quote from a Kozma paper comes to mind: "Intelligence is characterized by the flexible and creative pursuit of endogenously defined goals". I believe this quite well summarises the open-ended nature of the task at hand. As I understand it, indeed, goals, rewards, risks, hazards, they are all in the game and shape the decision-making of the agent, it's "policy". But the way they are formalized for each individual and situation, well, is probably subject to constant redefinition itself.

[JulianMorrison]

> All training algorithms, supervised, unsupervised or reinforcement learning with LSTM rely on the assumption that there is an "utility function" imposed by some external entity. Problem is, by doing so, we are taking away their capacity [of the machines] to make questions and create meaning.

It's just a simplification. We are consuming the algorithm's output, rather than it being for internal consumption in a larger mind. So we supply the goal, where in a larger more natural mind the goal comes from other parts of the mind - still computing ultimately against a fixed utility function. That function being our human instinctual basis and the "firmware" of culture it's primed to load.

[joe_the_user]

Well,

It's not just training that implies utility maximization, basic internal processes of a neural network, such as back propagation, are based on the utility function maximization. Back propagation is an approach that allows the network to do gradient descent on a utility function by propagating errors through the system.[1] Back propagation has been the way that deep networks have become tunable and effective. Without a utility function, it's hard to see how one would tune them.

And it's hard to think of what "mean maximization" could be but maximizing a utility function based on meaning.

And while you can construct a neural network that learns from a single example, the nn framework is fundamentally based on learning from multiple examples so such a construct is mostly meaningless.

But the point about machine learning being limited by the training methodology is good, it's just I'm pretty sure you'd different algorithms if you used a different methodology, the existing algorithms are used specifically because they fit the methodology.

[1]https://en.wikipedia.org/wiki/Backpropagation