Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Model-Free, Model-Based, and General Intelligence [pdf]

ijcai.org

Post reply on HN
Print view View original

21–30 of 32 posts

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#21

Post by eli_gottlieb »

naasking wrote: post #19
eli_gottlieb wrote: post #18

Earlier quoted context omitted.

Bayesian brain theories are more "in vogue", along with various other theories saying that the brain does some forms of statistical and causal learning and inference.

Sure, that seems reasonable. But I don't see why statistical and causal learning and inference really preclude the evolution of a system 1/2 dualism.

They don't preclude it, but they didn't happen to include it in our particular history. In particular, in the evolutionary history of the brain as an energy-optimizing controller of the body, a "System 1" would have been selected against extremely early on, when it directed the internal organs to act according to "heuristics" that wasted calories.

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#22

Post by JoshTriplett »

nmca wrote: post #8
nutjob2 wrote: post #5

It seems the fundamental problem with bottom up/learning AI is that it is opaque and essentially unknowable. I find it all very hackish. We can develop systems now which we can test and seem to work, but we don't know exactly why they work (eg: what parts of the training data they are promoting) and when (or why) they will fail. The effectiveness of adversarial inputs to trained vision systems illustrates this. Zoom…

One approach being considered is "AI Safety Via Debate"[0], which hopes to prevent deception by carefully constructing games in which a superhuman agent's best strategy is honesty. Note that this is the goal; much work to be done! [0] https://arxiv.org/abs/1805.00899

> which hopes to prevent deception by carefully constructing games in which a superhuman agent's best strategy is honesty

I'd be very hesitant to assume that an agent cannot learn under which circumstances it should be honest to gain a benefit without putting any innate value on honesty. A human agent is more than capable of reasoning like that, let alone a superhuman one.

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#23

Post by naasking »

eli_gottlieb wrote: post #21
naasking wrote: post #19

Earlier quoted context omitted.

Sure, that seems reasonable. But I don't see why statistical and causal learning and inference really preclude the evolution of a system 1/2 dualism.

They don't preclude it, but they didn't happen to include it in our particular history. In particular, in the evolutionary history of the brain as an energy-optimizing controller of the body , a "System 1" would have been selected against extremely early on, when it directed the internal organs to act according to "heuristics" that wasted calories.

> In particular, in the evolutionary history of the brain as an energy-optimizing controller of the body, a "System 1" would have been selected against extremely early on, when it directed the internal organs to act according to "heuristics" that wasted calories.

How are the calories are wasted exactly? Our hind brain triggers autonomous reactions to various inputs. Clearly not all such reactions are adaptive, sometimes staying very still and bearing some pain or discomfort is better than death. And so we evolved higher level cognitive faculties to make better choices, just a little slower than the hind brain. This is system 1.

I don't see why the exact same pressures couldn't work at this cognitive level as well. System 1 provided more adaptive reactions to a wider range of situations, but just a little slower than the hind brain. But even still, some metacognitive faculty would yield even better reactions in some circumstances, and so we evolved system 2.

But system 1 still has tremendous significance, because it's much better than our hind brain, is sufficient for most daily scenarios, and is not as calorically expensive as system 2.

The logic behind the efficiency gains is similar to the cache hierarchy in computers. We have more than one cache level because 2-3 cache levels is pretty close to optimal when trading off density, thermal considerations, and efficiency.

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#24

Post by eli_gottlieb »

naasking wrote: post #23
eli_gottlieb wrote: post #21

Earlier quoted context omitted.

They don't preclude it, but they didn't happen to include it in our particular history. In particular, in the evolutionary history of the brain as an energy-optimizing controller of the body , a "System 1" would have been selected against extremely early on, when it directed the internal organs to act according to "heuristics" that wasted calories.

> In particular, in the evolutionary history of the brain as an energy-optimizing controller of the body, a "System 1" would have been selected against extremely early on, when it directed the internal organs to act according to "heuristics" that wasted calories. How are the calories are wasted exactly? Our hind brain triggers autonomous reactions to various inputs. Clearly not all such reactions are adaptive, someti…

>How are the calories are wasted exactly? Our hind brain triggers autonomous reactions to various inputs. Clearly not all such reactions are adaptive, sometimes staying very still and bearing some pain or discomfort is better than death. And so we evolved higher level cognitive faculties to make better choices, just a little slower than the hind brain. This is system 1.

This story is false. The autonomic system isn't autonomous from the rest of the brain: its parameters, "set trajectories", are regulated by the rest of the brain (meaning: the limbic areas of the cortex), while the hypothalamus communicates with the endocrine system to predict and control the body from that angle. As you already say, a truly reactive body-regulator would get you killed very quickly.

Regulating the body by anticipating what it has to do is The Point of a brain. Further, the brain has six intrinsic networks to its functional connectivity, not two modular systems.

There's just no empirical evidence for a dual-process model. There's empirical evidence for an embodied predictive-control model. If you want to arrange this into "layers" from "animalistic" to "human", the way to do it would be to section off the particular cognitive functions which, at times, can be used for offline simulation of the environment, as a mode of metabolic reinvestment of surpluses.

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#25

Post by naasking »

eli_gottlieb wrote: post #24
naasking wrote: post #23

Earlier quoted context omitted.

> In particular, in the evolutionary history of the brain as an energy-optimizing controller of the body, a "System 1" would have been selected against extremely early on, when it directed the internal organs to act according to "heuristics" that wasted calories. How are the calories are wasted exactly? Our hind brain triggers autonomous reactions to various inputs. Clearly not all such reactions are adaptive, someti…

>How are the calories are wasted exactly? Our hind brain triggers autonomous reactions to various inputs. Clearly not all such reactions are adaptive, sometimes staying very still and bearing some pain or discomfort is better than death. And so we evolved higher level cognitive faculties to make better choices, just a little slower than the hind brain. This is system 1. This story is false. The autonomic system isn't…

> There's just no empirical evidence for a dual-process model.

I'm not sure the consensus is as strong as you imply:

https://en.wikipedia.org/wiki/Dual_process_theory#Evidence

https://scottbarrykaufman.com/wp-content/uploads/2014/04/dua...

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#26

Post by eli_gottlieb »

naasking wrote: post #25
eli_gottlieb wrote: post #24

Earlier quoted context omitted.

>How are the calories are wasted exactly? Our hind brain triggers autonomous reactions to various inputs. Clearly not all such reactions are adaptive, sometimes staying very still and bearing some pain or discomfort is better than death. And so we evolved higher level cognitive faculties to make better choices, just a little slower than the hind brain. This is system 1. This story is false. The autonomic system isn't…

> There's just no empirical evidence for a dual-process model. I'm not sure the consensus is as strong as you imply: https://en.wikipedia.org/wiki/Dual_process_theory#Evidence https://scottbarrykaufman.com/wp-content/uploads/2014/04/dua...

I don't see any direct, empirical neuroscientific evidence in that "Evidence" tab. Unfortunately, many psychology experiments allow you to fit any remotely reasonable theory to the data.

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#27

Post by naasking »

eli_gottlieb wrote: post #26
naasking wrote: post #25

Earlier quoted context omitted.

> There's just no empirical evidence for a dual-process model. I'm not sure the consensus is as strong as you imply: https://en.wikipedia.org/wiki/Dual_process_theory#Evidence https://scottbarrykaufman.com/wp-content/uploads/2014/04/dua...

I don't see any direct, empirical neuroscientific evidence in that "Evidence" tab. Unfortunately, many psychology experiments allow you to fit any remotely reasonable theory to the data.

So then you'd agree that there is as much empirical evidence for a dual process model as there is for many other models. This seems a little broader than your original claim that there is "just no empirical evidence for a dual process model", which suggests that there is evidence for a more realistic model which should be preferred.

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#28

Post by eli_gottlieb »

naasking wrote: post #27
eli_gottlieb wrote: post #26

Earlier quoted context omitted.

I don't see any direct, empirical neuroscientific evidence in that "Evidence" tab. Unfortunately, many psychology experiments allow you to fit any remotely reasonable theory to the data.

So then you'd agree that there is as much empirical evidence for a dual process model as there is for many other models. This seems a little broader than your original claim that there is "just no empirical evidence for a dual process model", which suggests that there is evidence for a more realistic model which should be preferred.

>So then you'd agree that there is as much empirical evidence for a dual process model as there is for many other models.

No, because a dual process model spreads itself too thin: it doesn't coherently explain many experiments with a single theory, but instead rewrites the theory for every distinct experiment.

Re: Model-Free, Model-Based, and General Intelligence [pdf]

#30

Post by krageon »

nutjob2 wrote: post #11
nmca wrote: post #8

Earlier quoted context omitted.

One approach being considered is "AI Safety Via Debate"[0], which hopes to prevent deception by carefully constructing games in which a superhuman agent's best strategy is honesty. Note that this is the goal; much work to be done! [0] https://arxiv.org/abs/1805.00899

Unfortunatley being dishonest or evil is just one example. Arguably the AI can develop new classes of deviancy, abuse or maladaptation that we haven't conceptualized yet. We supersize the ability, surely we supersize the problems. It leads to a scary question: what does a superhuman AI really want?

What does a human that is much smarter than you really want? It's a fundamental philosophical problem that hasn't been solved.
Post reply on HN