The first bit is why it is called Stochastic gradient decent. You follow the gradient of a randomly chosen minibatch at each step. It basically makes you "vibrate" down along the gradient.
62 karma · joined March 11, 2019
Gravity, in this framework, is an emergent property arising from the statistical behavior of the hypergraph's evolution, suggesting that gravity is an "entropic force" arising from the tendency of the system to minimize its computational complexity
A better way of thinking about qualia is like embeddings in a neural network. Every time you run the training from a random initialization you will get different resulting embedding, but given the same training data all embeddings will all essentially be equivalent under rotation. I.e., your internal representation of blue might be very different from mine in an absolute sense but our relation between the representation of different colors will be roughly the same.
AIXI is a definition of the optimal agent and is hence uncomputable but LLMs are approximations which are approaching AIXI. I recommend Fridman's interview with Hutter.
In my mind, that the known laws of nature do not permit "free will" in the way incompatibilists define it is trivial. (regardless of whether the universe is truly deterministic or also has some randomness sprinkled on top)
A quote from it: "Other compatibilists show less concern in rebutting the conclusion that the freedom to do otherwise is incompatible with determinism. Compatibilists of this stripe reject the idea that such freedom is necessary for meaningful forms of free will (e.g., Frankfurt 1969, 1971; Watson 1975, Dennett 1984)—the “varieties of free will worth wanting,” (Dennett 1984). And even more notably, some compatibilists simply deny that freedom of this sort is in any way connected to morally responsible agency (e.g., Fischer 1994, Fischer & Ravizza 1998, Scanlon 1998, Wallace 1994, Sartorio 2016)."
This is the position of the paper above essentially. It references Dennett. The kind of freedom that matters, and that ppl are talking about in everyday life, is not the type of freedom incompatibilists reject. It still makes sense to talk about freedom, as in "freedom of opinion" etc, even if agents could not have chosen differently.
The paper explains this position from the perspective of reinforcement learning, and also gives a theory for why it is beneficial for intelligent agents to model themselves as being able to have chosen differently even if they actually could not.
Defining free will: Compatibilists often define an instance of "free will" as one in which the agent had the freedom to act according to their own motivation. That is, the agent was not coerced or restrained. Arthur Schopenhauer famously said: "Man can do what he wills but he cannot will what he wills."[14] In other words, although an agent may often be free to act according to a motive, the nature of that motive is determined. This definition of free will does not rely on the truth or falsity of causal determinism.[2] This view also makes free will close to autonomy, the ability to live according to one's own rules, as opposed to being submitted to external domination. https://en.wikipedia.org/wiki/Compatibilism
"By clearly defining what freedom is, regardless of whether or not determinism holds, our conceptualization of value freedom allows for a compatibilist view on free will. We are not disagreeing with the classic incompatibilist argument (McKenna & Coates, 2021) that determinism does not allow for free will; rather, what we propose is that the kind of freedom lay people actually refer to in everyday language is not the same kind of freedom incompatibilists reject. This distinction between what we here call physical freedom vs. value freedom allows for several potentially interesting perspectives."
The confusion about free will comes from a confusion about what "free" means. The everyday use of the word free has little to do with determinism etc.
The thing is that I can't prove that you or anyone else is conscious either, including myself. With conscious acting AIs it will be the same, and these AIs will believe they are conscious in the same way as we do. So yes, we will have to treat them as if they are conscious.
Yes, I do. I guess it all boils down to whether or not you think the hard problem of consciousness is actually a problem or not. I doubt it is, but it is a totally respectable position if you do :)
> Where exactly in the cited paper is the claim that emotions are a belief? Can't find it.
As I wrote in my previous response, you can call it predictions too. Predictions are usually beliefs about the future. In the predictive brain literature, it is also used as predictions about the present. I use belief as the output of some inference algorithm that needs to deal with uncertainty.
Here is a quote from the paper: "The brain continually constructs concepts and creates categories to identify what the sensory inputs are, infers a causal explanation for what caused them, and drives action plans for what to do about them. When the internal model creates an emotion concept, the eventual categorization results in an instance of emotion."
But you are right, the paper talks more about how emotional categories are created, and dodges the question of how the "subjective experience of having an emotion" emerges. In my mind, the step is not far though, and boils down to, as said above, how you view the hard problem or consciousness. That we "feel" stuff is a result of an algorithm/model that describes ourselves as having experiences, which is a good model, since how else would we describe ourselves? My belief is that the progress in AI and neuroscience will prove me right :)
Of course it is real. When I feel pain from an injury, the injury is very real, my pain sensor's signals are real and my brain's interpretation of those signals is real, and my behavior of trying to avoid further injury is real.
> Again, is there any reliable source for that other than a whole book that probably doesn't even talk about what you claim? If it is in the book, can you give a detailed reference where to find it? And if so, does the book provide references to research underpinning those claims?
Lisa Barrett is a very famous professor of neuroscience, so sure there is a lot of research in that book :) Here is a short video where she explains it https://www.youtube.com/watch?v=0QfCvIJRtE0 I also recommend her interviews with Lex Fridman.
She talks more about emotions as predictions, but we mean the same thing.
Here is a highly cited paper where she explains the theory: https://pubmed.ncbi.nlm.nih.gov/27798257/
> Or did you just come up with that on your own? It sounds absolutely ridiculous to me. Any attempt to "model" the experience is prone to fail by lack of evidence because pain is, by definition, subjective.
Your pain is very subjective yes, but that just means that it is created by your internal model of the world and yourself, which has a unique subjective perspective of the world.
> If you just search for the definitions of belief, emotion and pain, you will quickly find that their meanings are distinct. Claiming that pain is a belief is absurd.
These words are normally not used in the way I use them, but the point is that emotions are outputs of our model of ourselves. These outputs you can call beliefs, predictions or interpretations if you like.
The system would believe it is experiencing pain. Its model of itself would use the pain concept to describe itself, and it would believe it is experiencing it. And this description would be accurate since it is getting damaged.
I would argue that pain, like any interpretation of sensory input, are believes. "experience of a feeling" is equal to interpreting sensory input in a certain way.
But due to the way we are constructed, we cannot unbelieve certain things. It is like when you see a tree, you cannot decide to not see a tree after you have seen it. Or when you see a 2D projection of a 3D cube. You have to see the cube, even if you know it is not there since it is actually just a bunch of lines on a paper. This makes this type of belief different from other types of beliefs we hold which we easily can change our mind about, like what the weather will be like tomorrow.
It's the same thing with pain. You cannot stop believing in the interpretation of the sensory data that we describe as pain. And this "pain belief" is tied to certain behaviors, like trying to avoid it, since it is built into our reward system that it is negative. The reason we cannot easily "unbelieve" pain is that it would be very dangerous if we could simply ignore it. This is also why we are wired so that love is so hard to unbelieve, reproduction would fail in a species that could easily change their mind on the "love belief" about their offspring or partners.
If you think about phantom limb pain, it is quite clear that it is just a belief/model, and in that case an inaccurate one.
For a system to experience what we describe as pain, it needs to receive sensory input which it predicts signals that something is damaging its internal state, and then model/describe/represent itself as something that has this experience. Such as system would, like you do, believe it is experiencing pain.
The open question is why and how humans can "feel" guilt. How can a machine have a subjective experience at all? What is a subjective experience? This is the so-called "hard problem".
It becomes a similar question as asking: if god or "something" created the universe, what created that something. Its turtles all the way down yes.
>"it is part of the natural system we find our selves in". Like exactly everything else that surrounds us :)
This view is based on a misunderstanding of what moral responsibility is. Moral responsibility is essentially credit assignment [1]. It exists as a concept not because it matters whether or not you actually could have done differently before, but rather as a mechanism to teach you to do differently in a similar situation in the future. "You" is just a machine making decisions. When you make a bad decision, according to the other machines surrounding you (society) or your own brain's reward system, you are assigned blame or feel guilty because you should avoid making the same decision again in a similar situation.
To say that moral responsibility requires something supernatural doesn't make sense, since we successfully use the concept all the time. Its a mechanism for multi-agent learning.
Here is a great comparison between different noninvasive brain stimulation techniques for depression: https://www.bmj.com/content/364/bmj.l1079
Here is a great comparison of noninvasive brain stimulation techniques for depression treatment: https://www.bmj.com/content/364/bmj.l1079
A company that offers tDCS as a home treatment for depression can be found here: https://flowneuroscience.com/