But, if you don’t post, what is left for those of us who appreciate your posts?
1 Like
Make no mistake, the passing on of Smorgasboard from these boards will be difficult on all 7 of us who read and participate here. Maybe TMF will let us open up another board in his honor where we can regale each other with tales of the high credibility of his posts.
1 Like
What is the point of having an online community where people can try and discuss things in an informed manner if the people that make those attempts (and Smorg clearly does, I’ve “known’“ him since Tesla was first recommended in RB… so a ~dozen years).
All you’ll be left with is this crap that flows below Yahoo!Finace stories where people sh!t talk poliics and blame liberals/conservatives/minorities/LGBTQ/”white guys” whatever for anything and everything.
I am basically only over here on the free side to see what he posts since skepticism and positivism based on media stories is both rampant and near useless on its own.
1 Like
I’ve been reading and posting here for a few months and can’t recall a single political post in all that time other than the one you just made. Can you point to one of these political posts (other than your own) that are supposedly bogging down the board?
Let’s discuss.
From “SelfAnnointedMostCrediblePoster”
We get statements like:
(my bolding)
Does @rtichy agree with this statement?
That AI, such as implemented with neural networks, is NOT just a complex set of mathematical formulas?
Please go find any expert that agrees with that statement (might take awhile).
I’ll provide the answer.
That statement is flat wrong.
Fundamentally, basically wrong.
Someone with an introductory understanding of the math of neural networks knows that this betrays a fundamental lack of understanding of neural networks.
But anyway, good luck discussing that in an informed manner.
I tried here:
I just checked the robotaxi tracker and it looks like they got rid of the data on number of rides and number of robotaxi rides (at least I can’t find it). But it is still showing 9 unsupervised robotaxis in total.
Not a whole lot of growth this quarter.
I believe that a distinction is being made here which you might not be getting. There are lots of things which involve formulas at some level in some form, but which are not regarded as formulaic in character. Such are neural nets. If formulaic, one might imagine “reaching in” and adjusting some value in an existing formula. Instead, one has to feed it new data.
Ok, let’s explore.
A neural network is certainly not a single, simple formula, of course (“network”).
Which aspect of a neural network does not have mathematical formulas?
Maybe we need to be more pedantic and write “mathematical equations” instead of “mathematical formulas”?
If true, one could just have said “well technically not “formulas,” in general there are “equations,”” or something along those lines.
And, for this:
One could absolutely manually adjust a specific portion or value(s) of the mathematical formulas in a neural network.
One certainly trains a neural network on data, which is essentially a statistical process of parameter estimation (and thus the parameter values are adjusted in that way using that process - which is also just more math and formulas).
(not sure exactly what “reach in” means, you might have to clarify that)
Even assuming you could identify the place to twiddle, doing so is counter to the whole idea of neural nets.
That’s all fine.
That’s also not my claim.
My claim is that AI such as neural networks is built with mathematical formulas/equations/expressions that make a function.
Model training is a process, also with mathematical formulas/equations/expressions, that modifies that function.
All of that math takes text, sound, image input, represents it digitally, and produces a model output/prediction.
That model output is a chatbot response or vehicle driving behavior or a weather forecast or projectile trajectory or etc.
Why anyone might question the role mathematics plays in all this AI, esp. machine learning/neural nets and inference therein is highly confusing — kinda like thinking there is only one form of engineering redundancy.
I don’t mean this to sound critical of smorgasbord1 because I know he is very knowledgeable on all this stuff and i am thinking maybe there is just a disconnect somewhere. Hopefully there is an easy way to understand without him leaving these boards. Heck, we both left the paid boards for our respective reasons. Despite frequent disagreements on specifics, I have learned a lot and increased my powwn understanding from him. Let’s all be on good behaviour
Here’s my thoughts on the math question. Way back in my undergraduate days I had lots of probability and statistics classwork. When it involved theory, it was mostly classes from the math dept; I also had a few applied probability/statistics classes from outside math dept - both Industrial engineering and business classes – but these still were even more math.
Probabilistic inference and statistical inference both seem to have widespread application. They both seem to have important roles in analyzing data then building, training and using AI within models for robotaxis. One statistical theory class had major focus on Bayes theorem and Bayesian pre-posterior analysis. The former seems to have ‘uuuuuuge applicability here and less for the pre-posterior analysis.
Bottom line is that I don’t see how you separate neural nets/machine learning and the inference provided from all the math behind it. In one of our best selling products from the former software company I co-founded, we used a rudimentary rules-based AI engine from the 1980s-90s era. Even though all was explicitly defined, I guess one could say wading through all that data was ultimately just a complex decision tree which undoubtedly was just a big math problem too.
1 Like
I agree with your post.
Thank you for the thoughtful reply.
I don’t think (most) anyone here would say any other contributor doesn’t have something to add.
Best not to take oneself or this discussion too seriously.
I try to follow the data and applicable theory when those trail markers are available.
With all of the new “wide awareness” of AI and machine learning (the foundational ideas are not so new), I thought it helpful to myself and hopefully others to recognize that there is some basic, very familiar math that underlies what might seem to otherwise be a mysterious set of tools.
Because characterizing something as formula-based conjures up an image where, if something doesn’t work quite right, one can “reach in and twiddle a knob” to adjust the behavior, but one can’t do such a thing with neural nets. “Fixing” a neural net application implies new training and hoping that the new learning will be an improvement.
I don’t disagree with that. But even though the term “formula” was used in a prior post, preceding exchanges seemed to minimize the underlying math that is ultimately involved. I have little doubt that underlying math is formulaic in nature. And not just the actual math performed within these platforms but all the math theory underlying it.
Meanwhile I would have assumed you knew it already but models can be edited to clear up weighting issues which result in observable discrepancies. My experiences with AI aren’t deep enough to comment on pros and cons of doing so, but some capabilites do exist and will probably grow in usage. Let’s face it – the probabilistic nature of AI as used in robotaxis has it s shortcomings already. Tweaking may not be as thorough as new training, but if done by a knowledgeable support person is probably as good and more efficient than a new training cycle.
Edit: My knowledge of current AI implementations is not deep. If anything is wrong in what my perspective is above, I would appreciate someone noting corrections
1 Like
Yeah, no one does that…(tweaking the weights of a trained model to “fix a bug”) except maybe researchers, etc.
Here is why. Let’s say you do have a behavior that you want to remove or to add and you “somehow” spend a lot of effort to discover a trained weight that fixes it. And you do this and you release it to lots of vehicles. And now next month you retrain the model. How do you incorporate that same fix to maintain the code? All the weights changed, slightly and the thing you changed probably moved locations.
Mike
2 Likes
I agree with what you say. My point is merely that it does happen in some cases and certainly can be problematic which I didnt note as I should have.
1 Like
Point being that many people are going to hear formulaic and identify that as something they know about with well defined terms and values where it is easy to fiddle with a term to adjust a behavior … and neural nets are just not like that. So, one needs to fight the vocabulary.
1 Like
I agree, the need to define terms is important.
The terms are well defined in many places; thats not the real issues. It’s very easy to forget despite how much math I have had what the exact distinctions are between expressions, functions, equations and/or formulas. Based on looking at such definitions, I would say “formula” appears to be the most appropriate term for the preceding discussion though once new or updated weightings are created/updated any discussion all seems moot.
1 Like