agree. fake sounds prone to erroneous data, while some systems are fairly easy to simulate, generating real data from a simulation. for more complex scenarios this is trickier, and a faulty simulation moght leave subtle issues in the networks functioning. if u dont have real.data to verify it with u will be in the dark about real accuracy. depending on the networks purpose, this moght be really dangerous.
What's weird about this to me is the flow of information, knowledge from one place to another.
We're using software that has encoded the rules we want to use to generate data. The ML is using the data to infer the rules. It seems like we need a better way to give our ML algorithms the rules that we already know rather than this processing-intensive process.
But perhaps that's the cost of using neural networks- they can learn any function, but they can only learn from data, examples.
I think "giving the rules we already know" is only a good idea if we are 100% sure that we know 100% of the rules necessary with 100% accuracy. The fuzzyness of NNs is what in many cases allows them to work as well as they do, and training on close-to-perfect data is going to preserve this aspect better than close-to-perfect rules could.
What you’re describing would be an Expert System, part of GOFAI.
Neural nets work by generalizing. This is important because GOFAI systems are brittle. We are never able to encode all the relevant rules for real world situations by hand.
Also, note that the synthetic data is generated from rules about how the world looks, not rules about what the model should do.
There may still be some direction for zapping the rules-of-the-world right into the circuitry of, say, a subnet of the model, without doing backprop. Something like a one-shot initialization of the network to approximate the given distribution.
I think what he's saying is that currently we can get one of these expert systems to create 1000 examples with the encoded rules, which a NN might then generalize correctly or it may come up with a completely batshit insane ruleset that overfits on that data. There's no way for us to actually feed the NN the rules directly, like one would give them to a human brain.
The main issue with expert systems appeared to be not the amount of necessary rules but pathway dependency on representation model for the problem domain. If it mapped well, it could perform exceptionally well. If not, you got into all sort of combinatory troubles real quick and had to wing it (or as they used to say in the 1980s, to use heuristic rules).
The combinatory bit due to its branching nature lends itself poorly to leveraging SIMD hardware as statistical solutions do. And getting a right representation model is a substantial challenge. Everyone knew the importance of that ('knowledge engineering') by the onset of AI Winter but nobody had any good methods.
We don't have the rules that we want the neural network to learn, if we did we could just directly use those rules, and there would be no need for ML to solve the problem. In this case we want the ML to learn how to infer spatial information from two dimensional images, and the process that generates the data it trains on cannot do this at all. It can create two dimensional images from spatial information, which is a much simpler and effectively solved process.
There are cases when we want a machine learning model to do the same thing as the process which generate it's data, like in the case of model's learning to replicate physics simulations, but even then the entire point is for the machine learning model to accomplish the same or a similar result but in a more computationally efficient way.
With (imho closely related) bayesian statistics you can add those rules into your model, the problem is that it quickly becomes computationally intractable to do exactly (in a calculate this N dimensional integral with N>>1000 kind of way).
So in the end you still end up fitting some simplified vaguely similar model/function to your data and/or using sampling to keep things tractable.
I think that's an insightful point. Personally, I believe an important part of ML architecture engineering is figuring out architectures that make giving "the wrong answer" to a problem impossible, or at least difficult. But sometimes just throwing a bunch of data at a generic architecture and hoping it can learn the "rules" is easier.
And of course, as alluded in the article, sometimes we don't really know the "rules", or don't know how to articulate them in a way we can feasibly synthesize data according to those rules. Then people might reach for ML to figure out those rules, implicitly or explicitly. So it becomes a bit of an ouroboros situation. But still somewhere in that loop there's some domain knowledge injected somewhere by whoever is engineering the pipeline.
But even us as humans (or anything with biological intelligence) infers the rules from data.
The only thing is — we start collecting data since birth and we also don’t have to pay for every little thing we put in front of our eyes.
If you’re training an AI, you have to pay for data. If you generate some of your training data, you pay a little less.
(Also I think we ask of AI more than we ask of humans. For example, it’s common to be face blind of ethnicities you’re not familiar with but we get pissed if an AI is face blind at all.)
We ask of AI more than we ask of humans because AI can have a much greater reach than one person at the same task.
If you're just one guy who's face blind, no one really cares and you can't cause that much damage unless you're a border agent or something. Even then, it's not like you're affecting every single person passing through your country's border, just the ones who happened to be in the right (wrong?) place at the right time.
But if you deploy a face-blind AI as a security system that discriminates against minorities in thousands of establishments, suddenly it's like having thousands of clones of the same face-blind human and it can affect millions of people. If we're effectively making that many clones of a human, it better be good at its job
CMU tried to build a self-learning expert system called NELL but it failed. Never-Ending Language Learning system (NELL) is a semantic machine learning system developed by a research team at Carnegie Mellon University, and supported by grants from DARPA, Google, NSF, and CNPq with portions of the system running on a supercomputing cluster provided by Yahoo!
In his 2019 book "Human Compatible", Stuart Russell commented that 'Unfortunately NELL has confidence in only 3 percent of its beliefs and relies on human experts to clean out false or meaningless beliefs on a regular basis—such as its beliefs that “Nepal is a country also known as United States” and "value is an agricultural product that is usually cut into basis."'
I know this is different than what the article is talking about, but I've wondered if something like a GPTZero would be possible. Train a model to reproduce random sequences, and see if it's able to learn any patterns that are also useful for language or other real-world tasks.
If there's nothing to be gained from training on pure noise, what is the most generic and generally useful sequence we could train on?
Is there not an algebraic function approach to neural networks instead of discrete datasets? Instead of multidimensional vectors, have balanced parametric equations or similar.
26 comments
[ 2.7 ms ] story [ 63.8 ms ] threadhttps://www.capitalone.com/tech/software-engineering/why-you...
Part of the reason was models generating synthetic data hone in on more key differentiable features
We're using software that has encoded the rules we want to use to generate data. The ML is using the data to infer the rules. It seems like we need a better way to give our ML algorithms the rules that we already know rather than this processing-intensive process.
But perhaps that's the cost of using neural networks- they can learn any function, but they can only learn from data, examples.
Neural nets work by generalizing. This is important because GOFAI systems are brittle. We are never able to encode all the relevant rules for real world situations by hand.
Also, note that the synthetic data is generated from rules about how the world looks, not rules about what the model should do.
There may still be some direction for zapping the rules-of-the-world right into the circuitry of, say, a subnet of the model, without doing backprop. Something like a one-shot initialization of the network to approximate the given distribution.
The combinatory bit due to its branching nature lends itself poorly to leveraging SIMD hardware as statistical solutions do. And getting a right representation model is a substantial challenge. Everyone knew the importance of that ('knowledge engineering') by the onset of AI Winter but nobody had any good methods.
There are cases when we want a machine learning model to do the same thing as the process which generate it's data, like in the case of model's learning to replicate physics simulations, but even then the entire point is for the machine learning model to accomplish the same or a similar result but in a more computationally efficient way.
So in the end you still end up fitting some simplified vaguely similar model/function to your data and/or using sampling to keep things tractable.
And of course, as alluded in the article, sometimes we don't really know the "rules", or don't know how to articulate them in a way we can feasibly synthesize data according to those rules. Then people might reach for ML to figure out those rules, implicitly or explicitly. So it becomes a bit of an ouroboros situation. But still somewhere in that loop there's some domain knowledge injected somewhere by whoever is engineering the pipeline.
The only thing is — we start collecting data since birth and we also don’t have to pay for every little thing we put in front of our eyes.
If you’re training an AI, you have to pay for data. If you generate some of your training data, you pay a little less.
(Also I think we ask of AI more than we ask of humans. For example, it’s common to be face blind of ethnicities you’re not familiar with but we get pissed if an AI is face blind at all.)
No, I don't agree with this. My point is that we can learn from higher level information than raw data.
I can give you the formula of a complex mathematical function over 3 variables, and provide no data points.
You'll be able to tell me with perfect accuracy what values match the formula and which don't.
If you're just one guy who's face blind, no one really cares and you can't cause that much damage unless you're a border agent or something. Even then, it's not like you're affecting every single person passing through your country's border, just the ones who happened to be in the right (wrong?) place at the right time.
But if you deploy a face-blind AI as a security system that discriminates against minorities in thousands of establishments, suddenly it's like having thousands of clones of the same face-blind human and it can affect millions of people. If we're effectively making that many clones of a human, it better be good at its job
In his 2019 book "Human Compatible", Stuart Russell commented that 'Unfortunately NELL has confidence in only 3 percent of its beliefs and relies on human experts to clean out false or meaningless beliefs on a regular basis—such as its beliefs that “Nepal is a country also known as United States” and "value is an agricultural product that is usually cut into basis."'
If there's nothing to be gained from training on pure noise, what is the most generic and generally useful sequence we could train on?