We know the input, we can set the model to save the weight in checkpoints during training and can view them any time, and we can see weights of the finished model, and we can see the code.
If what you said about LLMs being completely black box were true, we wouldn’t be able to reproduce models, and each model would be unique.
But we can control every step of the training process, and we can reproduce not just the finished model, but the model at every single step during training.
We created the math, we created the training sets, we created the code and we can see and modify the weights and any other property of the model.
Look, I understand why you think this. I thought this too when I was first beginning to learn machine learning and data science. But I’ve now been working with machine learning models including neural networks for nearly a decade, and the truth is that is nearly impossible to track the path of an input to a given output in machine learning models other than regression-based models and decision tree-based models.
There is an entire field of data science devoted to explaining how these models arrive at their conclusions. It’s called “explainable AI” or “xAI”, and I have a few papers that I’ve published in exploring the utility of them. The basic explanation for how they work is that we run hundreds of thousands of different models and then do statistical analysis to estimate why the models arrived at their conclusion. It isn’t an exact science, however.
Again, we have the input, we have the math and code that make it work, we have the weights, we have everything.
Would it take a lot of time to backtrack and check why we got a given output to an input? Yes, maybe an inordinate amount of time. But it can be done. It’s only black box because nobody has the time (likely years to decades) to wade through the layers of a finished model to check every node and weight.
The whole thing at its core is mathematics. It’s a series of steps, that can be listed and reviewed each step of the way if we wanted. It’s just that if would take too much time.
If what you said were true, we couldn’t reproduce models. And since we can…
It isn’t an exact science, however.
So if math and computer science isn’t an exact science, what is?
It’s only black box because nobody has the time (likely years to decades) to wade through the layers of a finished model to check every node and weight.
This is exactly correct, except you’re also not accounting for the insane amount of computational power that would be necessary to backtrack a single output of a single model. This is why it is a black box. It simply is not possible on a meaningful level.
So if math and computer science isn’t an exact science, what is?
Things that are reproducible with known inputs and outputs, allowing for all components to be studied and explained. As an example from my field: if you damage the dorsolateral prefrontal cortex in a fully grown adult, they will have the impulse control of a three-year old. We know this because we have observed damage to this area in multiple individuals, and can measure the effects based on the severity of that damage.
In contrast, if you provide the same billion-parameter neural network identical inputs, you will not receive identical outputs.
This is exactly correct, except you’re also not accounting for the insane amount of computational power that would be necessary to backtrack a single output of a single model. This is why it is a black box. It simply is not possible on a meaningful level.
It’s not practical, but it can be done. We simple don’t have the time or inclination to do it.
It’s like like saying we don’t understand how an internal combustion engine works. Every explosion is a bit different, it pushes the pistons a bit less or more, it leaves a bit more or less residue in different places. We can’t backtrack and check every cycle and every part on a meaningful level, but we understand of it works, and we could do it if we wanted to. It’s just not practical.
As an example from my field: if you damage the dorsolateral prefrontal cortex in a fully grown adult, they will have the impulse control of a three-year old. We know this because we have observed damage to this area in multiple individuals, and can measure the effects based on the severity of that damage.
Okay so explain how sudden savant syndrome works. Step by step, biochemical process by biochemical process.
In contrast, if you provide the same billion-parameter neural network identical inputs, you will not receive identical outputs.
If you take the same model, put it in a VM, give it an input, get an output and the restore the VM to the exact same state before and ensure there’s no randomness, the model will give you same output.
That’s not really the case for all ML systems, it the fact that programers can generate content that they themselves did not make, collect, or anticipate themselves by creating models that generate thier own decision trees based on input data.
We know the input, we can set the model to save the weight in checkpoints during training and can view them any time, and we can see weights of the finished model, and we can see the code.
If what you said about LLMs being completely black box were true, we wouldn’t be able to reproduce models, and each model would be unique.
But we can control every step of the training process, and we can reproduce not just the finished model, but the model at every single step during training.
We created the math, we created the training sets, we created the code and we can see and modify the weights and any other property of the model.
What exactly do we not understand?
Look, I understand why you think this. I thought this too when I was first beginning to learn machine learning and data science. But I’ve now been working with machine learning models including neural networks for nearly a decade, and the truth is that is nearly impossible to track the path of an input to a given output in machine learning models other than regression-based models and decision tree-based models.
There is an entire field of data science devoted to explaining how these models arrive at their conclusions. It’s called “explainable AI” or “xAI”, and I have a few papers that I’ve published in exploring the utility of them. The basic explanation for how they work is that we run hundreds of thousands of different models and then do statistical analysis to estimate why the models arrived at their conclusion. It isn’t an exact science, however.
Again, we have the input, we have the math and code that make it work, we have the weights, we have everything.
Would it take a lot of time to backtrack and check why we got a given output to an input? Yes, maybe an inordinate amount of time. But it can be done. It’s only black box because nobody has the time (likely years to decades) to wade through the layers of a finished model to check every node and weight.
The whole thing at its core is mathematics. It’s a series of steps, that can be listed and reviewed each step of the way if we wanted. It’s just that if would take too much time.
If what you said were true, we couldn’t reproduce models. And since we can…
So if math and computer science isn’t an exact science, what is?
This is exactly correct, except you’re also not accounting for the insane amount of computational power that would be necessary to backtrack a single output of a single model. This is why it is a black box. It simply is not possible on a meaningful level.
Things that are reproducible with known inputs and outputs, allowing for all components to be studied and explained. As an example from my field: if you damage the dorsolateral prefrontal cortex in a fully grown adult, they will have the impulse control of a three-year old. We know this because we have observed damage to this area in multiple individuals, and can measure the effects based on the severity of that damage.
In contrast, if you provide the same billion-parameter neural network identical inputs, you will not receive identical outputs.
It’s not practical, but it can be done. We simple don’t have the time or inclination to do it.
It’s like like saying we don’t understand how an internal combustion engine works. Every explosion is a bit different, it pushes the pistons a bit less or more, it leaves a bit more or less residue in different places. We can’t backtrack and check every cycle and every part on a meaningful level, but we understand of it works, and we could do it if we wanted to. It’s just not practical.
Okay so explain how sudden savant syndrome works. Step by step, biochemical process by biochemical process.
If you take the same model, put it in a VM, give it an input, get an output and the restore the VM to the exact same state before and ensure there’s no randomness, the model will give you same output.
|it’s not practical, but it can be done. We simple don’t have the time or inclination to do it
Is also supposed to be true if the human mind.
Not really, because there’s still some processes in the human brain we don’t understand.
For example, you can list the steps and processes for every step an AI makes. You have to, in order to code and run it.
But you can’t list every step or process taking place in cases of sudden savant syndrome, for example.
That’s not really the case for all ML systems, it the fact that programers can generate content that they themselves did not make, collect, or anticipate themselves by creating models that generate thier own decision trees based on input data.