Hacker News new | past | comments | ask | show | jobs | submit login

I was like this a week ago. Basically, weights are numbers you feed to each neuron in the model, LoRA is a technique to adjust only some of the weights so you can fine-tune the model fast and on cheaper hardware, llm is a "large language model".

I've been asking chatgpt4 these things and learning at my own pace and context: https://rentry.co/vafkn




What's fun is that the recent references to LoRA stands for Low-Rank Adaptation [1], not Low-Resource Adapters[2] (??? don't know if this even exists), but because Low-Rank Adaptation came out in 2021, chatgpt only explains what Low-Resource Adapters is, not Low-Rank Adaptation, which is what is being used in recent break throughs.

My own questioning of chatgpt on LoRA returns "Logistic Regression Algorithm" instead. Looks like it's too new to ChatGPT.

[1] https://arxiv.org/abs/2106.09685 [2] https://rentry.co/vafkn


Well, it’s more that the weights are the neurons. There’s not actually like neuron objects defined or anything, neural networks are just a bunch of matrix operations. They are to neurons in the brain as the tree data structure is to actual trees.


The weights aren't the neurons, they're the connection strengths between the neurons.


You mean the same neurons that don't really exist, other than as a connection from weight to weight, so quite reasonable to see as the weights themselves, because what else is available for us to reference as neurons in absentia those connections?


I disagree. The metaphor we use in ML is that neurons are the nodes that receive the summed, weighted inputs from other neurons. Yes, the weights are the strengths of the connections between them. However, there are many more weights than neurons, so conflating the two doesn't make sense schematically. Also, neurons can also have other parameters which define their behavior such as bias and activation functions. Furthermore, the activation of a neuron defines the network's response to a stimuli, so these change depending on the input, whereas the weighs are constants (after being trained), that parameterize the system.

The analogy is that weights are synapses, not neurons. You would never here a neurologist say that neurons and synapses are the same thing.


I would agree with most of this, but there is no direct analogy between all of the components of a neuron in neurology, which are typically umbrella'd under the name "neuron" and regarded as parts of one, versus ML's version.

Eg, if a weight can be a synapse, can't a weight be an axon? Axons also "connect" neurons, but their length is more related to the connection strength, so could be considered more analogous to a "weighting".

Yet, axons are not as obtusely "one-to-many" as synapses, but depending on the structure of the ML model, and the view of which aspect of it is more impactful to be highlighting by analogy, either take might be more appropriate.

I suppose it depends on the kind of structure you're working with, and whether you're training and inferring, or just one or the other. In all cases I think a good argument could be made for general neuron analogy abuse.


Oh that's interesting. I don't know too much about the neuroscience, just enough to agree that a real neuron is vastly more complex than a node in a "neural net". Based on your description, an axon is most highly analogous to the bias term, although it would be a multiplicative bias. I wonder if that's been tried.


> You mean the same neurons that don’t really exist,

“Neurons” are an abstraction that exists logically “between” the weights, but the weights themselves don’t have the features of neurons. (In that each weight is the weight of a connection between a neurons (or between a neuron and an input/output.) Weights are more synapses than neurons.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: