Weight Space
In interpretability research, the analytical object that directly studies the statistical structure and geometric properties of parameter matrices themselves; it is traditionally regarded as highly uninterpretable because of extensive symmetries and redundancy among parameters.
Timeline (last 90 days)
Taking the example in this article, a small multilayer perceptron with 500 inputs, 5 outputs, and hidden layers can have total weights on the order of hundreds of thousands, whereas a 270M-parameter model has 270 million floating-point numbers, which when flattened and concatenated layer by layer form a high-dimensional weight space.
Weight space is the high-dimensional space formed by all the learnable parameters of a neural network, and the weight space of modern large models can reach dimensions on the order of a hundred billion.
Weight space and activation space are two analysis objects at different levels in interpretability research.
LLMs may simply memorize patterns in the training data rather than truly understanding weight space.
All Facts (4)
Taking the example in this article, a small multilayer perceptron with 500 inputs, 5 outputs, and hidden layers can have total weights on the order of hundreds of thousands, whereas a 270M-parameter model has 270 million floating-point numbers, which when flattened and concatenated layer by layer form a high-dimensional weight space.
50%UnverifiedWeight space is the high-dimensional space formed by all the learnable parameters of a neural network, and the weight space of modern large models can reach dimensions on the order of a hundred billion.
50%UnverifiedWeight space and activation space are two analysis objects at different levels in interpretability research.
50%UnverifiedLLMs may simply memorize patterns in the training data rather than truly understanding weight space.
50%