The Quantization Model of Neural Scaling

Authors: Eric Michaud, Ziming Liu, Uzay Girit, Max Tegmark

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate this prediction on toy datasets, then study how scaling curves decompose for large language models. Using language model gradients, we automatically decompose model behavior into a diverse set of skills (quanta).
Researcher Affiliation Academia Eric J. Michaud , Ziming Liu, Uzay Girit, and Max Tegmark MIT & IAIFI
Pseudocode No The paper describes a method named 'Quanta Discovery from Gradients (QDG)' with sequential steps in paragraph format, but it does not present it as a formally structured pseudocode or algorithm block.
Open Source Code Yes Project code can be found at: https://github.com/ejmichaud/quantization-model.
Open Datasets Yes For our experiments, we use the Pythia model suite from Eleuther AI [29], a set of decoder-only transformers of varying size trained on approximately 300 billion tokens of The Pile [30].
Dataset Splits No The paper mentions training and testing on datasets but does not provide specific percentages or counts for training/validation/test splits, nor does it explicitly detail predefined splits with citations for reproducibility.
Hardware Specification Yes Availble GPUs include NVIDIA A100, RTXA6000, QUADRORTX6000, GEFORCERTX2080TI, GEFORCERTX2080, GEFORCEGTX1080TI, titan-x, and tesla-v100.
Software Dependencies No The paper mentions using 'scikit-learn [45]' but does not provide specific version numbers for this or any other software dependencies crucial for replication.
Experiment Setup Yes We use the Adam optimizer with a learning rate of 10-3. To study scaling with respect to the number of model parameters, we train networks of varying width by sampling batches online... For the results shown, we used ntasks = 500, n = 100, k = 3, α = 0.4, and a batch size of 20000. We vary training dataset size from 1e4 to 5e6 and vary hidden-layer width from 10 to 500 neurons. We train for 2e5 steps.