Summary of Unified Neural Network Scaling Laws and Scale-time Equivalence, by Akhilan Boopathy et al.
Unified Neural Network Scaling Laws and Scale-time Equivalence
by Akhilan Boopathy, Ila Fiete
First submitted to arxiv on: 9 Sep 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Machine Learning (stat.ML)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A novel theoretical characterization is presented to understand how three factors – model size, training time, and data volume – interact to determine the performance of deep neural networks. A scale-time equivalence is established, challenging current practice where large models are trained for small durations. This equivalence leads to a method for predicting the performance of large-scale networks from small-scale networks trained for extended epochs. A unified theoretical scaling law is obtained by combining scale-time equivalence with linear model analysis of double descent, which is confirmed through experiments across vision benchmarks and network architectures. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Deep neural networks are getting bigger, but we still don’t know how much they can really do. We’re trying to figure out if making the network bigger or having more training data makes a difference. Some experts have found that bigger models need more training time, while others think more data is what you need. Our research shows that these two things are actually connected – we call it “scale-time equivalence”. This means that big models can be trained for shorter times, and small models for longer times, to get the same results. We also found out why bigger models sometimes don’t work better just because they’re bigger. Our discoveries could make it easier and more efficient to train these powerful networks. |
