REVIEW 6 cited by
Tensor Programs VI: Feature Learning in Infinite-Depth Neural Networks
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
abstract
By classifying infinite-width neural networks and identifying the *optimal* limit, Tensor Programs IV and V demonstrated a universal way, called $\mu$P, for *widthwise hyperparameter transfer*, i.e., predicting optimal hyperparameters of wide neural networks from narrow ones. Here we investigate the analogous classification for *depthwise parametrizations* of deep residual networks (resnets). We classify depthwise parametrizations of block multiplier and learning rate by their infinite-width-then-depth limits. In resnets where each block has only one layer, we identify a unique optimal parametrization, called Depth-$\mu$P that extends $\mu$P and show empirically it admits depthwise hyperparameter transfer. We identify *feature diversity* as a crucial factor in deep networks, and Depth-$\mu$P can be characterized as maximizing both feature learning and feature diversity. Exploiting this, we find that absolute value, among all homogeneous nonlinearities, maximizes feature diversity and indeed empirically leads to significantly better performance. However, if each block is deeper (such as modern transformers), then we find fundamental limitations in all possible infinite-depth limits of such parametrizations, which we illustrate both theoretically and empirically on simple networks as well as Megatron transformer trained on Common Crawl.
Forward citations
Cited by 6 Pith papers
-
DeepLoop: Depth Scaling for Looped Transformers
Looped Transformers need residual-scaling exponent p=1/2 instead of DeepNorm's 1/4 when shared blocks are revisited with aligned visit-wise gradients.
-
Improving Neural Network Training by Decoupling the Magnitude and Direction of Weight Vectors
Splitting weight matrices into a fixed-norm direction and learnable per-row/column magnitudes improves LLM training over AdamW/Muon, removes weight decay and warmup, and transfers the optimal LR across width.
-
Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance
Falcon-H1 reports competitive benchmark scores for a 0.5B to 34B family of parallel hybrid attention/Mamba-2 models, claiming 2x to 4x parameter efficiency versus dense transformers.
-
Scale Weight Decay and Train Better
Muon with weight decay scaled by η/η_max reaches the same MoE validation loss ~30% faster than constant-decay Muon while preserving asymptotic stationarity of the unregularized objective.
-
SingLoRA: Low Rank Adaptation Using a Single Matrix
SingLoRA replaces LoRA's two matrices A and B with one matrix A and the symmetric update AA^T, cutting adapter parameters roughly in half while claiming more stable fine-tuning.
-
Sub-Scaling Laws: On the Role of Data Density and Training Strategies in LLMs
High data redundancy and over-training decelerate LLM performance gains, and the authors fit a sub-optimal scaling law with logistic correction terms to predict the slowdown.
Discussion (0). Sign in to comment.