Pith. sign in

REVIEW 8 cited by

Understanding the Difficulty of Training Transformers

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

arxiv 2004.08249 v3 pith:PVRUE43M submitted 2020-04-17 cs.LG cs.CLstat.ML

classification cs.LGcs.CLstat.ML
keywords trainingmodeltextbftransformersadminanalysisdependencyleads
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
abstract

Transformers have proved effective in many NLP tasks. However, their training requires non-trivial efforts regarding designing cutting-edge optimizers and learning rate schedulers carefully (e.g., conventional SGD fails to train Transformers effectively). Our objective here is to understand $\textit{what complicates Transformer training}$ from both empirical and theoretical perspectives. Our analysis reveals that unbalanced gradients are not the root cause of the instability of training. Instead, we identify an amplification effect that influences training substantially -- for each layer in a multi-layer Transformer model, heavy dependency on its residual branch makes training unstable, since it amplifies small parameter perturbations (e.g., parameter updates) and results in significant disturbances in the model output. Yet we observe that a light dependency limits the model potential and leads to inferior trained models. Inspired by our analysis, we propose Admin ($\textbf{Ad}$aptive $\textbf{m}$odel $\textbf{in}$itialization) to stabilize stabilize the early stage's training and unleash its full potential in the late stage. Extensive experiments show that Admin is more stable, converges faster, and leads to better performance. Implementations are released at: https://github.com/LiyuanLucasLiu/Transforemr-Clinic.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 8 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. One task to rule them all: A closer look at traffic classification generalizability

    cs.NI 2025-07 conditional novelty 7.0 of 10

    Traffic classifiers that seem near-perfect on their own datasets fall to 30-40% accuracy on another network's same-task data, and a 1-Nearest Neighbor baseline is competitive.

  2. Scalable Complexity Control Facilitates Reasoning Ability of LLMs

    cs.LG 2025-05 conditional novelty 6.0 of 10

    Controlling model complexity through smaller initialization rates and stronger weight decay improved LLM benchmark scores and made loss-versus-scale curves descend faster.

  3. One Rank at a Time: Cascading Error Dynamics in Sequential Learning

    cs.LG 2025-05 conditional novelty 6.0 of 10

    Errors from each rank-1 step in sequential low-rank learning compound through factors that grow when singular values are close, so early steps deserve more compute.

  4. Avoiding spurious sharpness minimization broadens applicability of SAM

    cs.LG 2025-02 conditional novelty 6.0 of 10

    SAM's failure in language modeling is traced to a dominant 'logit path' that minimizes sharpness spuriously, and the proposed Functional-SAM, which removes that path, improves validation loss over AdamW and SAM.

  5. On the Performance of Differentially Private Optimization with Heavy-Tail Class Imbalance

    cs.LG 2025-07 conditional novelty 5.0 of 10

    Under heavy-tail class imbalance, subtracting the DP noise variance from Adam's second moment (DP-AdamBC) substantially improves learning of rare classes compared with DP gradient descent.

  6. SoftSignSGD(S3): An Enhanced Optimizer for Practical DNN Training and Loss Spikes Minimization Beyond Adam

    cs.LG 2025-07 reject novelty 5.0 of 10

    S3, an optimizer with a p-th order momentum denominator, equal EMA coefficients, and Nesterov acceleration, is claimed to match AdamW's 100k-step perplexity at 50k steps while avoiding loss spikes.

  7. An Analysis for Reasoning Bias of Language Models with Small Initialization

    cs.CL 2025-02 conditional novelty 5.0 of 10

    Initialization scale controls whether a transformer learns compositional reasoning or memorized mappings, because reasoning tokens acquire more differentiated embeddings early in training.

  8. DBF-Net: A Dual-Branch Network with Feature Fusion for Ultrasound Image Segmentation

    eess.IV 2024-11 conditional novelty 3.0 of 10

    DBF-Net, a dual-branch network with body and boundary supervision plus feature fusion, reports Dice scores of 81.05%, 76.41%, and 87.75% on BUSI, UNS, and UHES ultrasound datasets.

Pith tools