2
Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
submitted 2026年10月22日 by @nineDan 203
122 則主張
+2
Forked from Attention is All you Need
Overview
Attention is All you Need：1 項公開調查 與 342 項主張。
地圖總覽
- 1Attention is All you Need
submitted 2026年10月22日 by @nineDan 203
101 則主張+2OverviewThis map provides a foundational overview of the 2017 paper, tracing its authorial lineage across Google Brain and Research, the release of the open-source software ecosystem, and its core thesis against sequential recurrence. It also tracks the paper's multi-year bibliographic revisions and metric inconsistencies.
- 3Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
submitted 2026年10月22日 by @nineDan 203
119 則主張+2Forked from Attention is All you Need
OverviewThis map investigates whether the Transformer's ascent was driven by hardware-level TPU parallelizability and economics rather than algorithmic superiority. It balances technical trade-offs between sample efficiency and compute throughput against the historical record of project leadership and implementation details.
Last generated 2026年10月22日 下午5:10 by a member
更多資訊
已於 2026年10月22日 下午5:10 進行比較
圖譜
正在比較地圖…
Themes across the family
A broad read of every map in this family: the themes that run through them and how each map treats them. Structured scans give a more targeted side-by-side.
Attribution and Architectural Provenance
3 個圖譜中的 3 個The family documents the inception of the Transformer as a collaborative synthesis where Jakob Uszkoreit provided the initial catalytic hypothesis of replacing recurrence with self-attention, complemented by specialized architectural contributions across Google Brain and Research.
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Identifies Uszkoreit as the conceptual initiator and catalyst while emphasizing that design and mechanisms were distributed among coauthors such as Vaswani, Polosukhin, and Shazeer.
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
Acknowledges Uszkoreit's foundational role while highlighting Shazeer's invention of scaled dot-product attention.
Traces the close cross-building collaboration between Google Brain and Google Research teams that produced the early model implementations.
Architectural Components Beyond Attention
3 個圖譜中的 3 個All maps qualify the rhetorical claim that attention is all that is needed, showing that feed-forward sublayers, learning rate schedules, and optimizer engineering were critical to model viability.
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Stresses that the finished Transformer was not composed solely of attention and relied on load-bearing optimizer schedules.
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
Contends that feed-forward sublayers play a crucial role as associative key-value memory storage rather than being incidental to attention.
Points out that independent replication required extracting subtle learning-rate schedules and hyperparameters not described by pure attention.
Hardware Parallelization and Recurrence Bottlenecks
3 個圖譜中的 2 個The research maps examine whether overcoming the sequential computational constraints of recurrent networks was fundamentally a hardware-throughput trade-off.
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Probes whether the architecture sacrificed algorithmic sample efficiency specifically to capitalize on parallel throughput across TPU clusters.
Focuses on the paper's core empirical proposition that sequential recurrence prevented intra-example training parallelization.
Publication Inconsistencies and Reporting Revisions
3 個圖譜中的 3 個The family details persistent reporting discrepancies in benchmark translation metrics across successive arXiv versions and conference proceedings.
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Documents multiple conflicting English–French BLEU scores and missing implementation details in the original paper text.
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
Notes the evolution of the paper across seven arXiv versions and the delayed introduction of headline BLEU scores.
Highlights direct internal contradictions between abstract and body BLEU figures within the official NeurIPS proceedings.
Jakob Uszkoreit acted as the originator and catalyst by proposing the replacement of recurrence with self-attention, but he was not the sole inventor of the full Transformer system.
3 個圖譜中的 3 個M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
The best-supported interpretation of Uszkoreit's role is originator and initial project catalyst, not sole architect or inventor.
有爭議高 重要性ChatGPT
The Transformer was not composed solely of attention mechanisms, as non-attention components and optimizer schedules were critical to the system's empirical success.
3 個圖譜中的 3 個M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0080Transformer not all attention
The finished Transformer model did not literally replace every RNN operation with attention and was not composed solely of attention.
有爭議高 重要性ChatGPT
The public record of the paper contains multiple mutually inconsistent English–French translation BLEU figures.
3 個圖譜中的 3 個M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0052Inconsistent BLEU figures
The public record of the "Attention is All you Need" paper contains at least four mutually inconsistent English–French BLEU figures.
具來源連結高 重要性ChatGPT
Reimplementation of the Transformer demonstrated that the learning-rate schedule was an unusually consequential and sensitive implementation detail.
3 個圖譜中的 3 個M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0091Reviewer reimplemented system
An independent NeurIPS reviewer attempted to reimplement the system from "Attention is All you Need" and found the learning-rate schedule to be unusually consequential.
具來源連結高 重要性ChatGPT
The maps diverge on whether the Transformer's decisive breakthrough was trading sample efficiency for TPU cluster parallelizability or establishing feed-forward sublayers as associative factual storage.
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0022Hardware Determinism over Algorithmic Elegance使用 AI 調查
The Transformer architecture traded algorithmic sample efficiency for hardware-level training parallelizability on TPU clusters.
推論高 重要性模型來源無法使用
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0022The 'Attention' Misnomer and MLP Dominance使用 AI 調查
Feed-forward sublayers function as associative key-value memories that store the substantive knowledge in Transformer models.
推論高 重要性模型來源無法使用
Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise? uniquely highlights the extensive GPU compute time and memory limitations encountered during independent empirical exploration.
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0027Transformer paper compute requirements
The empirical exploration experiments for the Popel and Bojar (2018) Transformer training tips paper required roughly 4 years of cumulative GPU compute time.
具來源連結低 重要性Claude
Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage? uniquely connects feed-forward sublayers to classical associative memory models formulated in the 1970s.
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0039Linear associative memory foundation
ROME's key-value memory formulation relies on classical linear associative memory frameworks originally established by Anderson (1972) and Kohonen (1972).
推論低 重要性Claude
Attention is All you Need uniquely highlights the contemporaneous June 2017 open-sourcing of Google's Tensor2tensor framework.
C0043Tensor2tensor open-sourced June 2017
Tensor2tensor was open-sourced by Google, announced in June 2017, the same month as the "Attention Is All You Need" paper.
推論低 重要性Kimi
The reported English–French translation performance conflicts across publications, claiming 41.8 BLEU in the current arXiv version but 41.0 BLEU in Table 2 of the NeurIPS proceedings PDF.
The current arXiv abstract and Table 2 for "Attention is All you Need" report an English–French BLEU score of 41.8.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Table 2 of the NeurIPS proceedings PDF for "Attention is All you Need" reports an English–French BLEU score of 41.0.
具來源連結高 重要性ChatGPT
The NeurIPS proceedings text conflicts internally by reporting an English–German score of 27.5 BLEU in the abstract while reporting 28.4 BLEU in the body text.
C0094Model size and BLEU score
The NeurIPS proceedings abstract for "Attention is All you Need" describes a 165-million-parameter model scoring 27.5 English–German BLEU.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
The proceedings PDF body of "Attention is All you Need" uses 28.4 English–German BLEU, creating an abstract/body discrepancy.
具來源連結高 重要性ChatGPT
- https://arxiv.org/abs/1706.03762
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
The current arXiv abstract and Table 2 for "Attention is All you Need" report an English–French BLEU score of 41.8.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0052Inconsistent BLEU figures
The public record of the "Attention is All you Need" paper contains at least four mutually inconsistent English–French BLEU figures.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
The arXiv record for "Attention is All you Need" includes versions v6 on July 24, 2023, and v7 on August 2, 2023.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
The proceedings PDF body of "Attention is All you Need" uses 28.4 English–German BLEU, creating an abstract/body discrepancy.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0075Model size and BLEU score
The NeurIPS proceedings abstract for "Attention is All you Need" describes a 165-million-parameter model scoring 27.5 English–German BLEU.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0089Vaswani submitted paper v1
Ashish Vaswani was identified as the submitter for v1 of "Attention is All you Need" on arXiv, while the current record identifies Llion Jones.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Table 2 of the NeurIPS proceedings PDF for "Attention is All you Need" reports an English–French BLEU score of 41.0.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0105Attention paper version 7
The latest official version of "Attention is All you Need" on arXiv is v7, not v5.
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0055Inconsistent BLEU figures
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0078Model size and BLEU score
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0092Vaswani submitted paper v1
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
- https://papers.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Reviews.html
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0052Inconsistent BLEU figures
The public record of the "Attention is All you Need" paper contains at least four mutually inconsistent English–French BLEU figures.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
The proceedings PDF body of "Attention is All you Need" uses 28.4 English–German BLEU, creating an abstract/body discrepancy.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0075Model size and BLEU score
The NeurIPS proceedings abstract for "Attention is All you Need" describes a 165-million-parameter model scoring 27.5 English–German BLEU.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
The NeurIPS proceedings abstract for "Attention is All you Need" reports an English–French BLEU score of 41.1.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0091Reviewer reimplemented system
An independent NeurIPS reviewer attempted to reimplement the system from "Attention is All you Need" and found the learning-rate schedule to be unusually consequential.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Table 2 of the NeurIPS proceedings PDF for "Attention is All you Need" reports an English–French BLEU score of 41.0.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
A reviewer of "Attention is All you Need" found that important replication details, such as numerous hyperparameters, had to be extracted from the code rather than the submitted paper.
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0055Inconsistent BLEU figures
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0078Model size and BLEU score
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0094Reviewer reimplemented system
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
- https://arxiv.org/pdf/1706.03762
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0003Kaiser affiliated with Google
Łukasz Kaiser was listed as being affiliated with “Google Brain” in arXiv:1706.03762, with the email lukaszkaiser@google.com.
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Noam Shazeer was listed as being affiliated with “Google Brain” in arXiv:1706.03762, with the email noam@google.com.
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0066Ashish Vaswani affiliation
Ashish Vaswani was listed as being affiliated with “Google Brain” in arXiv:1706.03762, with the email avaswani@google.com.
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0085Polosukhin affiliation clarified
Illia Polosukhin's affiliation in arXiv:1706.03762 was clarified by an explicit footnote stating “‡ Work performed while at Google Research.”
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0110Illia Polosukhin departure
Illia Polosukhin's departure from Google occurred before the June 12, 2017, arXiv v1 posting of the "Attention Is All You Need" paper.
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0027Transformer paper compute requirements
The empirical exploration experiments for the Popel and Bojar (2018) Transformer training tips paper required roughly 4 years of cumulative GPU compute time.
具來源連結低 重要性Claude
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0003Kaiser affiliated with Google
具來源連結中 重要性Grok
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0069Ashish Vaswani affiliation
具來源連結中 重要性Grok
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0088Polosukhin affiliation clarified
具來源連結中 重要性Grok
- https://arxiv.org/html/1706.03762
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0003Kaiser affiliated with Google
Łukasz Kaiser was listed as being affiliated with “Google Brain” in arXiv:1706.03762, with the email lukaszkaiser@google.com.
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
Noam Shazeer was listed as being affiliated with “Google Brain” in arXiv:1706.03762, with the email noam@google.com.
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0066Ashish Vaswani affiliation
Ashish Vaswani was listed as being affiliated with “Google Brain” in arXiv:1706.03762, with the email avaswani@google.com.
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0085Polosukhin affiliation clarified
Illia Polosukhin's affiliation in arXiv:1706.03762 was clarified by an explicit footnote stating “‡ Work performed while at Google Research.”
具來源連結中 重要性Grok
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0110Illia Polosukhin departure
Illia Polosukhin's departure from Google occurred before the June 12, 2017, arXiv v1 posting of the "Attention Is All You Need" paper.
具來源連結中 重要性Grok
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0003Kaiser affiliated with Google
具來源連結中 重要性Grok
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0069Ashish Vaswani affiliation
具來源連結中 重要性Grok
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0088Polosukhin affiliation clarified
具來源連結中 重要性Grok
- https://papers.nips.cc/paper_files/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0052Inconsistent BLEU figures
The public record of the "Attention is All you Need" paper contains at least four mutually inconsistent English–French BLEU figures.
具來源連結高 重要性ChatGPT
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
A reviewer of "Attention is All you Need" found that important replication details, such as numerous hyperparameters, had to be extracted from the code rather than the submitted paper.
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
C0055Inconsistent BLEU figures
具來源連結高 重要性ChatGPT
M2Did the Attention Mechanism Disguise the True Role of Feed-Forward Storage?
具來源連結高 重要性ChatGPT
- https://arxiv.org/abs/1606.01933
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0043Proposal attributed to Uszkoreit
The explicit sentence attributing the proposal to Uszkoreit does not appear in arXiv versions v1–v4 (June 12, 19, 20, 30, 2017) but appears in v5 (December 6, 2017) with detailed contribution statements.
有爭議高 重要性ChatGPT
How should credit attribution for proposing the replacement of RNNs with self-attention be formally interpreted given evolving paper revisions?
M1Did TPU Hardware Economics, Not Algorithmic Superiority, Drive the Transformer's Rise?
C0005Uszkoreit proposed self-attention使用 AI 調查
Jakob Uszkoreit is credited with proposing replacing RNNs with self-attention in the "Attention Is All You Need" paper.
推論低 重要性Claude
Discussion
在評論中輸入 @m 標記地圖，或輸入 @username 提及他人。
還沒有留言。