BZJZ_Material/文档润色流和知识库构建流/claude-scholar/skills/ml-paper-writing/references/knowledge/design-simplification-papers-kaiming-he.md

Design Simplification Papers: Less Is More

Source: Kaiming He et al., "Exploring Plain Vision Transformer Backbones for Object Detection" (ViTDet, 2022)

Paper Type: Design simplification / Minimal adaptations paper

Core Pattern: Challenge design assumptions → Minimize changes → Surprising effectiveness → Fair comparison

---

1. Abstract Structure: The "Surprisingly" Framework

Pattern: Conventional Practice → Simple Alternative → Unexpected Results

Template:

Abstract:
1. [Context] Standard practice in [domain] is [conventional design]
2. [Challenge] With [new technology], this faces [challenges]
3. [Common Solution] Most work addresses this by [abandoning philosophy / adding complexity]
4. [Our Direction] We explore [different direction]: [minimal approach]
5. [Surprisingly 1] Surprisingly, we observe: (i) [simple finding 1]
   and (ii) [simple finding 2]
6. [Surprisingly 2] More surprisingly, [stronger claim under conditions]
7. [Implications] This enables [benefit] without [traditional requirement]

ViTDet Abstract Example (annotated):

Modern object detectors consist of hierarchical backbone feature extractors
and detection-specific necks/heads (e.g., FPN, RPN).

With Vision Transformers (ViT) emerging as powerful backbones, their plain,
non-hierarchical nature poses challenges: How to address multi-scale objects?

One solution abandons the plain ViT philosophy, re-introducing hierarchical
designs (e.g., Swin).

We pursue a different direction: plain ViT backbones with minimal adaptations.

Surprisingly, we observe: (i) A simple feature pyramid from a single-scale map
is sufficient (without FPN), and (ii) Window attention without shifting is
sufficient (with a few propagation blocks).

More surprisingly, under some circumstances, our ViTDet can compete with
leading hierarchical detectors like Swin.

With MAE pre-training, ViTDet outperforms hierarchical counterparts,
especially for larger models.

This decouples pre-training from fine-tuning, maintaining independence of
upstream vs downstream tasks.

Key Techniques:

1. "Modern...consist of..." - Establish conventional practice
2. "With...emerging as..." - New technology, new challenge
3. "abandons the...philosophy" - Critique common solutions
4. "We pursue a different direction" - Clear positioning
5. "Surprisingly, we observe: (i)... and (ii)..." - First surprise
6. "More surprisingly..." - Second, deeper surprise
7. "under some circumstances" - Measured claim
8. "sufficient" - Scientific (not "optimal")
9. "without [common practice]" - Negative emphasis

---

2. Introduction: The "Challenge Assumptions" Framework

Pattern: Tradition → New Challenge → Common Compromise → Your Alternative → Philosophy

Structure:

1. [Traditional Practice] Established design in [field]
2. [Evolution] How this emerged historically ("For a long while...")
3. [New Challenge] [New technology] with [different characteristics]
4. [Philosophy Clash] Original [tech] has "'minimalist' pursuit"
   - Questions: "How can we...?" "Is [X] too inefficient?"
5. [Common Solution] One solution: [abandon philosophy] → [revert to old design]
   - Acknowledge: "has shown successful results"
6. [Your Direction] "we pursue a different direction"
   - Motivation: "If successful, enables [benefit]"
7. [Philosophy] "in part follows the [philosophy] of '[concept]'"
8. [Surprising Findings] "Surprisingly, we observe..."
9. [Implications] "More surprisingly..." → Competitive results

ViTDet Introduction Flow:

Traditional Practice (Establish Context)

Modern object detectors in general consist of a backbone feature extractor
that is agnostic to the detection task and a set of necks and heads that
incorporate detection-specific prior knowledge.

Common components in the necks/heads may include Region-of-Interest (RoI)
operations, Region Proposal Networks (RPN) or anchors, Feature Pyramid
Networks (FPN), etc.

Technique:

• "in general consist of" - Standard architecture
• "agnostic to" vs "detection-specific" - Clear division
• "may include" - Examples, not exhaustive

Historical Evolution

For a long while, these backbones have been multi-scale, hierarchical
architectures due to the de facto design of convolutional networks (ConvNet),
which has heavily influenced the neck/head design for detecting objects at
multiple scales (e.g., FPN).

Technique:

• "For a long while" - Historical dimension
• "due to...which has heavily influenced" - Causal chain
• "de facto design" - Established convention

New Technology Challenge

Over the past year, Vision Transformers (ViT) have been established as a
powerful backbone for visual recognition.

Unlike typical ConvNets, the original ViT is a plain, non-hierarchical
architecture that maintains a single-scale feature map throughout.

Its 'minimalist' pursuit is met with challenges when applied to object
detection—e.g., How can we address multi-scale objects in a downstream task
with a plain backbone from upstream pre-training? Is a plain ViT too
inefficient to use with high-resolution detection images?

Technique:

• "Over the past year...have been established as" - Timeframe
• "Unlike typical ConvNets" - Direct contrast
• "plain, non-hierarchical", "single-scale" - Key characteristics
• "'minimalist' pursuit" - Philosophy (in quotes)
• Two questions: Challenge reader to think

Common Solution (Acknowledge then Pivot)

One solution, which abandons this pursuit, is to re-introduce hierarchical
designs into the backbone.

This solution, e.g., Swin Transformers and related works, can inherit the
ConvNet-based detector design and has shown successful results.

Technique:

• "which abandons this pursuit" - Critique (respectful)
• "can inherit" - Acknowledge advantage
• "has shown successful results" - Don't deny effectiveness

Your Different Direction

In this work, we pursue a different direction: we explore object detectors
that use only plain, non-hierarchical backbones.

If this direction is successful, it will enable the use of original ViT
backbones for object detection; this will decouple the pre-training design
from the fine-tuning demands, maintaining the independence of upstream vs.
downstream tasks, as has been the case for ConvNet-based research.

Technique:

• "we pursue a different direction" - Clear positioning
• "If this direction is successful, it will enable..." - Motivation
• "decouple", "independence" - Philosophy keywords
• "as has been the case for..." - Historical precedent

Philosophy Elevation

This direction also in part follows the ViT philosophy of 'fewer inductive
biases' in the pursuit of universal features.

As the non-local self-attention computation can learn translation-equivariant
features, they may also learn scale-equivariant features from certain forms
of supervised or self-supervised pre-training.

Technique:

• "in part follows the...philosophy of" - Theoretical connection
• "fewer inductive biases" - Core concept
• Analogy: translation-equivariant → scale-equivariant
• "may also learn" - Speculation (honest)

Surprising Findings

Surprisingly, we observe: (i) it is sufficient to build a simple feature
pyramid from a single-scale feature map (without the common FPN design) and
(ii) it is sufficient to use window attention (without shifting) aided with
very few cross-window propagation blocks.

Technique:

• "Surprisingly, we observe:" - Marker
• (i) and (ii) - Structured list
• "sufficient to" - Not "optimal", scientific phrasing
• "without the common [X]" - Negative differentiation

Deeper Surprise

More surprisingly, under some circumstances, our plain-backbone detector,
named ViTDet, can compete with the leading hierarchical-backbone detectors
(e.g., Swin, MViT).

With Masked Autoencoder (MAE) pre-training, our plain-backbone detector can
outperform the hierarchical counterparts that are pre-trained on ImageNet-1K/21K
with supervision (Figure 3).

The gains are more prominent for larger model sizes.

Technique:

• "More surprisingly" - Progressive emphasis
• "under some circumstances" - Measured claim
• "named ViTDet" - Brand at results
• Specific comparison: MAE vs ImageNet supervised
• "The gains are more prominent for..." - Pattern observation

---

3. Methods Section: The "Minimal Adaptations" Narrative

Pattern: We Don't Aim to Invent, We Minimize

Structure:

1. [Declaration] We do NOT aim to develop new components
2. [Philosophy] Instead, make minimal adaptations sufficient to overcome challenges
3. [Specific] In particular, [what we actually do]
4. [Abandonment] This abandons [traditional component]
5. [Decoupling] Adaptations only during fine-tuning, do not alter pre-training
6. [Contrast] This is in contrast to [recent methods] that [what they do]
7. [Benefit] Our scenario enables [benefit], without [cost]

ViTDet Methods Narrative:

In our study, we do not aim to develop new components; instead, we make
minimal adaptations that are sufficient to overcome the aforementioned
challenges.

In particular, our detector builds a simple feature pyramid from only the
last feature map of a plain ViT backbone (Figure 1).

This abandons the FPN design and waives the requirement of a hierarchical
backbone.

These adaptations are made only during fine-tuning and do not alter pre-training.

This is in contrast to the recent methods that modify the attention computation
directly with backbone pre-training (e.g., Swin, MViT).

Our scenario enables us to use the original ViT backbone for detection, without
redesigning pre-training architectures.

Techniques:

• "we do not aim to develop new components" - Clear scope
• "minimal adaptations" - Philosophy
• "sufficient to" - Not maximal, necessary
• "This abandons..." - What you give up
• "only during fine-tuning" - Temporal boundary
• "do not alter pre-training" - Upstream independence
• "This is in contrast to" - Competitor positioning
• "enables us to use" - Practical benefit

---

4. Fair Comparison: The "Equal Effort" Declaration

Pattern: Admit Complexity → Claim Effort → Demonstrate Fairness

Structure:

1. [Admit] Modern systems involve [complexity]
2. [Claim] To compare as fairly as possible, we [effort]
3. [Specific 1] Use same [implementation] for all
4. [Specific 2] Different backbones get [appropriate treatment]
5. [Validation] Our results for [competitor] are [better/equal] to original
6. [Implication] Since we reproduce others well, comparisons are fair

ViTDet Fair Comparison Statement:

Modern detection systems involve many implementation details and subtleties.

To focus on comparing backbones under as fair conditions as possible, we
incorporate the Swin and MViTv2 backbones into our implementation.

We use the same implementation of Mask R-CNN and Cascade Mask R-CNN for all
ViT, Swin, and MViTv2 backbones.

We use FPN for the hierarchical backbones of Swin/MViTv2.

We search for optimal hyper-parameters separately for each backbone.

Techniques:

• "involve many implementation details and subtleties" - Admit difficulty
• "under as fair conditions as possible" - Effort disclaimer
• "incorporate...into our implementation" - What we did
• "use the same...for all" - Unified framework
• "search for optimal...separately" - Equal effort

Self-Validation

Our Swin results are better than their counterparts in the original paper;
our MViTv2 results are better than or on par with those reported in the
original paper.

Technique:

• Report self-results → Show competence → Imply fairness

---

5. Results: Multi-Factor Analysis

Pattern: Factors, Trends, Wall-Clock Time

Structure:

1. [Acknowledge Complexity] Comparisons involve [factors]
2. [Identify Trend] Our method presents better [trend behavior]
3. [Qualify] When [condition], our method [advantage]
4. [Expand] Moreover, [second dimension advantage]
5. [Explain] as [reason related to simplicity]

ViTDet Results Narrative:

Figure 3 plots the trade-offs.

The comparisons here involve two factors: the backbone and the pre-training
strategy.

Our plain-backbone detector, combined with MAE pre-training, presents better
scaling behavior.

When the models are large, our method outperforms the hierarchical
counterparts of Swin/MViTv2, including those using IN-21K supervised
pre-training.

Moreover, the plain ViT has a better wall-clock performance (Figure 3 right),
as the simpler blocks are more hardware-friendly.

Techniques:

• "involve two factors" - Analysis depth
• "presents better scaling behavior" - Trend, not just points
• "When the models are large" - Qualify claim
• "Moreover" - Second dimension
• "better wall-clock performance" - Practical metric
• "simpler blocks are more hardware-friendly" - Explain why

---

6. "Surprisingly" Usage: Multi-Level Pattern

Level 1: Basic Surprise (Abstract)

Surprisingly, we observe: (i) [simple sufficient without common practice]
and (ii) [simple sufficient without common practice]

Characteristics:

• Two findings (i) and (ii)
• "sufficient" not "optimal"
• "without [common practice]"
• Structured presentation

Level 2: Competitive Surprise (Introduction)

More surprisingly, under some circumstances, our [method] can compete
with the leading [competitors].

Characteristics:

• "More surprisingly" - Progressive
• "under some circumstances" - Measured
• "compete with" - Not "beat", competitive
• Name competitors specifically

Level 3: Superiority Surprise (Introduction)

With [specific condition], our [method] can outperform the [competitors]
that use [stronger condition].

The gains are more prominent for [specific condition].

Characteristics:

• Specific conditions compared
• "outperform" - Stronger claim here
• Pattern observation: "more prominent for"
• Shows understanding of when/where

---

7. Ablation Study: Incremental + Destructive

Pattern: Baseline → Incremental Additions → Sufficient

Table Design:

Table X: [Component] Ablation
┌──────────────────────────────────────────┐
│ Baseline          | Metric  | Δ         │
├──────────────────────────────────────────┤
│ no [component]    | 47.8    | -         │
│ (a) [common]      | 50.3    | +2.5      │
│ (b) [variant]     | 50.9    | +3.1      │
│ (c) ours: simple  | 51.2    | +3.4 ✓   │
├──────────────────────────────────────────┤
│ Conclusion: Our simple [X] is sufficient  │
└──────────────────────────────────────────┘

ViTDet Table 1 Example:

pyramid design              APbox   APmask
─────────────────────────────────────────
no feature pyramid          47.8    42.5
(a) FPN, 4-stage           50.3    44.9
(b) FPN, last-map          50.9    45.3
(c) simple feature pyramid  51.2    45.5

Techniques:

• Baseline: "no [X]" shows it's needed
• (a), (b), (c): Progressive variations
• Δ标注: (+2.5) - Show incremental gains
• Conclusion text: "our simple pyramid is sufficient"

---

8. "Interestingly" Usage: Pattern + Explanation

Pattern: Observation → Literature Support → Explanation

Structure:

Interestingly, [observation].

This is in line with the observation in [paper] that [their finding].

[Additional explanation or hypothesis].

ViTDet Example:

Interestingly, performing propagation in the last 4 blocks is nearly as
good as even placement.

This is in line with the observation in ViT [14] that ViT has longer
attention distance in later blocks and is more localized in earlier ones.

Techniques:

• "Interestingly" - Marker for unexpected
• Observation: Specific finding
• "in line with the observation in" - Literature support
• Explanation: Why it makes sense

---

9. Minimalism Keywords: Design Simplification Vocabulary

Philosophy Keywords:

• "minimal" - "minimal adaptations"
• "sufficient" - "is sufficient to" (not "optimal")
• "simple" - "simple feature pyramid"
• "plain" - "plain backbone"
• "decouple" - "decouple pre-training from fine-tuning"
• "independence" - "independence of upstream vs downstream"

Direction Keywords:

• "pursue a different direction" - Positioning
• "in contrast to" - Differentiation
• "abandons" - What you give up
• "enables" - What your approach allows

Measured Claim Keywords:

• "under some circumstances" - Not always
• "can compete with" - Competitive, not dominant
• "more prominent for" - When effect is stronger
• "is sufficient" - Necessary, not maximal

Surprise Markers (use in order):

1. "Surprisingly" - First finding
2. "More surprisingly" - Deeper finding
3. "Interestingly" - Pattern observation
4. "Notably" - Important detail
5. "It is worth noting that" - Caveat/clarification

---

10. Common Mistakes in Design Simplification Papers

❌ Don't:

• Claim your method is "optimal" - You're simplifying, not optimizing
• Attack common practices - Acknowledge their value first
• Overgeneralize - "under some circumstances" is honest
• Forget to show fair comparison - Prove you tried hard with baselines
• Hide complexity - Admit what you don't address

✅ Do:

• Use "sufficient" instead of "optimal"
• Say what you DON'T do ("do not aim to develop new components")
• Emphasize minimal changes ("minimal adaptations")
• Report when your method wins and when it doesn't
• Show "surprisingly" findings with proper qualification
• Demonstrate fair comparison effort
• Adapt only where necessary (fine-tuning, not pre-training)

---

11. Revision Checklist for Design Simplification Papers

Before Submission, Verify:

• Abstract has "Surprisingly, we observe: (i)... and (ii)..."
• Introduction establishes conventional practice first
• Common solution is acknowledged ("has shown successful results")
• "We pursue a different direction" is stated clearly
• Philosophy is elevated ("fewer inductive biases")
• "More surprisingly" used for deeper claim
• Methods section says "we do not aim to develop new components"
• "minimal adaptations" philosophy stated
• "only during fine-tuning" boundary specified
• Fair comparison effort described explicitly
• Self-validation shown (our reproduction of others is good)
• Multi-factor analysis in results (scaling, wall-clock)
• Ablations show incremental progression
• "sufficient" used, not "optimal"
• Under what conditions is stated ("under some circumstances")

---

12. Example: Applying This Pattern

Original Idea (Not Design Simplification):

"We propose a new feature pyramid that improves detection AP by 3%."

Design Simplification Version:

"Modern detectors use hierarchical backbones with FPN. With plain ViT emerging as powerful backbones, a common solution re-introduces hierarchy (abandoning the plain philosophy). We pursue a different direction: plain backbones with minimal adaptations. Surprisingly, we observe a simple feature pyramid from a single-scale map is sufficient (without FPN). More surprisingly, with MAE pre-training, ViTDet competes with hierarchical detectors, especially for larger models. This decouples pre-training from fine-tuning, maintaining upstream/downstream independence."

The Design Simplification Frame:

• Conventional: Hierarchy + FPN
• Challenge: Plain ViT is...plain
• Common: Swin (abandons philosophy)
• Ours: Minimal adaptations
• Surprise: Simple is sufficient
• Philosophy: Decoupling, independence

---

Paper Metadata

Title: Exploring Plain Vision Transformer Backbones for Object Detection (ViTDet)

Authors: Yanghao Li, Hanzi Mao, Kaiming He

Venue: ECCV 2022

arXiv: 2203.16527

Key Concepts:

• Plain ViT for detection (no hierarchy needed)
• Simple feature pyramid (no FPN needed)
• Minimal adaptations philosophy
• Decoupling pre-training from fine-tuning
• MAE pre-training synergy

---

Extracted by

Date: 2026-01-26

Source: Analysis of ViTDet paper (21 pages)

Extraction Focus: Design simplification paper writing patterns, "surprisingly" findings reporting, minimal adaptations philosophy, fair comparison strategies

For Integration: ml-paper-writing skill knowledge base