del cvpr

papers/petl-visual-recognition/blacklist.txt

@@ -1,2 +0,0 @@
-# Official repository
-https://github.com/OSU-MLB/PETL_Vision

papers/petl-visual-recognition/config.yaml

@@ -1,8 +0,0 @@
-title: "Lessons Learned from a Unifying Empirical Study of PETL in Visual Recognition"
-pdf_url: "https://arxiv.org/pdf/2409.16434.pdf"
-venue: "ICLR 2025"
-year: "2025"
-extra:
-  selection_index: 11
-  domain: "Computer Vision"
-  paradigm: "Empirical Comparison"

papers/petl-visual-recognition/paper.pdf

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:18c13e25306b3a5bab3917cc5206807538caaa519690f24b5eeefb056ceb2dec
-size 2479066

papers/petl-visual-recognition/sau.json

@@ -1,337 +0,0 @@
-{
-  "paper_id": "petl-visual-recognition",
-  "paper_title": "Parameter-Efficient Transfer Learning for Visual Recognition",
-  "D1": [
-    {
-      "id": "petl-visual-recognition-D1-001",
-      "claim": "Backbone: ViT-B/16 with 86M params, 12 Transformer layers, patch_size P=16, embedding_dim D determined by ViT-B/16 spec, num_heads Nh determined by ViT-B/16 spec, head_dim Dh = D/Nh, MLP expansion ratio=4",
-      "source": "Section 3 (Setup), Appendix B.1 (Vision Transformer)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-002",
-      "claim": "LayerNorm parameters: ~38K (0.04% of total 86M backbone)",
-      "source": "Appendix B.2.3 (Direct Selective Tuning, LayerNorm)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-003",
-      "claim": "VTAB-1K low-shot benchmark: 19 tasks in 3 groups (Natural 7, Specialized 4, Structured 8), 1000 training samples per dataset, train/val split 800/200 (80/20 split)",
-      "source": "Section 3 (Dataset)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-004",
-      "claim": "VTAB-1K training setup: AdamW optimizer, batch_size=64, epochs=100, cosine_decay lr scheduler, no data augmentation, drop_path_rate search [0, 0.1], PEFT param cap <=1.5% of ViT-B/16 (~Q/K/V of single MSA block)",
-      "source": "Section 3 (Setup), Appendix A.1 (VTAB-1K)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-005",
-      "claim": "VTAB-1K hyperparameter search: learning_rate range [0.001, 0.01], weight_decay range [0.0001, 0.001], grid search over method-specific ranges",
-      "source": "Appendix A.1 (VTAB-1K)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-006",
-      "claim": "VPT-Shallow: prompt_number range [5, 10, 50, 100, 200], param range [0.3K, 153K]",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-007",
-      "claim": "VPT-Deep: prompt_number range [5, 10, 50, 200], param range [46K, 921K]",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-008",
-      "claim": "BitFit: 102K tunable params, tunes bias terms in patch embedding projection, Q/K/V weights, MLP FC layers, LN blocks",
-      "source": "Appendix A.1 (Table 3), Appendix B.2.3 (BitFit)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-009",
-      "claim": "DiffFit: 140K tunable params (bias + LN tuning + feature scaling)",
-      "source": "Appendix A.1 (Table 3), Appendix B.2.3 (DiffFit)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-010",
-      "claim": "LayerNorm Tuning: 38K tunable params (0.04% of backbone)",
-      "source": "Appendix A.1 (Table 3), Appendix B.2.3 (LayerNorm)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-011",
-      "claim": "SSF (Scale & Shift Features): 205K tunable params, modulates features at positions h2, h3, h5, h7, h8, h9 per Transformer layer",
-      "source": "Appendix A.1 (Table 3), Appendix B.2.3 (SSF)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-012",
-      "claim": "Pfeiffer Adapter: scale_factor range [0.01, 0.1, 1, 10], bottleneck range [4, 8, 16, 32], param range [82K, 599K]",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-013",
-      "claim": "Houlsby Adapter: scale_factor range [0.01, 0.1, 1, 10], bottleneck range [4, 8, 16, 32], param range [165K, 1.198M]",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-014",
-      "claim": "AdaptFormer: scale_factor range [0.05, 0.1, 0.2], bottleneck range [4, 16, 32], param range [82K, 599K]",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-015",
-      "claim": "RepAdapter: scale_factor range [0.1, 0.5, 1, 5, 10], bottleneck range [8, 16, 32], param range [239K, 903K], group-wise transformation with G groups",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-016",
-      "claim": "Convpass: scale_factor range [0.01, 0.1, 1, 10, 100], bottleneck range [8, 16], Xavier init [true, false], param range [327K, 664K]",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-017",
-      "claim": "LoRA: bottleneck rank r range [1, 8, 16, 32] with r << D, applied to W_Q and W_V matrices, param range [36K, 1.179M]",
-      "source": "Appendix A.1 (Table 3), Appendix B.2.3 (LoRA)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-018",
-      "claim": "FacT-TT: scale_factor range [0.01, 0.1, 1, 10, 100], bottleneck range [8, 16, 32], param range [21K, 196K], stacks 12M weight matrices (12 layers x 6 matrices Q/K/V/O/W1/W2)",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-019",
-      "claim": "FacT-TK: scale_factor range [0.01, 0.1, 1, 10, 100], bottleneck range [16, 32, 64], param range [30K, 369K]",
-      "source": "Appendix A.1 (Table 3)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-020",
-      "claim": "Many-shot training setup: AdamW, batch_size=64, epochs=40, cosine_decay lr scheduler, lr range [0.0005, 0.001], weight_decay range [0.0001, 0.001]",
-      "source": "Section 5 (Setup), Appendix A.1 (Many-shot)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-021",
-      "claim": "Many-shot datasets: CIFAR-100 (50K train, 100 classes, horizontal_flip, 90/10 split), RESISC (25.2K train, 45 classes, horizontal+vertical flips, 90/10 split), Clevr-Distance (70K train, 6 classes, no augmentation)",
-      "source": "Section 5 (Dataset), Appendix A.2 (Many-shot Datasets)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-022",
-      "claim": "Many-shot PEFT parameter budget study: effective_param_percent range [2%, 5%]",
-      "source": "Section 5 (Results, Recipes)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-023",
-      "claim": "Robustness experiment: CLIP ViT-B/16 backbone, ImageNet-1K target with 100 shots/class, lr=3e-5, weight_decay=0.005, strong data augmentation (NPS), 80 class text prompts, distribution shift test sets: ImageNet-V2/ImageNet-R/ImageNet-S/ImageNet-A",
-      "source": "Section 7 (Setup), Appendix A.1 (Robustness Setup, Robustness Model)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-024",
-      "claim": "WiSE mixing coefficient alpha: linear interpolation coefficient in [0,1], evaluated at multiple points for target-robustness trade-off analysis",
-      "source": "Section 7 (WiSE for PEFT)"
-    },
-    {
-      "id": "petl-visual-recognition-D1-025",
-      "claim": "Pre-training datasets: ImageNet-21K (14M images), ImageNet-1K (1.3M images), LAION-5B (5B images)",
-      "source": "Section 1 (Introduction), Section 2.1"
-    },
-    {
-      "id": "petl-visual-recognition-D1-026",
-      "claim": "Computation hardware: 8x NVIDIA RTX 6000 Ada GPUs, 2x AMD EPYC 9554 64-Core CPUs, 800GB RAM",
-      "source": "Appendix A.1 (Computation)"
-    }
-  ],
-  "D2": [
-    {
-      "id": "petl-visual-recognition-D2-001",
-      "claim": "ViT Transformer Layer: Z'_m = MSA(LN(Z_{m-1})) + Z_{m-1}; Z_m = MLP(LN(Z'_m)) + Z'_m — forward pass with pre-norm, residual MSA, and residual MLP blocks. Each layer L_m maps Z_{m-1} in R^{D x (1+N)} to Z_m in R^{D x (1+N)}.",
-      "source": "Appendix B.1 (Vision Transformer)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-002",
-      "claim": "Single-Head Self-Attention: Q^{(i)} = W_Q^{(i)} Z; K^{(i)} = W_K^{(i)} Z; V^{(i)} = W_V^{(i)} Z where W_{Q/K/V}^{(i)} in R^{Dh x D}, Dh = D/Nh. Attn^{(i)}(Z) = V^{(i)} x Softmax(K^{(i)^T} Q^{(i)} / sqrt(D_h)) in R^{Dh x (1+N)}.",
-      "source": "Appendix B.1 (Vision Transformer)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-003",
-      "claim": "MSA Output: MSA(Z) = W_O [Attn^0(Z), ..., Attn^{N_h}(Z)] — concatenation of all head outputs projected by W_O in R^{D x (D_h * N_h)}",
-      "source": "Appendix B.1 (Vision Transformer)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-004",
-      "claim": "MLP Block: MLP(Z) = GELU(Z W_1 + b_1) W_2 + b_2 with W_1 in R^{D x 4D}, W_2 in R^{4D x D}, b_1 in R^{4D}, b_2 in R^{D}, GELU activation. Two FC layers (FC1, FC2) with expansion ratio 4.",
-      "source": "Appendix B.1 (Vision Transformer)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-005",
-      "claim": "ViT Input Construction: Z_0 = [x_0^{(Class)}, x_0^{(1)}, ..., x_0^{(N)}] + E_{pos} where N = H*W/P^2 patches. Each patch I^{(n)} in R^{P^2 x C} is flattened and embedded via trainable linear projection to x_0^{(n)} in R^D. A learnable class token x_0^{(Class)} is prepended, and position embeddings E_{pos} in R^{D x (1+N)} are added.",
-      "source": "Appendix B.1 (Vision Transformer)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-006",
-      "claim": "LoRA Weight Update: W_{Q/V} + DeltaW_{Q/V} = W_{Q/V} + W_{down}^{Q/V} W_{up}^{Q/V} — low-rank decomposition with W_{down}^{Q/V} in R^{r x D}, W_{up}^{Q/V} in R^{D x r}, r << D. Applied to Q and V matrices only; zero inference overhead via weight merging (DeltaW_{Q/V} merged into W_{Q/V}). W_{up}^{Q/V} initialized to zero so DeltaW = 0 at start of training.",
-      "source": "Appendix B.2.3 (Efficient Selective Tuning, LoRA)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-007",
-      "claim": "LoRA Forward Pass: h_3 = LoRA(h_2) + h_3 where h_3 = [Q, K, V] and LoRA(h_2) = [W_down^Q W_up^Q h_2, 0, W_down^V W_up^V h_2] — additive residual on Q and V only (K unchanged, residual=0). LoRA modifies the Q/V projections within the MSA block.",
-      "source": "Appendix B.2.3 (Efficient Selective Tuning, LoRA)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-008",
-      "claim": "Houlsby Adapter: two adapter modules placed at h_5 (after MSA) and h_9 (after MLP). Formally: h_5 = Adapter_1(h_5), h_9 = Adapter_2(h_9) where Adapter(h) = s * W_up * sigma(W_down * h) + h. The bottleneck module composes W_down in R^{r x D} (down-projection), nonlinear activation sigma (GELU/ReLU), W_up in R^{D x r} (up-projection), scale factor s, and residual skip connection. r << D for parameter efficiency.",
-      "source": "Appendix B.2.2 (Adapter-based Methods, Houlsby Adapter)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-009",
-      "claim": "Pfeiffer Adapter: single adapter placed only after MLP block. Formally: h_9 = Adapter(h_9) where Adapter(h) = s * W_up * sigma(W_down * h) + h. Same bottleneck module structure (W_down in R^{r x D}, sigma, W_up in R^{D x r}, scale factor s, skip connection) as Houlsby but more parameter-efficient: one adapter per layer instead of two.",
-      "source": "Appendix B.2.2 (Adapter-based Methods, Pfeiffer Adapter)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-010",
-      "claim": "AdaptFormer: parallel adapter at h_9 = h_9 + Adapter(h_7), where Adapter(h_7) = s * W_up * sigma(W_down * h_7). Domain-specific features from the adapter (taking pre-MLP feature h_7 as input) complement domain-agnostic MLP output h_9. The parallel design differs from Houlsby/Pfeiffer sequential placement — adapter runs alongside (not after) the MLP block.",
-      "source": "Appendix B.2.2 (Adapter-based Methods, AdaptFormer)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-011",
-      "claim": "Convpass: convolutional by-pass with Convpass(h) = s * W_up * sigma(Conv2d(sigma(W_down * h))). Architecture: W_down: 1x1 conv (D -> r), sigma activation, Conv2d: 3x3 conv (r -> r, same input/output channels), sigma activation, W_up: 1x1 conv (r -> D). Placed parallel: h_5 = Convpass_1(h_2) + h_5; h_9 = Convpass_2(h_7) + h_9. The 3x3 conv encodes visual inductive bias via hard-coded locality over nearby patch tokens.",
-      "source": "Appendix B.2.2 (Adapter-based Methods, Convpass)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-012",
-      "claim": "RepAdapter: linear re-parameterizable adapter without nonlinearity, placed sequentially: h_5 = RepAdapter_1(h_2) and h_7 = RepAdapter_2(h_7). Formally: RepAdapter(h) = s * phi_up(phi_down(h)) + h where phi_down(h) = W_down * h (W_down in R^{r x D}) and phi_up(tilde_h) = [W_{g1} * tilde_h_{g1}, ..., W_{gG} * tilde_h_{gG}] with G group-wise projections. Each group splits tilde_h into G chunks of size r/G, and W_{gi} in R^{D/G x r/G}. Includes internal skip connection +h. Can be merged into original weights for zero inference overhead due to linearity.",
-      "source": "Appendix B.2.2 (Adapter-based Methods, RepAdapter)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-013",
-      "claim": "VPT-Shallow: [tilde_P_1, Z_1] = L_1([P_0, Z_0]); [tilde_P_m, Z_m] = L_m([tilde_P_{m-1}, Z_{m-1}]) for m=2..M — learnable prompts P_0 in R^{l x D} (l prompts, D dimensional) prepended at first layer input. Prompt outputs tilde_P_{m-1} propagated through subsequent layers. Only P_0 is updated; backbone frozen.",
-      "source": "Appendix B.2.1 (Prompt-based Methods, VPT-Shallow)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-014",
-      "claim": "VPT-Deep: [_, Z_m] = L_m([P_{m-1}, Z_{m-1}]) for m=1..M — separate learnable prompts P_{m-1} in R^{l x D} per layer m. Previous layer prompt outputs discarded (not propagated, denoted by _), fresh prompts at each layer input. All P_{m-1} across M layers are updated during fine-tuning.",
-      "source": "Appendix B.2.1 (Prompt-based Methods, VPT-Deep)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-015",
-      "claim": "SSF Feature Modulation: h_i = SSF_i(h_i) = w^i odot h_i + b^i for i in {2, 3, 5, 7, 8, 9} — per-channel learnable scale w^i in R^D and shift b^i in R^D at 6 feature positions per Transformer layer. odot denotes channel-wise multiplication. SSF accommodates distribution difference between upstream and downstream datasets via linear transformation. Merged into weights at inference for zero overhead.",
-      "source": "Appendix B.2.3 (Direct Selective Tuning, SSF)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-016",
-      "claim": "DiffFit Feature Scaling: h_5 = gamma_1 * h_5; h_9 = gamma_2 * h_9 — learnable per-channel scale factors gamma_1, gamma_2 in R^D applied after MSA and MLP blocks respectively. Combined with BitFit (all bias terms) and LayerNorm tuning (LN weights and biases). Total: 140K tunable params for ViT-B/16. DiffFit = BitFit + LN-Tune + feature scaling factors.",
-      "source": "Appendix B.2.3 (Direct Selective Tuning, DiffFit)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-017",
-      "claim": "BitFit: fine-tunes only bias terms b in all network components. Per Transformer layer: Q = W_Q * Z + b_Q (tune b_Q, freeze W_Q); K = W_K * Z + b_K; V = W_V * Z + b_V; FC_attn output bias b_attn; FC1: Z * W_1 + b_1 (tune b_1 in R^{4D}, freeze W_1); FC2: GELU(.) * W_2 + b_2 (tune b_2 in R^D). LN blocks: LN(h) = W_LN * (h - mu)/sigma + b_LN (tune W_LN, b_LN). Also tunes bias in patch embedding projection. All weight matrices remain frozen. 102K tunable params for ViT-B/16.",
-      "source": "Appendix B.2.3 (Direct Selective Tuning, BitFit)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-018",
-      "claim": "LayerNorm-Only Tuning: updates LN weight and bias per LN block (2 LN per layer). Formally: LN_1(h_1) = W_{LN_1} * (h_1 - mu_1)/sigma_1 + b_{LN_1} (before MSA, at h_2); LN_2(h_6) = W_{LN_2} * (h_6 - mu_2)/sigma_2 + b_{LN_2} (before MLP, at h_7). Each LN contains 2 trainable parameters {W_LN, b_LN} in R^D. Total: 4DM params (~38K for ViT-B/16 with M=12, D=768; 0.04% of 86M backbone).",
-      "source": "Appendix B.2.3 (Direct Selective Tuning, LayerNorm)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-019",
-      "claim": "FacT-TT (Tensor-Train): stacks Q/K/V/O/W1/W2 across M layers into 3D tensor W_{FacT} in R^{12M x D x D}. Weight update: DeltaW_{FacT} = s * Sigma x_2 U^T x_3 V^T with factor matrices U in R^{D x r}, V in R^{D x r} and TT-core Sigma in R^{12M x r x r}. x_j denotes mode-j product. s is the scaling factor.",
-      "source": "Appendix B.2.3 (Efficient Selective Tuning, FacT)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-020",
-      "claim": "FacT-TK (Tucker): DeltaW_{FacT} = s * A x_1 B^T x_2 U^T x_3 V^T — core tensor A in R^{r x r x r}, factor matrices B in R^{12M x r}, U in R^{D x r}, V in R^{D x r}. Tucker decomposition generalizes TT by allowing interactions across all three modes via the core tensor A. s is the scaling factor.",
-      "source": "Appendix B.2.3 (Efficient Selective Tuning, FacT)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-021",
-      "claim": "FacT Forward Pass: modifies h_3, h_5, h_8, h_9 via extracted weight updates from DeltaW_{FacT}. For h_5: h_5 = h_4 * DeltaW_O + h_5 where DeltaW_O is the W_O update extracted from the full stacked tensor DeltaW_{FacT}, which contains updates for all 6 weight matrices (Q/K/V/O/W1/W2) across all M layers. Similarly, DeltaW_Q/K/V modify h_3, DeltaW_1/2 modify h_8, h_9.",
-      "source": "Appendix B.2.3 (Efficient Selective Tuning, FacT)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-022",
-      "claim": "WiSE Head Interpolation: W_head = alpha * W_head^{ft} + (1 - alpha) * W_{zero-shot} — mixes fine-tuned prediction head with zero-shot CLIP text-embedding head (columns are class-name text embeddings from 80 ensembled prompts). alpha in [0,1] controls the blend between fine-tuned and zero-shot knowledge.",
-      "source": "Section 7 (How Robust are PEFT Methods to Distribution Shifts?)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-023",
-      "claim": "WiSE for Adapter-Based Methods: Adapter_{wise}(h) = alpha * s * W_up * sigma(W_down * h) + h — scales the adapter module contribution by alpha to control domain-specific vs. domain-agnostic feature blending. Since most adapter-based methods include residual connections (skip connection +h), WiSE functions as a feature ensemble where alpha controls how strongly domain-specific adapter features blend with domain-agnostic backbone features.",
-      "source": "Section 7 (How Robust are PEFT Methods to Distribution Shifts?)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-024",
-      "claim": "WiSE for Efficient Selective Tuning (LoRA, FacT): W_{wise} = W + alpha * DeltaW — scales the additive residual DeltaW by alpha in [0,1]. For LoRA: W_{Q/V}^{wise} = W_{Q/V} + alpha * W_down^{Q/V} W_up^{Q/V}. For FacT: the extracted per-matrix update DeltaW_O (and similarly for other matrices) is scaled by alpha before merging.",
-      "source": "Section 7 (How Robust are PEFT Methods to Distribution Shifts?)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-025",
-      "claim": "WiSE for Direct Selective Tuning: param_{wise} = alpha * param_{tuned} + (1 - alpha) * param_{pretrained} — linearly interpolates tuned and pre-trained parameters for BitFit (bias terms: b_{wise} = alpha * b_tuned + (1-alpha) * b_pretrained), LayerNorm (W_LN, b_LN), DiffFit (bias + LN + scale factors gamma_{wise} = alpha * gamma_tuned + (1-alpha) * 1). Exploits the fact that fine-tuned parameters remain near the original loss basin.",
-      "source": "Section 7 (How Robust are PEFT Methods to Distribution Shifts?)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-026",
-      "claim": "Ensemble Logit Averaging: hat_y = argmax(1/K * sum_{k=1}^{K} f_k(x)) — averages logits from K PEFT methods (K=14) fine-tuned on the same dataset for final prediction. Each PEFT method produces logits f_k(x) in R^{C} (C classes), and the averaged logit vector is argmax'ed. Leverages prediction diversity across PEFT methods for consistent accuracy gains.",
-      "source": "Section 4 (Different PEFT Approaches Offer Complementary Information)"
-    },
-    {
-      "id": "petl-visual-recognition-D2-027",
-      "claim": "CLIP Zero-Shot Classification: hat_y = argmax_j <g(x), h(s_j)> — cosine similarity between image embedding g(x) and class-name text embeddings h(s_j) for caption 'a photo of a c_j'. The text encoder produces text embeddings for 80 ensembled prompts per class. Equivalent to f(x) = g(x)^T W_{zero-shot} where W_{zero-shot} columns are averaged text embeddings from 80 prompts. The prediction head is initialized from W_{zero-shot} for PEFT robustness experiments.",
-      "source": "Section 2.1 (Large pre-trained models), Section 7"
-    }
-  ],
-  "D3": [
-    {
-      "id": "petl-visual-recognition-D3-001",
-      "claim": "P1 — Low-Shot PEFT Comparison: Compare 14 PEFT methods + Linear Probing + Full Fine-Tuning on VTAB-1K benchmark (19 datasets, 3 groups) using ViT-B/16 pre-trained on ImageNet-21K. Hyperparameter tuning via grid search over method-specific ranges (lr, weight_decay, drop_path). Train on 800/200 (80/20) split for HP search, retrain on full 1000 images after selection. Primary metric: Top-1 Accuracy (%). PEFT param cap: <=1.5% of backbone.",
-      "source": "Section 3 (PEFT Methods in Low-Shots Regime)"
-    },
-    {
-      "id": "petl-visual-recognition-D3-002",
-      "claim": "P2 — Prediction Diversity Analysis: Quantify prediction complementarity among 14 PEFT methods on VTAB-1K using (a) prediction similarity matrix (percentage test samples where method i and j agree), (b) high-confidence correct prediction overlap (Venn diagram of top 5K most confident correct), (c) low-confidence wrong prediction overlap (Venn diagram of 5K least confident wrong), (d) within-category vs. cross-category similarity analysis. Primary metric: prediction overlap percentage.",
-      "source": "Section 4 (Different PEFT Approaches Offer Complementary Information)"
-    },
-    {
-      "id": "petl-visual-recognition-D3-003",
-      "claim": "P3 — Ensemble Evaluation: Average logits across all 14 PEFT methods per sample then argmax for final prediction. Evaluated on all 19 VTAB-1K test sets. Baseline: worst-performing PEFT method per dataset (relative baseline = 0). Primary metric: relative accuracy gain over worst PEFT per dataset.",
-      "source": "Section 4 (Different PEFT Approaches Offer Complementary Information)"
-    },
-    {
-      "id": "petl-visual-recognition-D3-004",
-      "claim": "P4 — Many-Shot PEFT Evaluation: Evaluate PEFT methods on full-size datasets (CIFAR-100 50K/100 classes, RESISC 25.2K/45 classes, Clevr-Distance 70K/6 classes) using ViT-B/16 ImageNet-21K. AdamW, batch_size=64, epochs=40, cosine_decay, lr [0.0005, 0.001], weight_decay [0.0001, 0.001]. Study parameter budgets at 2%, 5% and higher. Baseline: Linear Probing and Full Fine-Tuning. Primary metric: Top-1 Accuracy vs. tunable parameter count.",
-      "source": "Section 5 (PEFT Methods in Many-Shot Regime)"
-    },
-    {
-      "id": "petl-visual-recognition-D3-005",
-      "claim": "P5 — Why PEFT Works Analysis: Analyze PEFT accuracy relative to Linear Probing and Full Fine-Tuning across VTAB-1K tasks. Case 1 (Full FT > Linear Probing): backbone update needed to close domain gap (e.g., RESISC, Clevr-Distance). Case 2 (Linear Probing > Full FT): pre-trained features sufficient; updating risks over-fitting (e.g., CIFAR-100). Per-task line plots with Linear Probing < PEFT < Full FT ordered by tunable parameter count.",
-      "source": "Section 6 (Why Do PEFT Methods Work?)"
-    },
-    {
-      "id": "petl-visual-recognition-D3-006",
-      "claim": "P6 — Robustness to Distribution Shift: Fine-tune CLIP ViT-B/16 with PEFT on ImageNet-1K (100 shots/class), freeze visual encoder, initialize head from CLIP text embeddings (80 ensembled prompts). Evaluate on ImageNet-1K test + 4 distribution shift datasets (ImageNet-V2, ImageNet-R, ImageNet-S, ImageNet-A). Training: lr=3e-5, weight_decay=5e-3, strong augmentation. Compare Full FT, CLIP Zero-Shot, 8 PEFT methods, and WiSE variants. Analyze target-robustness Pareto frontier via alpha sweep.",
-      "source": "Section 7 (How Robust are PEFT Methods to Distribution Shifts?)"
-    },
-    {
-      "id": "petl-visual-recognition-D3-007",
-      "claim": "P7 — DINOv2 vs. ImageNet-21K Comparison: Compare 5 PEFT methods (SSF, Houl. Adapter, AdaptFormer, Convpass, LoRA) on DINOv2 (self-supervised) vs. ImageNet-21K (supervised) ViT-B/16 backbones across selected VTAB-1K tasks (4 per group). Same training setup as P1. Compute Delta = DINOv2 accuracy - IN21k accuracy per (method, dataset). Analyze architecture interaction: parallel vs. sequential adapter designs on self-supervised features.",
-      "source": "Appendix C (Performance comparison between DINOv2 and IN21k)"
-    },
-    {
-      "id": "petl-visual-recognition-D3-008",
-      "claim": "P8 — Drop Path Rate Ablation: Quantify impact of stochastic depth regularization on PEFT with drop_path=0 vs. drop_path=0.1 across all 14 PEFT methods on VTAB-1K. Primary metric: accuracy gain (positive delta = regularization helps).",
-      "source": "Appendix C (Drop-path-rate), Figure 11"
-    },
-    {
-      "id": "petl-visual-recognition-D3-009",
-      "claim": "P9 — Data Augmentation Ablation: Validate no-augmentation decision for VTAB-1K by comparing no augmentation vs. simple DA (RandomResizedCrop + RandomVerticalFlip + RandomHorizontalFlip) on 9 selected datasets across all 3 groups. Metric: accuracy delta (negative = augmentation hurts). Finding: augmentation does not uniformly benefit all VTAB-1K datasets.",
-      "source": "Appendix A.1 (Table 4)"
-    }
-  ],
-  "D4": [
-    {
-      "id": "petl-visual-recognition-D4-001",
-      "claim": "WiSE-Merging Procedure (P6). Entry condition: PEFT model fine-tuned on ImageNet-1K (100 shots/class) with CLIP ViT-B/16 backbone, zero-shot CLIP text head W_{zero-shot} available, alpha in [0,1] selected. Step 1 (ALL methods, must execute first) — Head Interpolation: compute W_head = alpha * W_head^{ft} + (1 - alpha) * W_{zero-shot}, mixing the fine-tuned FC head with the CLIP zero-shot text-embedding head. Exit: interpolated head ready. Then branch to one of 3 method-specific procedures based on PEFT category. Step 2A (Adapter-Based: Houlsby, Pfeiffer, AdaptFormer, Convpass, RepAdapter) — Adapter Scaling: scale each adapter module output by alpha: Adapter_{wise}(h) = alpha * s * W_up * sigma(W_down * h) + h, controlling domain-specific vs. domain-agnostic feature blending. Exit: scaled adapter model ready for evaluation. Step 2B (Efficient Selective Tuning: LoRA, FacT) — Residual Scaling: scale additive residual by alpha: W_{wise} = W + alpha * DeltaW, where DeltaW is the low-rank/tensor update. Exit: merged model ready for evaluation. Step 2C (Direct Selective Tuning: BitFit, LayerNorm, DiffFit) — Parameter Merging: linearly interpolate tuned parameters with pre-trained parameters: param_{wise} = alpha * param_{tuned} + (1 - alpha) * param_{pretrained} for bias terms, LN weights/biases, and scaling factors. Exit: merged model ready for evaluation. Steps 2A, 2B, 2C are mutually exclusive (exactly one applies per PEFT method).",
-      "source": "Section 7 (How Robust are PEFT Methods to Distribution Shifts?)"
-    },
-    {
-      "id": "petl-visual-recognition-D4-002",
-      "claim": "Prediction Diversity Analysis Pipeline (P2, 4 parallel procedures). Entry condition: 14 PEFT methods fine-tuned and evaluated on each VTAB-1K dataset (19 datasets), prediction logits and confidence scores collected per test sample, per method. Step 1 — Compute Prediction Similarity Matrix: for each VTAB-1K dataset, compute entry (i, j) = (number of test samples where method i and j predict the same class) / (total test samples). Output: 14 x 14 similarity matrix per dataset. Step 2 — High-Confidence Correct Prediction Overlap: for each method, select top 5K most confident correct predictions per dataset; compute Venn diagram overlap of these sets across 3 methods (LoRA, SSF, Adapter). Output: overlap counts and Venn visualization. Step 3 — Low-Confidence Wrong Prediction Overlap: for each method, select 5K least confident wrong predictions per dataset; compute Venn diagram overlap. Output: overlap counts and Venn visualization. Step 4 — Within-Category vs. Cross-Category Similarity: group methods into adapter-based, selective-tuning, prompt-based categories; compute average prediction overlap within each category vs. across categories. Output: within-group and cross-group similarity scores. Exit: all diversity metrics and visualizations generated. Steps 1-4 are independently executable (no data dependencies); can run in parallel on the same set of trained models.",
-      "source": "Section 4 (Different PEFT Approaches Offer Complementary Information), Appendix C"
-    },
-    {
-      "id": "petl-visual-recognition-D4-003",
-      "claim": "Robustness Evaluation Pipeline (P6, 3 procedures with partial ordering). Entry condition: PEFT model fine-tuned on ImageNet-1K (100 shots/class) with CLIP ViT-B/16, prediction head initialized from CLIP text embeddings, test sets ImageNet-1K, ImageNet-V2, ImageNet-R, ImageNet-S, ImageNet-A available. Step 1 — Target Distribution Accuracy: evaluate fine-tuned model on ImageNet-1K test set; compute Top-1 Accuracy (%). Exit: target accuracy score (numeric). Step 2 — Distribution Shift Accuracy: evaluate on 4 shift datasets (ImageNet-V2, ImageNet-R, ImageNet-S, ImageNet-A); compute Top-1 Accuracy per dataset and average across 4 datasets. Exit: shift accuracy scores (4 per-dataset values + 1 average, all numeric). Steps 1 and 2 are parallelizable (independent model evaluations on different test sets, no data dependency). Step 3 — WiSE Trade-off Analysis: for each PEFT method, sweep alpha over [0, 1] at multiple evaluation points. For each alpha: run WiSE merging — head interpolation (mixing fine-tuned head with zero-shot CLIP text head) + the corresponding method-specific procedure (adapter scaling, residual scaling, or parameter merging, per WiSE-Merging), then evaluate on target (Step 1 above) and shift (Step 2 above) test sets. Plot target accuracy (x-axis) vs. average shift accuracy (y-axis) for each PEFT method; compare with Full FT + WiSE to identify Pareto frontier. Exit: Pareto frontier plot showing trade-off curves for all PEFT methods and Full FT. Step 3 depends on results from both Step 1 and Step 2 (requires per-alpha target and shift accuracy values generated by WiSE variants, which are parameterized versions of the models evaluated in Steps 1-2).",
-      "source": "Section 7 (How Robust are PEFT Methods to Distribution Shifts?)"
-    }
-  ]
-}