1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
|
"""
Render Figure 1: Four-panel audit hero figure.
Panel (arch): Architecture diagram (3arc.pdf, merged from external)
Panel A: Standard pair (accuracy × aggregate Γ) — all 3 methods in the green zone
Panel B: Per-layer cosine — FA and DFA form an X-cross, BP flat at 1.0
Panel C: Per-layer ||g_l|| — BP flat, FA gentle decay, DFA cliff
"""
import os
import json
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
from matplotlib.backends.backend_pdf import PdfPages
import numpy as np
from PIL import Image
import subprocess
REPO_ROOT = "/home/yurenh2/fa"
# ─── DATA ────────────────────────────────────────────────────────────────
# Panel A: accuracy and aggregate Γ (3-seed means where available)
# BP: from protocol_audit (d=256 L=4, 3-seed)
# FA: from fa_main_audit (d=256 L=4, 3-seed)
# DFA: from protocol_audit + paper
panel_a = {
"BP": {"acc": 0.6147, "gamma": 1.0},
"FA": {"acc": 0.401, "gamma": 0.250},
"DFA": {"acc": 0.306, "gamma": 0.100},
}
# Panel B: per-layer cosine (4 blocks, l=0..3)
# BP: by definition cos(bp_grad, bp_grad) = 1.0
# FA: 3-seed mean from fa_main_audit
# DFA: from d=512 L=4 s42 final (pattern robust across d/L)
panel_b = {
"BP": [1.0, 1.0, 1.0, 1.0],
"FA": [0.016, 0.072, -0.085, 0.997],
"DFA": [0.400, 0.001, -0.0004, -0.002],
}
# Panel C: per-layer ||g_l|| (5 layers, l=0..4)
# All from d=256 L=4 s42 (protocol_audit for BP/DFA, fa_main_audit for FA)
panel_c = {
"BP": [4.40e-4, 4.71e-4, 4.79e-4, 4.53e-4, 3.70e-4],
"FA": [1.79e-5, 1.21e-6, 8.85e-7, 8.89e-7, 8.89e-7],
"DFA": [4.39e-7, 4.19e-9, 4.18e-9, 4.17e-9, 4.17e-9],
}
CLAMP_EPS = 1e-8 # PyTorch F.cosine_similarity default eps
# ─── STYLE ───────────────────────────────────────────────────────────────
COLORS = {"BP": "#2166ac", "FA": "#e08214", "DFA": "#b2182b"}
MARKERS = {"BP": "o", "FA": "s", "DFA": "D"}
plt.rcParams.update({
"font.size": 9,
"axes.labelsize": 10,
"axes.titlesize": 10,
"legend.fontsize": 8,
"xtick.labelsize": 8,
"ytick.labelsize": 8,
"font.family": "serif",
})
fig, axes = plt.subplots(1, 3, figsize=(10.5, 4.5))
fig.subplots_adjust(wspace=0.38, left=0.06, right=0.97, bottom=0.30, top=0.94)
# ─── PANEL A: Standard pair (x=Γ, y=acc) ────────────────────────────────
ax = axes[0]
ax.set_title("(A) Standard reporting pair", fontsize=9, fontweight="bold", loc="left")
# Axes: x = aggregate Γ (cosine), y = test accuracy
x_lim = (-0.22, 1.12)
y_lim = (-0.02, 0.72)
# Four quadrants: boundaries at x=0 (cos=0) and y=0.10 (chance)
# Upper-right (green): cos > 0 AND acc > chance
ax.fill_between([0, x_lim[1]], 0.10, y_lim[1], color="#c8e6c9", alpha=0.5, zorder=0)
# Lower-left (red): cos < 0 AND acc < chance
ax.fill_between([x_lim[0], 0], y_lim[0], 0.10, color="#ffcdd2", alpha=0.5, zorder=0)
# Upper-left (light gray): cos < 0 AND acc > chance
ax.fill_between([x_lim[0], 0], 0.10, y_lim[1], color="#f5f5f5", alpha=0.6, zorder=0)
# Lower-right (light gray): cos > 0 AND acc < chance
ax.fill_between([0, x_lim[1]], y_lim[0], 0.10, color="#f5f5f5", alpha=0.6, zorder=0)
# Quadrant boundary lines
ax.axvline(0, color="gray", lw=0.6, ls="--", zorder=1)
ax.axhline(0.10, color="gray", lw=0.6, ls="--", zorder=1)
# Quadrant labels
ax.text(-0.11, 0.41, "cos < 0,\nacc > chance", fontsize=6.5, color="#888",
ha="center", va="center", style="italic", rotation=90)
ax.text(0.55, 0.04, "cos > 0,\nacc < chance", fontsize=6.5, color="#888",
ha="center", va="center", style="italic")
ax.text(0.55, 0.41, '"looks like\n learning"', fontsize=7, color="#388e3c",
ha="center", va="center", fontweight="bold")
ax.text(-0.11, 0.04, "neither", fontsize=6.5, color="#c62828",
ha="center", va="center", style="italic")
# Points: x=gamma, y=acc
for method in ["BP", "FA", "DFA"]:
d = panel_a[method]
ax.scatter(d["gamma"], d["acc"], c=COLORS[method], marker=MARKERS[method],
s=70, zorder=5, edgecolors="k", linewidths=0.5)
# Labels
ax.annotate("BP", (panel_a["BP"]["gamma"], panel_a["BP"]["acc"]),
xytext=(-8, 8), textcoords="offset points", fontsize=8, fontweight="bold",
color=COLORS["BP"])
ax.annotate("FA", (panel_a["FA"]["gamma"], panel_a["FA"]["acc"]),
xytext=(8, 5), textcoords="offset points", fontsize=8, fontweight="bold",
color=COLORS["FA"])
ax.annotate("DFA", (panel_a["DFA"]["gamma"], panel_a["DFA"]["acc"]),
xytext=(8, -8), textcoords="offset points", fontsize=8, fontweight="bold",
color=COLORS["DFA"])
ax.set_xlabel("Aggregate $\\Gamma$ (cosine)")
ax.set_ylabel("Test accuracy")
ax.set_xlim(*x_lim)
ax.set_ylim(*y_lim)
# ─── PANEL B: Per-layer cosine ──────────────────────────────────────────
ax = axes[1]
ax.set_title("(B) Per-block cosine $\\cos(a_l, g_l)$", fontsize=9, fontweight="bold", loc="left")
blocks = np.arange(4)
for method in ["BP", "FA", "DFA"]:
vals = panel_b[method]
ax.plot(blocks, vals, color=COLORS[method], marker=MARKERS[method],
markersize=5, linewidth=1.8, label=method, zorder=3)
ax.axhline(0, color="gray", lw=0.5, ls=":", zorder=1)
ax.set_xlabel("Block $l$")
ax.set_ylabel("$\\cos(a_l,\\, \\nabla_{h_l} \\mathcal{L})$")
ax.set_xticks(blocks)
ax.set_xticklabels([f"$l={l}$" for l in blocks])
ax.set_ylim(-0.25, 1.12)
# ─── PANEL C: Per-layer ||g_l|| ─────────────────────────────────────────
ax = axes[2]
ax.set_title("(C) Per-layer $\\|g_l\\|$ (BP gradient)", fontsize=9, fontweight="bold", loc="left")
layers = np.arange(5)
for method in ["BP", "FA", "DFA"]:
vals = panel_c[method]
ax.semilogy(layers, vals, color=COLORS[method], marker=MARKERS[method],
markersize=5, linewidth=1.8, label=method, zorder=3)
ax.set_xlabel("Layer $l$")
ax.set_ylabel("$\\|\\partial\\mathcal{L}/\\partial h_l\\|_2$ (median)")
ax.set_xticks(layers)
ax.set_xticklabels([f"$h_{l}$" for l in layers])
ax.set_ylim(5e-12, 5e-2)
# ─── GRID (all panels) ───────────────────────────────────────────────────
for ax in axes:
ax.grid(True, which="major", color="#d0d0d0", linewidth=0.4, linestyle=":")
ax.set_axisbelow(True)
# Panel C also needs minor grid for log scale
axes[2].grid(True, which="minor", color="#e8e8e8", linewidth=0.3, linestyle=":")
# ─── CAPTION BOXES below each panel ─────────────────────────────────────
captions = [
"With standard reporting pair, FA and\nDFA reached non-trivial accuracy and\npositive cosine alignment in this setting",
"Aggregated cosine lies: shallow layers\nof FA and deep layers of DFA are not\nlearning or aligned well",
"Reference also fails: DFA collapses to\nnumerical noise at depth, FA decays\n2 orders of magnitude across layers",
]
fig.canvas.draw()
box_h = 0.13 # height of caption box in figure coords
box_gap = 0.12 # gap between axes bottom and box top
for i, (ax, txt) in enumerate(zip(axes, captions)):
bbox = ax.get_position()
bx0 = bbox.x0
bx1 = bbox.x1
by_top = bbox.y0 - box_gap
by_bot = by_top - box_h
# Draw rounded rectangle
fancy = mpatches.FancyBboxPatch(
(bx0, by_bot), bx1 - bx0, box_h,
boxstyle="round,pad=0.008",
facecolor="#f7f7f7", edgecolor="#aaaaaa", linewidth=0.7,
transform=fig.transFigure, clip_on=False)
fig.patches.append(fancy)
# Text centered in the box
fig.text((bx0 + bx1) / 2, by_bot + box_h / 2, txt,
ha="center", va="center", fontsize=9.5, style="italic",
transform=fig.transFigure)
# ─── SAVE ────────────────────────────────────────────────────────────────
out = os.path.join(REPO_ROOT, "paper/figures/fig1_audit_hero.pdf")
fig.savefig(out, bbox_inches="tight", dpi=300)
out_png = out.replace(".pdf", ".png")
fig.savefig(out_png, bbox_inches="tight", dpi=200)
print(f"Saved: {out}")
print(f"Saved: {out_png}")
|
