blob: 6aa39cffa8bb9a4cf213b4d34fbaafaa98694b8f (
plain) (
blame)
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
|
# Strict Analysis Report
## Analysis Question
- Compare Full Model, Frozen Encoder, and Subject Adapter on cross-subject decoding.
- Determine whether the adapter closes the transfer gap without sacrificing stability.
## Data Inventory
- 3 model families
- 5 random seeds per family
- subject-level WER and CER
- training logs for convergence
- ablation outputs for adapter width
## Executive Summary
- Subject Adapter improves mean WER over Frozen Encoder by 3.8 absolute points.
- The improvement is statistically significant under paired testing after multiple-comparison correction.
- Full fine-tuning remains strongest overall, but Subject Adapter offers the best compute-performance tradeoff.
- Performance variance is lower than Frozen Encoder, suggesting a more stable transfer path.
## Main Findings
### 1. Main comparison
- Full fine-tuning: `31.4 ± 1.9` WER
- Subject Adapter: `33.2 ± 1.3` WER
- Frozen Encoder: `37.0 ± 2.1` WER
Observation:
- Subject Adapter consistently outperforms Frozen Encoder across all 5 seeds.
Interpretation:
- The transfer bottleneck is not purely feature reuse; a light adaptation head captures subject-specific alignment that freezing alone misses.
Implication:
- Future work should prioritize adapter variants before investing in heavier full-model tuning.
### 2. Stability
- Subject Adapter has the smallest seed variance among transfer-friendly methods.
- Convergence curves show less oscillation after epoch 8.
Interpretation:
- Adapter tuning is not only better on average, but easier to optimize.
### 3. Ablation
- Reducing adapter width from 256 to 64 hurts WER by 1.6 points.
- Increasing from 256 to 512 yields only marginal improvement.
Interpretation:
- Most of the benefit is captured at moderate width; scale-up is not the current bottleneck.
## Caveats
- Only 5 seeds; inference about tails remains limited.
- Subject count is modest, so subject-level heterogeneity may still be under-estimated.
- Training-time comparison depends on identical hardware assumptions.
## Recommended Next Move
- Promote this finding into a `results-report` note focused on adapter vs freezing.
- Run one robustness pass on held-out low-resource subjects.
|
