gradient_analysis_BarrierTrust_1.md

Loss Values at Convergence

Loss	Mean	Std
nll	0.007336	0.004056
pairwise	0.087583	0.034836
trust	0.000404	0.001692
anchor	0.006579	0.001478
residual	0.058852	0.014836

Lambda-weighted Gradient Magnitudes

Parameter group: aleatoric (unc_head.ale_head.weight, unc_head.ale_head.bias)

Loss	Mean ||grad||	Std ||grad||	CV
nll	0.000166	0.000114	69.0%
pairwise	0.000827	0.000214	25.9%
trust	0.000000	0.000000	∞
anchor	0.000180	0.000045	24.8%
residual	0.000000	0.000000	∞

Parameter group: calibration (unc_head.calibration_head.weight, unc_head.calibration_head.bias)

Loss	Mean ||grad||	Std ||grad||	CV
nll	0.000684	0.000391	57.1%
pairwise	0.003024	0.000640	21.2%
trust	0.000000	0.000000	∞
anchor	0.000692	0.000162	23.3%
residual	0.000000	0.000000	∞

Parameter group: probe (unc_head.alpha, unc_head.probe_conv.weight, unc_head.probe_to_class.weight)

Loss	Mean ||grad||	Std ||grad||	CV
nll	0.000228	0.000102	44.8%
pairwise	0.004244	0.001395	32.9%
trust	0.000052	0.000127	244.1%
anchor	0.002462	0.000288	11.7%
residual	0.006048	0.001423	23.5%

Parameter group: ranking (unc_head.err_a, unc_head.err_b, unc_head.err_c)

Loss	Mean ||grad||	Std ||grad||	CV
nll	0.000000	0.000000	∞
pairwise	0.000207	0.000117	56.6%
trust	0.000000	0.000000	∞
anchor	0.000984	0.000111	11.3%
residual	0.000000	0.000000	∞

Gradient Cosine Similarities

Parameter group: aleatoric (unc_head.ale_head.weight, unc_head.ale_head.bias)

Loss pair	Mean cos	Std cos	Negative %
nll_vs_pairwise	+0.042	0.296	50%
nll_vs_anchor	-0.162	0.208	78%
pairwise_vs_anchor	+0.452	0.183	0%

Parameter group: calibration (unc_head.calibration_head.weight, unc_head.calibration_head.bias)

Loss pair	Mean cos	Std cos	Negative %
nll_vs_pairwise	-0.172	0.041	100%
nll_vs_anchor	+0.149	0.071	0%
pairwise_vs_anchor	+0.926	0.028	0%

Parameter group: probe (unc_head.alpha, unc_head.probe_conv.weight, unc_head.probe_to_class.weight)

Loss pair	Mean cos	Std cos	Negative %
nll_vs_pairwise	+0.749	0.313	8%
nll_vs_trust	-0.580	0.336	100%
nll_vs_anchor	+0.963	0.037	0%
nll_vs_residual	-0.928	0.039	100%
pairwise_vs_trust	-0.410	0.321	92%
pairwise_vs_anchor	+0.647	0.343	8%
pairwise_vs_residual	-0.613	0.285	92%
trust_vs_anchor	-0.573	0.336	100%
trust_vs_residual	+0.604	0.377	0%
anchor_vs_residual	-0.921	0.056	100%

Parameter group: ranking (unc_head.err_a, unc_head.err_b, unc_head.err_c)

Loss pair	Mean cos	Std cos	Negative %
pairwise_vs_anchor	-0.213	0.132	92%