Experiments

Neural Networks Paper: New Experiments

Notes 19.04.2021

• Vordefinierte Filter
  • Wavelet lernen
    • vollkommen frei
    • DoG
    • wavelet space? Difference-of-Beta-Space am flexibelsten?
  • See if correlation plot sinus curves stem from frequency of Gabors
• Effect of MH Width on Order
• Coverage
  • HP Optim mit SemLC
• Autoencoder/Classification
  • RDMs vergleichen
• Farben
  • Vergleich schwarz-weißmodelle

Proof of Concept

Strategies

All strategies use the default circular padding without a self-connection. Second set of accuracies refers to results with random augmentations of the test set.

Strategy	Accuracy	Uncertainty	Accuracy	Uncertainty	N
Baseline Alexnet	83.33	0.05	81.04	0.05	180
Baseline AlexCMap	83.36	0.08	81.10	0.10	60
SSI Frozen	84.03	0.07	81.82	0.08	60
SSI Adaptive	84.16	0.07	81.96	0.08	60
Converged Frozen	84.07	0.08	81.78	0.08	60
Converged Adaptive	84.21	0.07	81.93	0.08	60
Parametric	83.98	0.09	81.85	0.14	30

Self-Connection

Strategy	Accuracy	Uncertainty	Accuracy	Uncertainty	N
Converged Frozen	84.09	0.12	81.76	0.12	30
Converged Adaptive	84.20	0.09	81.96	0.14	30
Parametric	83.98	0.09	81.85	0.14	30
Converged Frozen Self	84.09	0.26	81.82	0.17	10
Converged Adaptive Self	83.99	0.15	82.00	0.15	10
Parametric Self	84.05	0.21	81.77	0.17	10

Padding

Strategy	Accuracy	Uncertainty	Accuracy	Uncertainty	N
SSI Adaptive	84.16	0.07	81.96	0.08	60
SSI Frozen	84.03	0.07	81.82	0.08	60
Converged Frozen	84.07	0.08	81.78	0.08	60
Converged Adaptive	84.21	0.07	81.93	0.08	60
Parametric	83.98	0.09	81.85	0.14	30
SSI Adaptive Zero	84.18	0.07	81.97	0.09	60
SSI Frozen Zero	84.10	0.19	81.75	0.25	10
Converged Frozen Zero	84.02	0.07	81.81	0.08	60
Converged Adaptive Zero	84.19	0.06	81.96	0.09	60
Parametric Zero	83.98	0.07	81.79	0.09	60

Inhibition Coverage

The Converged Adaptive Coverage Experiments used scope=27, width=3 and damp=0.1 in all layers.

Strategy	Accuracy	Uncertainty	Accuracy	Uncertainty	N
Parametric full	84.09	0.09	81.92	0.09	30
Parametric @1	83.98	0.09	81.85	0.14	30
Parametric @1,2	84.04	0.11	81.77	0.11	30
Parametric @1,2,3	83.92	0.12	81.82	0.11	30
Converged Adaptive full	84.33	0.09	82.11	0.10	30
Converged Adaptive @1	84.21	0.07	81.93	0.08	60
Converged Adaptive @1,2	84.18	0.11	82.06	0.11	30
Converged Adaptive @1,2,3	84.24	0.13	82.05	0.09	30

Competitors

Strategy	Dataset	Accuracy	Uncertainty	Accuracy	Uncertainty	N
VGG-19	CIFAR-10	90.18	0.16			30
VGG-19 + Best	CIFAR-10	90.21	0.15			30
AlexNet	CIFAR-10	83.36	0.08	81.10	0.10	60
AlexNet + Best	CIFAR-10	84.43	0.07	82.28	0.08	60
AlexNet	ImageNet
AlexNet + Best	ImageNet
CapsNet	MNIST