CV Classification Loss Function Design

Research Question

Design a novel classification loss function for deep convolutional neural networks that improves test accuracy across different architectures and datasets.

Background

The cross-entropy loss is the standard training objective for classification networks, but it has known limitations: it treats all misclassifications equally, assigns high confidence to correct predictions without margin, and does not adapt to training dynamics. Researchers have proposed alternatives:

• Label Smoothing (Szegedy et al., 2016): Softens hard targets to prevent overconfidence, CE with targets = (1-eps)*one_hot + eps/C
• Focal Loss (Lin et al., ICCV 2017): Down-weights easy examples via (1-pt)^gamma modulation
• PolyLoss (Leng et al., ICLR 2022): Extends CE with polynomial correction terms, CE + eps*(1-pt)

However, these methods are either static or address only specific failure modes. There is room to design loss functions that combine multiple insights: confidence calibration, curriculum-style epoch adaptation, class-count awareness, or learned temperature scaling.

What You Can Modify

The compute_loss(logits, targets, config) function (lines 165-185) in custom_loss.py. This function receives raw logits, integer targets, and a config dict, and must return a differentiable scalar loss.

You can modify:

• The loss formulation (cross-entropy variants, margin losses, etc.)
• Confidence-based reweighting schemes
• Epoch-dependent curriculum strategies using config['epoch'] and config['total_epochs']
• Class-count-dependent behavior using config['num_classes']
• Temperature or logit scaling
• Auxiliary regularization terms (entropy, logit penalties, etc.)

The config dict provides: num_classes (int), epoch (int, 0-indexed), total_epochs (int).

Important: The evaluation loss (for test_loss reporting) always uses standard cross-entropy. Your loss function only affects training.

Evaluation

• Metric: Best test accuracy (%, higher is better)
• Architectures & datasets:
  • ResNet-56 on CIFAR-100 (deep residual, 100 classes)
  • VGG-16-BN on CIFAR-100 (deep non-residual with BatchNorm, 100 classes)
  • MobileNetV2 on FashionMNIST (lightweight inverted-residual, 10 classes) — hidden, evaluated on final submission only
• Training: SGD (lr=0.1, momentum=0.9, wd=5e-4), cosine annealing, 200 epochs
• Data augmentation: RandomCrop(32, pad=4) + RandomHorizontalFlip