In weakly supervised learning settings—such as Positive-Unlabeled (PU) learning, complementary-label learning, or learning from label proportions—we are tasked with training a classifier without access to full, clean ground-truth labels. To bypass this limitation, a common strategy is to construct an unbiased risk estimator. By mathematically rewriting the true classification risk function, researchers can derive an objective function that relies solely on the available weak supervision.
However, this mathematical elegance hides a dangerous practical vulnerability when paired with modern deep learning: the negative empirical risk problem.
The Mechanics of Unbiased Risk Estimation
To understand why unbiased risk estimators fail, we must first look at how they are constructed. When we lack explicit negative labels (as in PU learning) or explicit positive labels (as in complementary-label learning), we cannot directly compute the error rate for that missing class. Instead, we estimate it indirectly by taking a broader data pool and subtracting out the components we can measure.
For instance, in PU learning, the risk of the negative class is typically estimated by evaluating the model's performance on the entire unlabeled dataset and subtracting its weighted performance on the positive dataset. Because we are subtracting risk components to correct for the lack of direct supervision, the resulting mathematical formula contains negative weights.
In the population limit—meaning if we had infinite data—this formulation is perfectly sound; the true expected risk is always non-negative ($\ge 0$). However, in practice, models are trained on finite mini-batches. Due to random sampling noise, the empirical risk calculated on a specific batch can easily fluctuate below zero, yielding a mathematically valid but physically impossible negative loss.
Why Deep Networks Hack the Loss
When empirical risk drops below zero, high-capacity models like deep neural networks encounter a catastrophic failure mode.
Because deep networks are incredibly flexible, they can easily exploit these negative terms in the rewritten risk estimators. As Masahiro Kato et al. describe in the context of density ratio estimation, this leads to a phenomenon known as train-loss hacking. Instead of learning the underlying data distribution, the network minimizes the objective function by overfitting drastically to specific finite-sample artifacts, driving the empirical loss toward negative infinity.
As Takashi Ishida et al. demonstrate, when the empirical risk becomes negative, the optimization process effectively begins to reverse the learning of the target task. The model begins to memorize label noise and misclassify clean data to satisfy the broken objective, resulting in a total collapse in generalization.
This vulnerability is directly tied to model capacity. Nan Lu et al. show that while simple linear models are largely unaffected because they lack the flexibility to exploit these non-physical risk regions, the impact is severely pronounced in deep multi-layer perceptrons and residual networks. Consequently, stabilizing these estimators is a prerequisite for using deep learning in weakly supervised settings.
Standard Supervised Loss:
[ Positive Risk ] + [ Negative Risk ] ===> Always ≥ 0
Unbiased Weakly Supervised Loss:
[ Unlabeled Risk ] - [ Positive Risk ] ===> Can drop below 0 due to batch sampling noise
===> Deep networks exploit this to drive loss to -∞
The Landscape of Solutions
To prevent deep models from exploiting negative empirical risk, researchers have developed several core strategies to enforce non-negativity and stabilize the loss landscape.
1. Rectification and Hard Clipping
The most intuitive fix is to wrap the vulnerable risk components in a rectification operator, such as a Rectified Linear Unit (ReLU), which forces any negative risk value directly to zero. As Lei Feng et al. demonstrate, bounding the empirical risk ensures stable convergence under weak supervision.
However, hard clipping introduces a new problem: the "dying gradient." When a risk component is clipped to zero, its gradient vanishes, causing the optimization process to stall. To maintain active gradients, Nan Lu et al. propose a leaky non-negative correction that maintains a small slope for negative values. Similarly, Taira Tsuchiya et al. utilize non-monotonic activation functions, such as a generalized leaky rectifier, to prevent negative risk while ensuring continuous gradient flow.
2. Absolute Value Correction
An alternative to zero-clipping is the absolute value transformation. Zayd Hammoudeh et al. show that taking the absolute value of a negative risk estimate acts as a natural mathematical penalty. Instead of ignoring the negative region, it forces the model to move back toward valid, positive risk territory without requiring complex hyperparameter tuning. Furthermore, Zhongnian Li et al. confirm that absolute value correction preserves vital gradient signals in negative risk regions far better than hard thresholding.
3. Gradient Ascent Regularization
Rather than altering the loss function itself, some approaches modify the optimization path. Takashi Ishida et al. introduce a dynamic switching strategy: whenever a mini-batch yields a negative risk value, the optimizer temporarily switches from gradient descent to gradient ascent. This acts as a self-correcting regularizer, actively pushing the empirical risk back toward zero whenever it enters non-physical territory.
4. Intrinsically Non-Negative Formulations
Because heuristic corrections (like clipping or absolute values) alter the original optimization landscape and can introduce non-convexity, some researchers argue for abandoning rewritten unbiased estimators altogether. For instance, Shuqi Liu et al. propose formulating alternative risk estimators that are inherently non-negative by design, such as through order-preserving losses, eliminating the root cause of the instability.
The Over-Correction Trap
While enforcing a non-negative floor prevents model divergence, it introduces a subtle secondary challenge.
In PU learning, when a negative risk estimator is strictly clamped to zero, the model can become biased. Because errors on the unlabeled data are artificially discounted whenever the estimator hits the zero floor, the model begins to treat false negatives as "cheaper" than false positives. As Ximing Li et al. observe, this creates a severe imbalance between false negative and false positive rates, causing the model to over-prioritize negative class identification at the expense of positive class accuracy.
To resolve this, the authors introduce a dual-constraint mechanism that tracks an exponential moving average of the positive class risk. This moving average is used to set a dynamic lower bound for the negative class risk rather than a hard floor of zero, preventing the estimator from collapsing and ensuring balanced error rates throughout the entire deep learning training process.