# can_gradient_clipping_mitigate_label_noise__635bb714.pdf

Published as a conference paper at ICLR 2020

CAN GRADIENT CLIPPING MITIGATE LABEL NOISE?

Aditya Krishna Menon, Ankit Singh Rawat, Sashank J. Reddi, Sanjiv Kumar Google Research New York, NY USA {adityakmenon,ankitsrawat,sashank,sanjivk}@google.com

Gradient clipping is a widely-used technique in the training of deep networks, and is generally motivated from an optimisation lens: informally, clipping controls the dynamics of iterates, thus enhancing the rate of convergence to a local minimum. This intuition has been made precise in a line of recent works, which show that suitable clipping can yield significantly faster convergence than vanilla gradient descent. In this paper, we propose a new lens for studying gradient clipping, namely, robustness: informally, one expects clipping to mitigate the effects of noise, since one does not overly trust any single sample. Surprisingly, we prove that for the common problem of label noise in classification, standard gradient clipping does not in general provide robustness. On the other hand, we show that a simple variant of gradient clipping is robust, and is equivalent to suitably modifying the underlying loss function. As a special case, this yields a simple, noise-robust modification of the standard cross-entropy loss which performs well empirically.

1 INTRODUCTION: TWO FACES OF GRADIENT CLIPPING

Gradient clipping is a widely-used technique in the training of deep networks (Mikolov, 2012; Pascanu et al., 2012; 2013; Mnih et al., 2015; Gehring et al., 2017; Merity et al., 2018; Peters et al., 2018), and in its simplest form, involves capping the global parameter gradient norm at a specified threshold. Constraining the gradient norm is paramount in training models such as recurrent neural networks, which may otherwise suffer from an exploding gradient problem (Bengio et al., 1994). Clipping has seen broader value in other neural models, and is abstractly understood as being valuable from an optimisation lens: intuitively, by ensuring that the gradient norm is not too large, the dynamics of iterates are more well-behaved (Hazan et al., 2015; Levy, 2016; Zhang et al., 2019). A distinct line of work has also explored the value of gradient clipping from a privacy lens (Abadi et al., 2016; Pichapati et al., 2019): intuitively, it prevents any single instance from overly influencing parameter updates, which is a core requirement in ensuring differential privacy (Dwork et al., 2006).

In this paper, we propose a new lens with which to study gradient clipping, namely, robustness: intuitively, clipping the gradient prevents over-confident descent steps, which is plausibly beneficial in the presence of noise. Given this intuition, our interest is whether gradient clipping can mitigate the problem of label noise in classification, which has received significant recent interest (Scott et al., 2013; Natarajan et al., 2013; Menon et al., 2015; Liu & Tao, 2016; Patrini et al., 2017; Ghosh et al., 2017; Han et al., 2018; Zhang & Sabuncu, 2018; Song et al., 2019; Thulasidasan et al., 2019; Charoenphakdee et al., 2019). We study this question, and provide three main contributions:

(a) we show that gradient clipping alone does not endow label noise robustness to even simple models. Specifically, we show that under stochastic gradient descent with linear models, gradient clipping is related to optimising a Huberised loss (Lemma 1, 2). While such Huberised losses preserve classification calibration (Lemma 3), they are not robust to label noise (Proposition 4). (b) we propose composite loss-based gradient clipping, a variant that does have label noise robustness. Specifically, for losses comprising a base loss composed with a link function, (e.g., softmax crossentropy), we only clip the contribution of the base loss. The resulting partially Huberised loss preserves classification calibration (Lemma 6), while being robust to label noise (Proposition 7). (c) we empirically verify that on both synthetic and real-world datasets, partially Huberised versions of standard losses (e.g., softmax cross-entropy) perform well in the presence of label noise ( 5).

Published as a conference paper at ICLR 2020

φ(v) Logistic Huberised Partially Huberised

0.2 0.4 0.6 0.8

Log-loss Huberised log-loss

Figure 1: Illustration of the Huberised and partially Huberised logistic loss (left panel). Recall that the logistic loss comprises the log-loss (or cross-entropy) composed with a sigmoid link. The Huberised loss arises from loss-based gradient clipping, and linearises the entire logistic loss beyond a threshold. The partially Huberised loss arises from composite loss-based gradient clipping, and linearises only the base loss (i.e., log-loss) beyond a threshold (right panel). Once combined with the sigmoid link, the overall partially Huberised loss asymptotically saturates. Our analysis in 3, 4 establishes that Huberised losses are not robust to label noise, but partially Huberised losses are (Proposition 4, 7).

To illustrate the difference between standard and composite loss-based gradient clipping in (a) and (b), consider the pervasive softmax cross-entropy loss, viz. the logistic loss for binary classification. In (a), we relate gradient clipping to loss Huberisation , which linearises the logistic loss when it exceeds a fixed threshold. In (b), we introduce composite loss-based gradient clipping, which is equivalent to a partial loss Huberisation that linearises only the cross-entropy loss but leaves the sigmoid link untouched. Figure 1 illustrates the Huberised and partially Huberised logistic loss.

2 BACKGROUND AND RELATED WORK

We provide background on gradient clipping, loss functions for classification, and label noise.

Gradient clipping. Consider a supervised learning task over instances X and labels Y, where we have a family of models indexed by θ Θ, and the quality of a particular model is measured by a loss function ℓθ : X Y R. The gradient for a mini-batch {(xn, yn)}N n=1 is

n=1 ℓθ(xn, yn). (1)

One may instead compute the clipped gradient, which for user-specified threshold τ > 0 is

gτ(θ) .= clipτ(g(θ)) clipτ(w) .=

( τ w ||w||2 if ||w||2 τ w else. (2)

Employing gτ(θ) for optimisation corresponds to clipped gradient descent. This is closely related to normalised gradient descent (NGD) (Hazan et al., 2015), wherein one employs g(θ) .= g(θ) ||g(θ)||2 . Hazan et al. (2015) showed that NGD can lead to convergence for a wider class of functions than standard gradient descent. Levy (2016) showed that NGD can escape saddle points for non-convex problems. Zhang et al. (2019) showed that gradient clipping can lead to accelerated convergence over gradient descent. Gradient clipping has also been explored for privacy (Abadi et al., 2016; Pichapati et al., 2019), motivated by it preventing any single instance from dominating parameter updates.

Loss functions. In binary classification, one observes samples from a distribution D over X { 1}, and seeks a predictor f : X R with low risk R(f) .= E ℓ(y, f(x)) according to a loss which, in an overload of notation, we denote ℓ: { 1} R R+. For the zero-one loss ℓ01(y, f) = Jy f < 0K, R(f) is known as the misclassification risk. Rather than directly using ℓ01, for computational convenience one often employs a margin loss ℓ(y, f) = φ(y f), for convex φ: R R+.

We say that φ is classification calibrated (Zhang, 2004; Bartlett et al., 2006) if driving the excess risk over the Bayes-optimal predictor for φ to zero also drives the excess risk for ℓ01 to zero; that is,

Published as a conference paper at ICLR 2020

minimising the φ-risk is statistically consistent for classification. A canonical example is the hinge loss φ(z) = [1 z]+. We call φ admissible if it is well-behaved in the sense of being bounded from below, strictly convex, continuously differentiable, non-increasing, and classification calibrated.

We say that φ is proper composite (Reid & Williamson, 2010; Nock & Nielsen, 2009) (or for brevity composite) if its minimising scores can be interpreted as probabilities. Composite here refers to such losses comprising a base loss ϕ composed with an invertible link function F : R [0, 1]. While φ accepts as input real-valued scores (e.g., the final layer logits of a neural network), these are internally converted to probabilities via F. A canonical example is the logistic loss φ(z) = log F(z) with F : z 7 σ(z) for sigmoid σ(z) .= (1+e z) 1. The multiclass analogue is the softmax cross-entropy loss, wherein the sigmoid becomes a softmax. See Appendix B for a more technical discussion.

Learning under label noise. In classification under label noise, one has samples from distribution D where P D(y | x) is a noisy version of PD(y | x), e.g., all labels are corrupted with a fixed constant probability. The goal remains to ensure low risk with respect to the clean D. This problem has a long history in statistics (Ekholm & Palmgren, 1982; Copas, 1988), and has emerged as a topic of recent interest in machine learning. There has been particular interest in the problem of learning under symmetric label noise, wherein all instances have a constant probability of their labels being flipped uniformly to any of the other classes. Scott et al. (2013); Katz-Samuels et al. (2019) proposed a framework for analysing the more general setting of class-conditional label noise. Takenouchi & Eguchi (2004); Natarajan et al. (2013); Cid-Sueiro et al. (2014); Patrini et al. (2017); Natarajan et al. (2018); van Rooyen & Williamson (2018) studied loss-correction for obtaining unbiased risk estimators. Recent works have explored ideas ranging from to model (dis)agreement (Malach & Shalev-Shwartz, 2017; Han et al., 2018; Song et al., 2019) to abstention (Thulasidasan et al., 2019).

3 GRADIENT CLIPPING AND HUBERISED LOSSES

We first show that gradient clipping in general does not endow robustness to label noise, even in simple settings. Specifically, we establish that stochastic gradient clipping with linear models is equivalent to modifying the underlying loss (Lemma 1). This modified loss is closely related to a Huberised loss (Lemma 2), which is equivalent to loss-based gradient clipping (L-gradient clipping). Unfortunately, Huberised losses, and thus L-gradient clipping, are not robust to label noise (Proposition 4).

3.1 FROM GRADIENT TO L-GRADIENT CLIPPING

Consider a binary classification problem, with Y = { 1}. Suppose we use a scoring model sθ(x) for θ Θ with margin mθ(x, y) .= y sθ(x), and margin loss ℓθ(x, y) .= φ(mθ(x, y)) for admissible φ: R R+. Then, θℓθ(x, y) = θmθ(x, y) φ (mθ(x, y)). (3)

Now suppose we perform stochastic gradient descent (i.e., we use N = 1 in (1)) and use a linear scorer sθ(x) = θTx. Here, gradient clipping of (3) is equivalent to modifying the loss as follows.

Lemma 1. Pick any admissible margin loss φ, and τ > 0. Then, for loss ℓθ(x, y) .= φ(mθ(x, y)) with linear scorer, the clipped loss gradient (2) is equivalent to the gradient of a modified loss ℓ:

clipτ ( ℓθ(x, y)) = ℓθ(x, y),

with ℓθ(x, y) .=

( τ mθ(x,y) 2 mθ(x, y) if |φ (mθ(x, y))| τ mθ(x,y) 2 φ(mθ(x, y)) else.

The loss ℓin (4) is intuitive: if the derivative of the original loss φ exceeds a certain effective clipping threshold, we replace φ with a linear loss function. This effective threshold τ ( mθ(x, y) 2) 1 takes into account the instance-dependent margin gradient, viz. x 2 for linear sθ(x).

We will comment in 4.4 as to the properties of gradient clipping in more general scenarios. For the moment, we note that Lemma 1 is closely related to the Huber loss (Huber, 1964),

ℓhuber(x, y; θ) = τ |sθ(x) y| if |sθ(x) y| τ 1 2 (sθ(x) y)2 else. (5)

Published as a conference paper at ICLR 2020

This replaces the extremes of the square loss with the absolute loss. Both the absolute and Huber losses are widely employed in robust regression (Huber, 1981, Chapter 3), (Hampel et al., 1986, pg. 100). Note that (4) slightly differs from (5) as the effective clipping threshold is instance-dependent. One may nonetheless arrive at Huber-style losses via a variant of gradient clipping; this connection will prove useful for our subsequent analysis. Per (3), gradient clipping involves using

clipτ ( ℓθ(x, y)) = clipτ ( mθ(x, y) φ (mθ(x, y))) , (6)

wherein the clipping considers the margin gradient. Consider now the following loss-based gradient clipping (L-gradient clipping), wherein we clip only the contribution arising from the loss:

l-clipτ ( ℓθ(x, y)) .= mθ(x, y) clipτ (φ (mθ(x, y))) . (7)

Compared to (6), we effectively treat the margin gradient norm mθ(x, y) 2 as constant across instances, and focus on bounding the loss derivative. The latter may be a significant component in the gradient norm; e.g., for linear models, mθ(x, y) 2 = x 2 is often bounded. Observe further that for linear models with x 2 R across instances, (7) is a rescaled version of the clipped gradient:

clipτ ( ℓθ(x, y)) = τ y x φ (mθ(x, y)) τ (R |φ (mθ(x, y))|) = mθ(x, y) clipτ/R (φ (mθ(x, y))) .

Interestingly, L-gradient clipping equivalently uses the following Huberised version of the loss. Lemma 2. Pick any admissible margin loss φ: R R+ with Fenchel conjugate φ , and τ > 0. Then, the clipped gradient in (7) is equivalent to employing a Huberised loss function φτ such that

clipτ (φ (z)) = φ τ(z) with φτ(z) .= τ z φ ( τ) if φ (z) τ φ(z) else. (8)

Evidently, φτ linearises φ once its derivative is sufficiently large, akin to the Huber loss in (5). One may verify that for φ(z) = (1 z)2, one arrives exactly at (5).

Example 3.1: For the logistic loss φ(z) = log(1 + e z), the Huberised loss for τ (0, 1) is

φτ(z) = τ z log(1 τ) τ σ 1 (τ) if z σ 1 (τ) log(1 + e z) else.

Per Figure 1a, this linearises φ beyond a fixed threshold. See Appendix C for further illustrations.

The use of Huberised losses in classification is not new (Zhang et al., 2002; Zhang & Johnson, 2003; Rosset & Zhu, 2007). However, we now provide a novel study of their label noise robustness.

3.2 CONSISTENCY AND LABEL-NOISE ROBUSTNESS OF HUBERISED LOSSES

Before studying the effect of label noise on Huberised losses, it is apposite to consider whether they are suitable for use even in the absence of noise. One way to formalise this is to ask whether the losses maintain classification calibration, in the sense of 2. This is a minimal requirement on a loss to be useful for classification (Zhang, 2004; Bartlett et al., 2006). One may further ask whether the losses preserve class-probabilities, i.e., are proper composite in the sense of 2. It is desirable to preserve this key trait of losses such as the logistic loss. The following clarifies both points. Lemma 3. Pick any admissible margin loss φ and τ > 0. Then, the Huberised loss φτ in (8) is classification calibrated. If φ is proper composite and τ φ (0), then φτ is also proper composite.

L-gradient clipping is thus benign for classification, generalising Rosset & Zhu (2007, Section 3.4), which was for the square-hinge loss. Interestingly, for composite φ and small τ, the proof reveals that φτ has a non-invertible link function. Intuitively, such φτ are effectively linear, and linear losses are unsuited for estimating probabilities (van Rooyen et al., 2015; Charoenphakdee et al., 2019).

We now turn to our central object of inquiry: does gradient clipping endow robustness to label noise? To study this, we consider L-gradient clipping, which as noted above is a special case of gradient clipping under linear models with constant mθ(x, y) 2. Since L-gradient clipping is in turn equivalent to using a Huberised loss, we may study the robustness properties of this loss.

Published as a conference paper at ICLR 2020

Surprisingly, when our loss is convex (e.g., softmax cross-entropy), Huberised losses are not robust to even very simple forms of label noise. Essentially, since these losses are still convex, they can still be affected by errant outlier observations. Formally, using a result of Long & Servedio (2010) (see Appendix E.1), we arrive at the following. Proposition 4. Pick any admissible margin loss φ and τ > 0. Then, a separable distribution for which the optimal linear classifier under φτ is equivalent to random guessing under symmetric noise.

To situate Proposition 4 in a broader context, we note that for regression problems, it is well known that the Huber loss is susceptible to high leverage outliers (Hampel et al., 1986, pg. 313), (Rousseeuw & Leroy, 1987, pg. 13), (Maronna et al., 2019, pg. 104), i.e., extremal instances which dominate the optimal solution. Proposition 4 complements these for the case of label noise in classification.

Given that gradient clipping does not endow label noise robustness, how else might we proceed? In a regression context, the outlier vulnerability of the Huber loss can be addressed by using a trimmed average of the loss (Rousseeuw & Leroy, 1987; Bhatia et al., 2015; Yang et al., 2018). Such ideas have been successfully explored for label noise problems (Shen & Sanghavi, 2019). We will however demonstrate that a simple variant of clipping yields a loss that does possess label noise robustness.

4 COMPOSITE LOSS-BASED GRADIENT CLIPPING

We now show that noise robustness can be achieved with CL-gradient clipping, a variant wherein for composite losses (e.g., softmax cross-entropy), we perform partial Huberisation of the base loss only.

4.1 FROM HUBERISATION TO PARTIAL HUBERISATION

Consider a composite margin loss ℓθ(x, y) .= φ(mθ(x, y)), where φ = ϕ F for some base loss ϕ and invertible link F : R [0, 1]; e.g., the logistic loss has ϕ(u) = log u and F(z) = σ(z). We can interpret pθ(x, y) .= F(mθ(x, y)) as a probability estimate; e.g., pθ(x, 1) = σ(mθ(x, 1)) is the probability of x being positive. Now, rewriting ℓθ(x, y) = ϕ(pθ(x, y)), we may express (3) as:

ℓθ(x, y) = pθ(x, y) ϕ (pθ(x, y)).

L-gradient clipping in (7) was defined as only clipping φ = (ϕ F) above, which ensures that the resulting Huberised loss is Lipschitz. Typically, however, F is already Lipschitz; e.g., this is the case for the commonly used sigmoid or softmax link. This suggests the following composite loss-based gradient clipping (CL-gradient clipping), wherein we only clip the derivative for the base loss ϕ:

cl-clipτ ( ℓθ(x, y)) .= pθ(x, y) clipτ (ϕ (pθ(x, y))) . (9)

As before, CL-gradient clipping corresponds to optimising a new, partially Huberised loss. Lemma 5. Pick any admissible, composite margin loss φ = ϕ F, and τ > 0. Then, the clipped gradient in (9) is equivalent to employing a partially Huberised loss φτ = ϕτ F, where

ϕτ(u) .= τ u ϕ ( τ) if ϕ (u) τ ϕ(u) else. (10)

Compared to the Huberised loss (7), the partially Huberised loss only linearises the base loss ϕ, while retaining the link F. Consequently, the composite loss φτ behaves like the link beyond a certain threshold, and will thus be bounded: indeed, φτ(z) = τ F(z) ϕ ( τ) if ϕ (F(z)) τ.

Example 4.1: For the logistic loss, φ(z) = ϕ(F(z)) with ϕ(u) = log u and F(z) = σ(z),

φτ(z) = τ σ(z) + log τ + 1 if z σ 1 1

log(1 + e z) else. (11)

Note that partial Huberisation readily generalises to a multi-class setting. Indeed, suppose we have softmax probability estimates pθ(x, y) exp(mθ(x, y)). Then, whereas the softmax cross-entropy employs ℓθ(x, y) = log pθ(x, y), our partially Huberised softmax cross-entropy for τ > 1 is

ℓθ(x, y) = τ pθ(x, y) + log τ + 1 if pθ(x, y) 1

τ log pθ(x, y) else. (12)

Published as a conference paper at ICLR 2020

4.2 CONSISTENCY AND LABEL NOISE ROBUSTNESS OF PARTIALLY HUBERISED LOSSES

Following 3.2, we establish that CL-gradient clipping is always benign from a classification perspective, and provided τ is sufficiently large, from a probability estimation perspective as well. As before, we do this by exploiting the equivalence of CL-gradient clipping to a partially Huberised loss. Lemma 6. Pick any admissible composite margin loss φ = ϕ F and τ > 0. Then, the loss φτ in (10) is classification calibrated. If further τ ϕ (1/2), then ϕτ is also proper composite.

We now show that partially Huberised losses have an important advantage over Huberised losses: under symmetric label noise, the optimal solution on the clean distribution cannot be too far away from the optimal solution on the noisy distribution. This implies that label noise (such as that considered in Proposition 4) cannot have an excessively deleterious influence on the loss. Proposition 7. Pick any proper loss ϕ and τ > 0. Let f be the risk minimiser of ϕτ on the clean distribution. For any non-trivial level of symmetric label noise, let regτ(f ) denote the excess risk of f with respect to ϕτ on the noisy distribution. Then, there exists C > 0 such that regτ(f ) C.

Note that by van Rooyen et al. (2015, Proposition 4), it is impossible for the above bound to hold with C = 0 without using a linear loss. Nonetheless, by virtue of partially Huberised losses being partially linear, we are able to bound the degradation under label corruption. The saturating behaviour of the partially Huberised loss also implies robustness to outliers in feature space; see Appendix D.

4.3 RELATION TO EXISTING WORK

The partially Huberised log loss in (10), (11) can be related to a family of losses studied in several works (Ding & Vishwanathan, 2010; Hung et al., 2018; Zhang & Sabuncu, 2018; Sypherd et al., 2019; Amid et al., 2019b;a): for α (0, 1], define the generalised cross-entropy ϕα : [0, 1] R+ by ϕα(u) .= (1 uα)/α. See Appendix C.3 for an illustration of ϕα. There are two similarities between (11) and ϕα. First, both proposals interpolate between the log and linear losses: when α 0+, ϕα approaches the log loss, and when α = 1, ϕα equals the linear loss. Second, both proposals modify the base loss, allowing the link F to be chosen independently. In particular, one may use the heavy-tailed F of Ding & Vishwanathan (2010); Amid et al. (2019b) in conjunction with our partially Huberised loss.

One difference between (11) and ϕα is that the partially Huberised loss is exactly linear for a suitable region; consequently, it is guaranteed to be Lipschitz, unlike ϕα. This can be understood in terms of the loss gradients: for a class-probability estimate pθ(x, y), let ℓθ(x, y) .= ϕ(pθ(x, y)). For the generalised cross-entropy, ℓθ(x, y) = pθ(x, y)α 1 pθ(x, y), while for the partially Huberised loss, ℓθ(x, y) = (pθ(x, y) τ) 1 pθ(x, y). Both gradients thus take into account whether a sample is informative , in the sense of being poorly-predicted (pθ(x, y) 0). Further, to guard against such samples being the result of label noise, both ensure this influence is not overwhelming, but in different ways: the generalised cross-entropy softens the influence, while still allowing it to be unbounded as pθ(x, y) 0. On the other hand, the partially Huberised loss enforces a hard cap of τ 1 on the influence. This is to be contrast with a truncated loss also considered in Zhang & Sabuncu (2018), which enforces a hard cap on the loss, thus completely discarding poorly-predicted instances.

As shown in Figure 1a, the partially Huberised loss in (10) exhibits saturating behaviour at the extremes. This by itself is hardly new in the context of robust statistics. For example, Tukey s biweight loss replaces the Huber loss linear behaviour with hard saturation (Gross & Tukey,

1973), and is a member of the broader family of redescending M-estimators (Andrews et al., 1972;

Shevlyakov et al., 2008). In a similar spirit, Masnadi-Shirazi & Vasconcelos (2009) proposed the bounded Savage loss ϕ(v) = (σ(2v) 1)2, while Yang et al. (2010); Menon & Williamson (2018) theoretically studied loss clipping. Robust versions of logistic regression derived by variants of the Huber loss have been explored, e.g., Pregibon (1982, Section 4.1), Copas (1988, Section 3.1), (Feng et al., 2014).

The distinction of the above proposals to our work is two-fold. First, we explicitly consider the classification setting, wherein we show consistency and robustness to label noise for the partially Huberised loss. Second, our proposal explicitly highlights the connection to gradient clipping.

Finally, while our focus has been on label noise, there are other notions of robustness . Indeed, several works have proposed means of making gradient descent robust to heavy-tailed distributions

Published as a conference paper at ICLR 2020

Clipping type Clipped gradient form Equivalent loss Reference Label noise robust?

Gradient clipτ ( m(θ) φ (m(θ))) Equation 4 Equation 6 (Proposition 4) L-Gradient m(θ) clipτ (φ (m(θ))) Huberised Equation 7 (Proposition 4) CL-Gradient p(θ) clipτ (ϕ (p(θ))) Partially Huberised Equation 9 (Proposition 7)

Table 1: Summary of types of gradient clipping considered in this paper. We consider binary classification problems involving a labelled example (x, y), parametrised scoring function sθ(x) with margin m(θ) .= y sθ(x), and differentiable composite margin loss φ(z). This loss internally converts scores to probabilities p(θ) .= F(m(θ)) for link function F( ), which is evaluated with some base loss ϕ; i.e., φ = ϕ F. Gradient clipping applies to the full loss, i.e., φ(m(θ)). L-gradient clipping applies only to the composite loss, leaving the score untouched; this is equivalent to using a Huberised loss. CL-gradient clipping applies only to the base loss, leaving the link untouched; this is equivalent to using a partially Huberised loss. Only the latter has robustness guarantee under symmetric label noise.

or outliers (Brownlees et al., 2015; Hsu & Sabato, 2016; Lugosi & Mendelson, 2016; Lecué et al., 2018; Prasad et al., 2018; Holland & Ikeda, 2019), typically by passing the gradient through a robust operator (Tukey, 1960; Hampel et al., 1986; Rousseeuw & Leroy, 1987; Lerasle & Oliveira, 2011; Catoni, 2012; Minsker, 2015). By contrast, we analyse the (simpler) gradient clipping approach, showing that while it is inadequate to cope with label noise, a simple modification rectifies this. Since our proposal only modifies the underlying loss, it may be combined with any of these methods.

4.4 DISCUSSION AND IMPLICATIONS

Table 1 summarises our results, highlighting the perspective of gradient clipping as equivalently modifying the loss. Before proceeding, we make some qualifying comments. First, our analysis has assumed symmetric label noise. Often, one encounters asymmetric or instance-dependent noise (Menon et al., 2018). While corresponding guarantees for the linear loss may be ported over to the partially Huberised loss, they require stronger distributional assumptions (Ghosh et al., 2015).

Second, Proposition 4 exhibits a specific distribution which defeats the Huberised loss under linear models. In practice, distributions may be more benign (Patrini et al., 2016), and models are often nonlinear, meaning that Huberised losses (and thus gradient clipping) are thus unlikely to succumb as extremely to label noise as Proposition 4 suggests. The aim of 4 is however to establish that a simple modification of clipping avoids worst-case degradation, without adding significant complexity.

Third, for minibatch size N > 1, the effect of clipping is not a simple loss modification, since the loss gradients for each sample will be modified by the entire minibatch loss gradient norm g(θ) 2. Since this minibatch is randomly drawn, one cannot mimic gradient clipping by a simple modification of the loss function. However, the results for N = 1 suffice to establish that gradient clipping cannot in general endow robustness. One may use our partially Huberised loss in conjunction with minibatch gradient clipping, to potentially obtain both robustness and optimisation benefits.

Finally, partially Huberised losses such as (12) require setting a hyperparameter τ (e.g., by crossvalidation), similar to α in the generalised cross-entropy per 4.3. Intuitively, the optimal τ trades off the noise-robustness of the linear loss, and the gradient informativeness of the base loss (per the discussion in 4.3). Setting τ to be large tacitly assumes that one s samples are largely noise-free.

5 EXPERIMENTAL ILLUSTRATION

We now present experiments illustrating that: (a) we may exhibit label noise scenarios that defeat a Huberised but not partially Huberised loss, confirming Propositions 4, 7, and (b) partially Huberised versions of existing losses perform well on real-world datasets subject to label noise.

Synthetic datasets. Our first experiments involve two synthetic datasets, which control for confounding factors. We begin with a setting from Long & Servedio (2010), comprising a 2D linearly separable distribution. (See Appendix E.1 for an illustration.) We draw N = 1, 000 random samples from this distribution, and flip each label with ρ = 45% probability. We train a linear classifier to minimise one of several losses, and evaluate the classifier s accuracy on 500 clean test samples.

Published as a conference paper at ICLR 2020

Partial Huberised

Test accuracy

(a) Symmetric label noise.

2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 θ

Logistic Huber Partial Huber

(b) Outliers in feature space.

Figure 2: Results on synthetic data. In (b), the solid (dashed) curve denotes empirical risk without (with) outliers.

Dataset Loss function ρ = 0.0 ρ = 0.2 ρ = 0.4 ρ = 0.6

CE 98.7 0.0 97.7 0.2 96.4 0.3 91.1 0.6 CE + clipping 98.6 0.1 97.5 0.1 96.8 0.1 88.9 0.8 Linear 98.3 0.0 88.2 0.1 78.8 0.0 9.6 0.1 GCE 98.8 0.1 98.6 0.1 98.2 0.0 97.6 0.1 PHuber-CE τ = 10 98.5 0.0 98.7 0.0 98.5 0.1 97.6 0.0 PHuber-GCE τ = 10 98.6 0.1 98.5 0.1 98.4 0.1 97.8 0.0

CE 92.4 0.1 72.4 0.2 61.4 0.8 40.6 0.3 CE + clipping 88.3 0.3 68.7 0.2 56.7 0.6 44.2 2.1 Linear 91.3 0.1 88.2 0.1 83.8 0.2 69.3 0.2 GCE 91.7 0.0 88.4 0.2 80.8 0.3 60.0 0.3 PHuber-CE τ = 2 91.6 0.1 88.6 0.0 83.6 0.1 72.2 0.0 PHuber-GCE τ = 10 92.0 0.1 88.5 0.1 80.8 0.1 62.6 0.2

CE 66.6 1.4 49.7 0.3 29.9 0.9 11.4 0.2 CE + clipping 28.8 0.1 20.6 0.4 14.7 0.6 9.0 0.4 Linear 12.1 1.6 6.6 1.2 5.7 0.9 3.6 0.1 GCE 70.1 0.1 63.9 0.1 52.0 0.2 29.9 0.5 PHuber-CE τ = 10 66.2 1.5 56.2 2.2 44.4 0.7 18.5 0.4 PHuber-GCE τ = 10 69.8 0.2 64.4 0.4 52.4 0.2 31.5 0.8

Table 2: Test set accuracy where the training labels are corrupted with probability ρ. We report the mean and standard error over 3 trials. The highlighted cells are the best performing loss at a given ρ. PHuber here refers to our partial Huberisation from 4, which is equivalent to a variant of gradient clipping.

We compare the logistic loss with its Huberised and partially Huberised versions, using τ = 1 and τ = 2 respectively. Figure 2a presents the results over 500 independent trials. The logistic loss and its Huberised counterpart suffer significantly under noise, while the partially Huberised loss often achieves perfect near-discrimination. This confirms that in the worst case, L-gradient clipping may succumb to noise, while CL-gradient clipping performs well in the same scenario.

We next consider a 1D setting based on Ding (2013, Section 3.2.3), comprising 10, 000 linearly separable inliers and 50 outliers . Assuming the use of a linear model parameterised by scalar θ R, we plot the empirical risk of the samples with and without the outlier observations as θ is varied. Figure 2b shows that the logistic loss and its Huberised variant are strongly affected by the outliers: their optimal solution goes from θ = + to θ = 0. However, the partially Huberised loss is largely immune to the outliers. Appendix E contains additional details on this experiment.

Real-world datasets. We now demonstrate that partially Huberised losses perform well with deep neural networks trained on MNIST, CIFAR-10 and CIFAR-100 (Krizhevsky & Hinton, 2009). For MNIST, we train a Le Net (Lecun et al., 1998) using Adam with batch size N = 32, and weight decay

Published as a conference paper at ICLR 2020

of 10 3. For CIFAR-10 and CIFAR-100, we train a Res Net-50 (He et al., 2016) using SGD with momentum 0.1, weight decay of 5 10 3, batch normalisation, and N = 64, 128 respectively.

For each dataset, we corrupt the training labels with symmetric noise at flip probability ρ {0.0, 0.2, 0.4, 0.6}. We compare the test set accuracy of various losses combined with a softmax link. Our baseline is the cross-entropy loss (CE). As representative noise-robust losses, we consider the linear or unhinged loss (van Rooyen et al., 2015; Ghosh et al., 2017), and the generalised cross-entropy (GCE) with α = 0.7, following Zhang & Sabuncu (2018). We additionally assess global gradient clipping (with τ = 0.1) of the CE, which per 3 is akin to a Huberised loss.

We apply our partial Huberisation of (10) to the CE (12) ( PHuber-CE ), and the GCE ( PHuber GCE ). The latter highlights that partial Huberisation is not tied to the cross-entropy, and is applicable even on top of existing noise-robust losses. Recall that partial Huberisation offers a choice of tuning parameter τ, similar to the α parameter in GCE, and noise-rate estimates in loss-correction techniques more generally. For each dataset, we pick τ {2, 10} (equivalently corresponding to probability thresholds 0.5 and 0.1 respectively) so as to maximise accuracy on a validation set of noisy samples with the maximal noise rate ρ = 0.6; the chosen value of τ was then used for each noise level. Tuning τ separately for each setting of the noise rate ρ can be expected to help performance, at the expense of increased computational cost. Recall also that as τ 1, partial Huberisation mimics using the base loss, while as τ + , partial Huberisation mimics using the linear loss; our hypothesis is that an intermediate τ can attain a suitable balance between noise robustness, and gradient informativeness.

Table 2 shows that in the noise-free case (ρ = 0.0), all methods perform comparably. However, when injecting noise, accuracy for the CE degrades dramatically. Further, gradient clipping sometimes offers improvements under high noise; however, the performance is far inferior to other losses, which is in keeping with their robustness guarantees. Indeed, the linear loss, which is provably robust to symmetric noise, generally performs well even when ρ = 0.6. However, optimisation under this loss is more challenging, since the gradient does not account for instances importance (per 4.3). This is particularly reflected on the CIFAR-100 dataset, where this loss suffers to learn even under no noise. The GCE and partially Huberised losses do not suffer from this issue, even at high noise levels.

Generally, the partially Huberised losses are competitive with or improve upon the counterpart losses they build upon. In particular, the partially Huberised CE performs much better than the CE under high noise, while the partially Huberised GCE slightly bumps up the GCE numbers on CIFAR-100. This indicates that partially Huberised may be useful in combining with generic base losses to cope with noise. We reiterate here that our partially Huberised loss may be used in conjunction with other ideas, e.g., pruning (Zhang & Sabuncu, 2018), consensus (Malach & Shalev-Shwartz, 2017; Han et al., 2018), or abstention (Thulasidasan et al., 2019). We leave such exploration for future work.

6 CONCLUSION AND FUTURE WORK

We established that gradient clipping by itself does not suffice to endow even simple models with label noise robustness; however, a simple variant resolves this issue. Experiments confirm that our composite loss-based gradient clipping performs well on datasets corrupted with label noise. One interesting direction for future work is to analyse the behaviour of gradient-clipping inspired losses for the more general problem of distributionally robust learning (Shafieezadeh-Abadeh et al., 2015; Namkoong & Duchi, 2016; 2017; Hu et al., 2018; Sinha et al., 2018; Duchi & Namkoong, 2019).

Martin Abadi, Andy Chu, Ian Goodfellow, H. Brendan Mc Mahan, Ilya Mironov, Kunal Talwar, and Li Zhang. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, CCS 16, pp. 308 318, New York, NY, USA, 2016. ACM. ISBN 978-1-4503-4139-4.

Ehsan Amid, Manfred K. Warmuth, Rohan Anil, and Tomer Koren. Robust bi-tempered logistic loss based on Bregman divergences. Co RR, abs/1906.03361, 2019a.

Ehsan Amid, Manfred K. Warmuth, and Sriram Srinivasan. Two-temperature logistic regression based on the Tsallis divergence. In The 22nd International Conference on Artificial Intelligence and Statistics, AISTATS 2019, 16-18 April 2019, Naha, Okinawa, Japan, pp. 2388 2396, 2019b.

Published as a conference paper at ICLR 2020

D. F. Andrews, P. J. Bickel, F. R. Hampel, P. J. Huber, W. H. Rogers, and J. W. Tukey. Robust Estimates of Location. Princeton University Press, Princeton NJ, 1972.

Peter L Bartlett, Michael I Jordan, and Jon D Mc Auliffe. Convexity, classification, and risk bounds. Journal of the American Statistical Association, 101(473):138 156, 2006.

Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is difficult. Trans. Neur. Netw., 5(2):157 166, March 1994. ISSN 1045-9227.

Kush Bhatia, Prateek Jain, and Purushottam Kar. Robust regression via hard thresholding. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS 15, pp. 721 729, Cambridge, MA, USA, 2015. MIT Press.

Christian Brownlees, Emilien Joly, and Gábor Lugosi. Empirical risk minimization for heavy-tailed losses. Ann. Statist., 43(6):2507 2536, 12 2015.

Andreas Buja, Werner Stuetzle, and Yi Shen. Loss functions for binary class probability estimation and classification: Structure and applications. Technical report, UPenn, 2005.

Olivier Catoni. Challenging the empirical mean and empirical variance: A deviation study. Ann. Inst. H. Poincaré Probab. Statist., 48(4):1148 1185, 11 2012.

Nontawat Charoenphakdee, Jongyeong Lee, and Masashi Sugiyama. On symmetric losses for learning from corrupted labels. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 961 970, Long Beach, California, USA, 09 15 Jun 2019. PMLR.

Jesús Cid-Sueiro, Darío García-García, and Raúl Santos-Rodríguez. Consistency of losses for learning from weak labels. In Toon Calders, Floriana Esposito, Eyke Hüllermeier, and Rosa Meo (eds.), Machine Learning and Knowledge Discovery in Databases, pp. 197 210, Berlin, Heidelberg, 2014. Springer Berlin Heidelberg.

J. B. Copas. Binary regression models for contaminated data. Journal of the Royal Statistical Society. Series B (Methodological), 50(2):225 265, 1988. ISSN 00359246.

Nan Ding. Statistical machine learning in the t-exponential family of distributions. Ph D thesis, Purdue University, 2013.

Nan Ding and S.V.N. Vishwanathan. t-logistic regression. In J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta (eds.), Advances in Neural Information Processing Systems 23, pp. 514 522. Curran Associates, Inc., 2010.

John Duchi and Hongseok Namkoong. Variance-based regularization with convex objectives. Journal of Machine Learning Research, 20(68):1 55, 2019.

Cynthia Dwork, Frank Mc Sherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In Proceedings of the Third Conference on Theory of Cryptography, TCC 06, pp. 265 284, Berlin, Heidelberg, 2006. Springer-Verlag. ISBN 3-540-32731-2, 978-3-540-327318.

Anders Ekholm and Juni Palmgren. A model for a binary response with misclassifications. In Robert Gilchrist (ed.), GLIM 82: Proceedings of the International Conference on Generalised Linear Models, pp. 128 143, New York, NY, 1982. Springer New York. ISBN 978-1-4612-5771-4.

Jiashi Feng, Huan Xu, Shie Mannor, and Shuicheng Yan. Robust logistic regression and classification. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27, pp. 253 261. Curran Associates, Inc., 2014.

Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, ICML 17, pp. 1243 1252. JMLR.org, 2017.

Aritra Ghosh, Naresh Manwani, and P. S. Sastry. Making risk minimization tolerant to label noise. Neurocomputing, 160:93 107, July 2015.

Published as a conference paper at ICLR 2020

Aritra Ghosh, Himanshu Kumar, and P. S. Sastry. Robust loss functions under label noise for deep neural networks. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, AAAI 17, pp. 1919 1925. AAAI Press, 2017.

Tilmann Gneiting and Adrian E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477):359 378, 2007.

I. M. Gross and John W. Tukey. The estimators of the Princeton robustness study. Technical report, series 2 38, Princeton, 1973.

F.R. Hampel, E.M. Ronchetti, P.J. Rousseeuw, and W.A. Stahel. Robust Statistics: the Approach Based on Influence Functions. Wiley New York, 1986.

Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor W. Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS 18, pp. 8536 8546, USA, 2018. Curran Associates Inc.

Elad Hazan, Kfir Y. Levy, and Shai Shalev-Shwartz. Beyond convexity: Stochastic quasi-convex optimization. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS 15, pp. 1594 1602, Cambridge, MA, USA, 2015. MIT Press.

K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770 778, June 2016.

Matthew Holland and Kazushi Ikeda. Better generalization with less data using robust gradient descent. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 2761 2770, Long Beach, California, USA, 09 15 Jun 2019. PMLR.

Daniel Hsu and Sivan Sabato. Loss minimization and parameter estimation with heavy tails. Journal of Machine Learning Research, 17(18):1 40, 2016.

Weihua Hu, Gang Niu, Issei Sato, and Masashi Sugiyama. Does distributionally robust supervised learning give robust classifiers? In Jennifer Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp. 2029 2037, Stockholmsmässan, Stockholm Sweden, 10 15 Jul 2018. PMLR.

Peter J. Huber. Robust estimation of a location parameter. Ann. Math. Statist., 35(1):73 101, 03 1964.

Peter J. Huber. Robust Statistics. John Wiley & Sons, 1981.

Hung Hung, Zhi-Yu Jou, and Su-Yun Huang. Robust mislabel logistic regression without modeling mislabel probabilities. Biometrics, 74(1):145 154, 2018.

Julian Katz-Samuels, Gilles Blanchard, and Clayton Scott. Decontamination of mutual contamination models. Journal of Machine Learning Research, 20(41):1 57, 2019.

A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Master s thesis, Department of Computer Science, University of Toronto, 2009.

Guillaume Lecué, Matthieu Lerasle, and Timothée Mathieu. Robust classification via MOM minimization. ar Xiv e-prints, art. ar Xiv:1808.03106, Aug 2018.

Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278 2324, Nov 1998.

M. Lerasle and R. I. Oliveira. Robust empirical mean Estimators. ar Xiv e-prints, art. ar Xiv:1112.3914, Dec 2011.

Kfir Y. Levy. The power of normalization: Faster evasion of saddle points. Co RR, abs/1611.04831, 2016.

Published as a conference paper at ICLR 2020

Tongliang Liu and Dacheng Tao. Classification with noisy labels by importance reweighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3):447 461, 2016.

Philip M. Long and Rocco A. Servedio. Random classification noise defeats all convex potential boosters. Machine Learning, 78(3):287 304, 2010. ISSN 0885-6125.

Gabor Lugosi and Shahar Mendelson. Risk minimization by median-of-means tournaments. ar Xiv e-prints, art. ar Xiv:1608.00757, Aug 2016.

Eran Malach and Shai Shalev-Shwartz. Decoupling "when to update" from "how to update". In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS 17, pp. 961 971, USA, 2017. Curran Associates Inc. ISBN 978-1-5108-6096-4.

R. A. Maronna, D. Martin, and V. J. Yohai. Robust Statistics: Theory and Methods. John Wiley & Sons, New York, 2nd edition, 2019.

Hamed Masnadi-Shirazi and Nuno Vasconcelos. On the design of loss functions for classification: theory, robustness to outliers, and Savage Boost. In D. Koller, D. Schuurmans, Y. Bengio, and L. Bottou (eds.), Advances in Neural Information Processing Systems 21, pp. 1049 1056. Curran Associates, Inc., 2009.

Aditya Krishna Menon and Robert C. Williamson. A loss framework for calibrated anomaly detection. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett (eds.), Advances in Neural Information Processing Systems 31, pp. 1487 1497. Curran Associates, Inc., 2018.

Aditya Krishna Menon, Brendan van Rooyen, Cheng Soon Ong, and Bob Williamson. Learning from corrupted binary labels via class-probability estimation. In International Conference on Machine Learning (ICML), pp. 125 134, 2015.

Aditya Krishna Menon, Brendan van Rooyen, and Nagarajan Natarajan. Learning from binary labels with instance-dependent noise. Machine Learning, 107(8-10):1561 1595, 2018.

Stephen Merity, Nitish Shirish Keskar, and Richard Socher. Regularizing and optimizing LSTM language models. In International Conference on Learning Representations, 2018.

Tomáš Mikolov. Statistical Language Models Based on Neural Networks. Ph D thesis, Brno University of Technology, 2012.

Stanislav Minsker. Geometric median and robust estimation in banach spaces. Bernoulli, 21(4): 2308 2335, 11 2015.

Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529 533, 02 2015.

H. Namkoong and J. Duchi. Variance regularization with convex objectives. In Advances in Neural Information Processing Systems (Neur IPS), 2017.

Hongseok Namkoong and John C Duchi. Stochastic gradient methods for distributionally robust optimization with f-divergences. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp. 2208 2216. Curran Associates, Inc., 2016.

Nagarajan Natarajan, Inderjit S Dhillon, Pradeep D Ravikumar, and Ambuj Tewari. Learning with noisy labels. In Advances in Neural Information Processing Systems (NIPS), pp. 1196 1204, 2013.

Nagarajan Natarajan, Inderjit S. Dhillon, Pradeep Ravikumar, and Ambuj Tewari. Cost-sensitive learning with noisy labels. Journal of Machine Learning Research, 18(155):1 33, 2018.

Richard Nock and Frank Nielsen. On the efficient minimization of classification calibrated surrogates. In D. Koller, D. Schuurmans, Y. Bengio, and L. Bottou (eds.), Advances in Neural Information Processing Systems 21, pp. 1201 1208. Curran Associates, Inc., 2009.

Published as a conference paper at ICLR 2020

Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. Understanding the exploding gradient problem. Co RR, abs/1211.5063, 2012.

Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In Proceedings of the 30th International Conference on International Conference on Machine Learning - Volume 28, ICML 13. JMLR.org, 2013.

Giorgio Patrini, Frank Nielsen, Richard Nock, and Marcello Carioni. Loss factorization, weakly supervised learning and label noise robustness. In International Conference on Machine Learning (ICML), pp. 708 717, 2016.

Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: a loss correction approach. In Computer Vision and Pattern Recognition (CVPR), pp. 2233 2241, 2017.

Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pp. 2227 2237, New Orleans, Louisiana, June 2018. Association for Computational Linguistics.

Venkatadheeraj Pichapati, Ananda Theertha Suresh, Felix X. Yu, Sashank J. Reddi, and Sanjiv Kumar. Ada Cli P: Adaptive Clipping for Private SGD. ar Xiv e-prints, art. ar Xiv:1908.07643, Aug 2019.

Adarsh Prasad, Arun Sai Suggala, Sivaraman Balakrishnan, and Pradeep Ravikumar. Robust Estimation via Robust Gradient Estimation. ar Xiv e-prints, art. ar Xiv:1802.06485, Feb 2018.

Daryl Pregibon. Resistant fits for some commonly used logistic models with medical applications. Biometrics, 38(2):485 498, 1982. ISSN 0006341X, 15410420.

Mark D. Reid and Robert C. Williamson. Composite binary losses. Journal of Machine Learning Research, 11:2387 2422, 2010.

Saharon Rosset and Ji Zhu. Piecewise linear regularized solution paths. Ann. Statist., 35(3):1012 1030, 07 2007.

Peter J. Rousseeuw and Annik M. Leroy. Robust Regression and Outlier Detection. John Wiley and Sons, 1987.

Leonard J. Savage. Elicitation of personal probabilities and expectations. Journal of the American Statistical Association, 66(336):783 801, 1971. ISSN 01621459.

Mark J. Schervish. A general method for comparing probability assessors. Ann. Statist., 17(4): 1856 1879, 12 1989.

Clayton Scott, Gilles Blanchard, and Gregory Handy. Classification with asymmetric label noise: consistency and maximal denoising. In Conference on Learning Theory (COLT), pp. 489 511, 2013.

Soroosh Shafieezadeh-Abadeh, Peyman Mohajerin Esfahani, and Daniel Kuhn. Distributionally robust logistic regression. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS 15, pp. 1576 1584, Cambridge, MA, USA, 2015. MIT Press.

Yanyao Shen and Sujay Sanghavi. Learning with bad training data via iterative trimmed loss minimization. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 5739 5748, Long Beach, California, USA, 09 15 Jun 2019. PMLR.

Georgy Shevlyakov, Stephan Morgenthaler, and Alexander Shurygin. Redescending m-estimators. Journal Of Statistical Planning And Inference, 138:2906 2917, 2008.

Aman Sinha, Hongseok Namkoong, and John C. Duchi. Certifying some distributional robustness with principled adversarial training. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings, 2018.

Published as a conference paper at ICLR 2020

Hwanjun Song, Minseok Kim, and Jae-Gil Lee. SELFIE: Refurbishing unclean samples for robust deep learning. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 5907 5915, Long Beach, California, USA, 09 15 Jun 2019. PMLR.

T. Sypherd, M. Diaz, L. Sankar, and P. Kairouz. A tunable loss function for binary classification. In 2019 IEEE International Symposium on Information Theory (ISIT), pp. 2479 2483, July 2019.

Takashi Takenouchi and Shinto Eguchi. Robustifying Ada Boost by adding the naive error rate. Neural Computation, 16(4):767 787, 2004.

Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof. Combating label noise in deep learning using abstention. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 6234 6243, Long Beach, California, USA, 09 15 Jun 2019. PMLR.

J. W. Tukey. A survey of sampling from contaminated distributions. Contributions to probability and statistics, 2:448 485, 1960.

Brendan van Rooyen and Robert C. Williamson. A theory of learning with corrupted labels. Journal of Machine Learning Research, 18(228):1 50, 2018.

Brendan van Rooyen, Aditya Krishna Menon, and Robert C. Williamson. Learning with symmetric label noise: the importance of being unhinged. In Advances in Neural Information Processing Systems (NIPS), pp. 10 18, 2015.

Robert C. Williamson, Elodie Vernet, and Mark D. Reid. Composite multiclass losses. Journal of Machine Learning Research, 17(222):1 52, 2016.

Eunho Yang, Aurélie C. Lozano, and Aleksandr Aravkin. A general family of trimmed estimators for robust high-dimensional data analysis. Electron. J. Statist., 12(2):3519 3553, 2018.

Min Yang, Linli Xu, Martha White, Dale Schuurmans, and Yao liang Yu. Relaxed clipping: A global training method for robust regression and classification. In J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta (eds.), Advances in Neural Information Processing Systems 23, pp. 2532 2540. Curran Associates, Inc., 2010.

Jingzhao Zhang, Tianxing He, Suvrit Sra, and Ali Jadbabaie. Analysis of gradient clipping and adaptive scaling with a relaxed smoothness condition. ar Xiv e-prints, art. ar Xiv:1905.11881, May 2019.

Tong Zhang. Statistical behavior and consistency of classification methods based on convex risk minimization. Annals of Statistics, 32(1):56 85, 2004.

Tong Zhang and David Johnson. A robust risk minimization based named entity recognition system. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003 - Volume 4, CONLL 03, pp. 204 207, Stroudsburg, PA, USA, 2003. Association for Computational Linguistics.

Tong Zhang, Fred Damerau, and David Johnson. Text chunking based on a generalization of winnow. J. Mach. Learn. Res., 2:615 637, March 2002. ISSN 1532-4435.

Zhilu Zhang and Mert R. Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS 18, pp. 8792 8802, USA, 2018. Curran Associates Inc.

Published as a conference paper at ICLR 2020

A PROOFS OF RESULTS

Proof of Lemma 1. Following Equation 3, ℓθ(x, y) = mθ(x, y) φ (mθ(x, y)). Similarly, for the linear loss function ℓlin(x, y; θ) .= mθ(x, y), we have ℓlin(x, y; θ) = mθ(x, y). To compute the normalised gradient, we need ℓθ(x, y) 2 = mθ(x, y) 2 |φ (mθ(x, y))|. Thus, ℓθ(x, y) ℓθ(x, y) 2 = 1 mθ(x, y) 2 mθ(x, y) φ (mθ(x, y))

|φ (mθ(x, y))|

= 1 mθ(x, y) 2 mθ(x, y) sign(φ (mθ(x, y)))

= 1 mθ(x, y) 2 ℓlin(x, y; θ),

since φ (z) < 0 by assumption that φ is admissible, and thus decreasing. Consequently, if N = 1,

g(θ) = clipτ( ℓθ(x, y)) =

( τ mθ(x,y) 2 ℓlin(x, y; θ) if |φ (mθ(x, y))| τ mθ(x,y) 2 ℓθ(x, y) else.

Assuming a linear scorer s(x; θ) = θTx, the score gradient s(x; θ) = x, which is independent of θ. The clipped gradient thus corresponds to the gradient under a Huberised loss function

( τ mθ(x,y) 2 ℓlin(x, y; θ) if |φ (mθ(x, y))| τ mθ(x,y) 2 ℓθ(x, y) else.

Proof of Lemma 2. Since φ is strictly convex and decreasing, it must be strictly decreasing. We have

clipτ (φ (z)) =

( τ φ (z) |φ (z)| if |φ (z)| τ φ (z) else

= τ sign(φ (z)) if |φ (z)| τ φ (z) else

= τ if φ (z) τ φ (z) else since φ is strictly decreasing

= τ if z (φ ) 1( τ) φ (z) else,

where (φ ) 1 exists since φ is strictly inceasing by definition of strict convexity.

Now, by definition, the Huberised loss is

φτ(z) = τ z φ ( τ) if φ (z) τ φ(z) else,

and so φ τ(z) = τ if φ (z) τ φ (z) else, which exactly equals clipτ (φ (z)). We remark here that φτ is continuous, since for any convex conjugate, φ ( u) = u (φ ) 1( u) φ((φ ) 1( u)). Plugging in u = τ, at the intersection point z0 .= (φ) 1( τ) of the two pieces of the function, φτ(z0) = τ (φ ) 1( τ) φ ( τ) = φ((φ ) 1( τ)) = φ(z0).

Published as a conference paper at ICLR 2020

Proof of Lemma 3. For admissible φ, the Huberised loss φτ of (8) is trivially convex, differentiable everywhere, and decreasing. In particular, we must have φ τ(0) < 0, and so φτ must be classification calibrated by Bartlett et al. (2006, Theorem 2). As an illustration, Figure 3 shows the minimiser of the conditional risk,

z (η) .= argmin z R

h η φτ(z) + (1 η) φτ( z) i .

Note that this quantity must be non-negative if and only if η > 1/2 for a loss to be classification calibrated. This is easily verified to be true for the Huberised logistic loss, regardless of τ.

0.0 0.2 0.4 0.6 0.8 1.0 η

τ = 0. 1 τ = 0. 5 τ = 0. 9

Figure 3: Risk minimiser for the Huberised logistic loss.

Now suppose that φ is a proper composite loss. To show that φτ is proper composite, it suffices by Williamson et al. (2016, Corollary 12) to show that φ τ (z) φ τ ( z) is strictly monotone and continuous. Continuity is immediate; for monotonicity, observe that by definition,

φ τ(z) = τ if z (φ ) 1( τ) φ (z) else.

For brevity, let z0 .= (φ ) 1( τ), where we take z0 = if τ / ran(φ ). Then,

φ τ(z) = φ (z0) if z z0 φ (z) else.

φ τ( z) = φ (z0) if z z0 φ ( z) else. Thus, the quantity of interest is

φ τ(z) φ τ( z) =

φ (z0) φ ( z) if z z0 φ (z) φ ( z) if z0 z z0 1 if z0 z z0 φ (z) φ (z0) if z z0.

Since φ is strictly convex, φ is strictly decreasing and thus invertible. Thus, the ratio φ τ (z) φ τ ( z) is invertible, provided z0 0, i.e., τ φ (0). Consequently, φτ is proper composite when τ φ (0). To intuit the need for the restriction on τ, observe that by (Reid & Williamson, 2010, Corollary 12) the tentative link function for the loss is

1 φ (z) φ ( z)

φ ( z) φ (z0)+φ ( z) if z z0 F(z) if z0 z z0 1 2 if z0 z z0 φ (z0) φ (z0)+φ (z) if z z0.

Published as a conference paper at ICLR 2020

When τ < φ (0), the above is seen to be non-invertible. See also Appendix C for an illustration of this link function for the logistic loss.

Proof of Proposition 4. In order to apply Long & Servedio (2010, Theorem 2), we simply need to check that the loss φτ is a convex potential in the sense of Long & Servedio (2010, Definition 1). This requires that φτ is convex, non-increasing, continuously differentiable, and asymptotes to zero (or equally, is bounded from below). Each of these is satisfied by assumption of φ being admissible.

Proof of Lemma 5. By Lemma 2, we may write clipτ (ϕ (u)) as the derivative of a partially Huberised base loss ϕτ given by

ϕτ(u) .= τ u ϕ ( τ) if ϕ (u) τ ϕ(u) else

This induces a composite margin loss φτ .= ϕτ F. Now let pθ(x, y) = F(mθ(x, y)), and define ℓτ(x, y; θ) .= φτ(mθ(x, y)). The gradient under this loss is

θ ℓτ(x, y; θ) = θ φτ(mθ(x, y)) = θ ϕτ(pθ(x, y))

= θpθ(x, y) ϕ τ(pθ(x, y))

= θpθ(x, y) clipτ (ϕ (pθ(x, y))) = cl clipτ( ℓθ(x, y)).

Thus, CL-gradient clipping is equivalent to using the loss ℓτ.

Proof of Lemma 6. We proceed in a similar manner to Lemma 3: to show that the loss is proper, we must establish that φ τ (z) φ τ ( z) is invertible. By Reid & Williamson (2010, Corollary 14), for the margin loss φ to be proper composite with link F, it must be true that F satisfies the symmetry condition F( z) = 1 F(z). We thus have

φ τ(z) = F (z) ϕ τ(F(z)) φ τ( z) = F ( z) ϕ τ(F( z))

= F (z) ϕ (1 F(z)).

Consequently,

φ τ(z) φ τ( z) = ϕ τ(F(z)) ϕ τ(1 F(z)).

Since F is invertible by assumption, the above is invertible if and only if ϕ τ (u) ϕ τ (1 u) is invertible. Observe that

ϕ τ(u) = τ if u (ϕ ) 1( τ) ϕ (u) else.

Let u0 .= (ϕ ) 1( τ) for brevity. Then,

ϕ τ(u) ϕ τ(1 u) =

τ ϕ (1 u) if u u0 (1 u0)

ϕ (u) ϕ (1 u) if u0 u 1 u0 1 if 1 u0 u u0 ϕ (u)

τ if u u0 (1 u0)

This quantity is invertible provided u0 1

2, i.e., τ ϕ (1/2). A subtlety, however, is that the above does not necessarily span the entire range of [0, + ]; consequently, ϕτ itself is proper composite, with a link function of its own.

Even when τ is small, one may verify that the loss φτ is nonetheless classification calibrated: this is because for any η [0, 1], the minimiser z (η) of the conditional risk

Published as a conference paper at ICLR 2020

must satisfy the stationarity condition

φ τ(z ) φ τ( z ) = η 1 η .

We thus need to find a suitable z such that the left hand side equates to a given constant. Now, for any C = 1 there is a unique u such that ϕ τ (u) ϕ τ (1 u) = C. Thus, provided η = 1

2, there is a unique

z = F 1(u ) such that φ τ (F (z )) φ τ (1 F (z )) = C. One may verify that this z > 0 η > 1

2; for example, see Figure 4, which visualises the risk minimiser for various values of τ. Thus, the sign of the minimising score conveys whether or not the positive class-probability is dominant; thus, the loss is classification calibrated.

0.0 0.2 0.4 0.6 0.8 1.0 η

τ = 100. 0 τ = 2. 0 τ = 1. 1

Figure 4: Risk minimiser for the partially Huberised logistic loss.

Proof of Proposition 7. Let Rτ(f) denote the risk on the clean distribution of a predictor f with respect to the partially Huberised loss ϕτ with parameter τ. Similarly, let Rτ(f) denote the risk on the noisy distribution. Following Ghosh et al. (2015); van Rooyen et al. (2015), we have Rτ(f) = (1 2 ρ) Rτ(f) + α E x [ ϕτ(f(x)) + ϕτ(1 f(x))] .

That is, the risk on the noisy distribution equals a scaled version of the risk on the clean distribution, plus an additional term. This term is a constant independent of f if and only if ϕτ satisfies the symmetry condition of Ghosh et al. (2015) , namely, ϕτ(u) + ϕτ(1 u) = C for some constant C.

Even when the symmetry condition does not hold, one may nonetheless aim to bound this additional term as follows. For simplicity, we restrict attention here to ϕ being the log or cross-entropy loss ϕ(u) = log u. By definition, for any f (0, 1),

ϕτ(f) + ϕτ(1 f) =

τ f + log τ + 1 log(1 f) if p 1

τ f τ (1 f) + 2 log τ + 2 if 1 1

τ log f log(1 f) if 1

log f τ (1 f) + log τ + 1 if p 1

τ f + log τ + 1 log(1 f) if p 1

τ + 2 log τ + 2 if 1 1

τ log f log(1 f) if 1

log f τ (1 f) + log τ + 1 if p 1

Evidently, all piecewise functions involved are bounded on the respective intervals. For example,

ϕτ(f) + ϕτ(1 f)

log τ + 1 log 1

τ + 2 log τ + 2 if 1 1

τ + log τ + 1 if p 1

Published as a conference paper at ICLR 2020

By taking the maximum of each of the quantities on the right hand side which are constants depending on τ we may thus find constants C1, C2 such that

Rτ(f) (1 2 ρ) Rτ(f) + C1 Rτ(f) (1 2 ρ) Rτ(f) + C2. Now let f denote the minimiser of the clean risk Rτ(f), and f the minimiser of the noisy risk Rτ(f). Then, using each of the above inequalities,

Rτ(f ) Rτ( f ) (1 2 ρ) (Rτ(f ) Rτ( f )) + C1 C2 (C1 C2),

where the last inequality is because Rτ(f ) Rτ( f ) by definition of f . The claim follows.

Published as a conference paper at ICLR 2020

B PROPER AND PROPER COMPOSITE LOSSES

Beyond requiring classification-calibration, it is often desirable to use classifier outputs as valid probabilities. Proper losses (Savage, 1971; Schervish, 1989; Buja et al., 2005) ϕ: { 1} [0, 1] R+ are the core losses of such class-probability estimation tasks, for which

( p [0, 1]) argmin u [0,1] E y Bern(p) [ϕ(y, u)] = p. (13)

Equation 13 stipulates that when using ϕ to distinguish positive and negative labels, it is optimal to predict the positive class-probability.

Typically, it is more useful to work with losses that accept real-valued scores, e.g. as output by the pre-activation of the final layer of a neural network. To this end, proper composite losses (Reid & Williamson, 2010; Nock & Nielsen, 2009) are those ℓ: { 1} R R+ are such that ℓ F 1 is proper, for invertible link function F : R [0, 1]. Proper composite losses are inherently classificationcalibrated (Reid & Williamson, 2010, Theorem 16). Canonical examples are the logistic loss ℓ(y, v) = log(1 + e yv) with F : v 7 σ(v), and exponential loss ℓ(y, v) = e yv with F : v 7 σ(2 v). When ℓis differentiable, we have (Reid & Williamson, 2010, Corollary 12)

( v R) F(v) = 1 ℓ (+1, v)

Given a proper loss ϕ and symmetric link F with F( v) = 1 F(v), the loss ℓ(y, v) = ϕ(y, F(v)) defines a margin loss (Reid & Williamson, 2010, Corollary 14).

Proper composite losses may also be extended to multiclass settings in the natural way (Gneiting & Raftery, 2007; Williamson et al., 2016): one now defines a proper loss ϕ: [K] K R+, where K is the number of classes and K denotes the K-simplex. A proper composite loss may be defined using a link F : RK K, such as the softmax operator Fk(v) = exp(vk) P

k exp(vk ). Combined with the log-loss ϕ(y, p) = log py, this yields the standard softmax cross-entropy loss.

Published as a conference paper at ICLR 2020

C ILLUSTRATION OF LOSSES

We illustrate the Huberised, partially Huberised, and generalised cross-entropy losses as their underlying tuning parameters are varied. Additionally, we illustrate the link functions that are implicit in each of the losses, which illustrates that they may be non-invertible if τ is too large.

C.1 HUBERISED LOSS

Figure 5 illustrates the Huberised version of the logistic loss, and its derivative. Following the proof of Lemma 3, for τ (0, 1) and z0 = σ 1(τ), the Huberised loss φτ has an implicit link function (see Figure 6)

( z R) Fτ(z) =

σ(z) σ(z0)+σ(z) if z z0 σ(z) if z0 z z0 1 2 if z0 z z0 σ(z0) σ(z0)+σ(z) if z z0.

Compared to the standard sigmoid, the Huberised link saturates more slowly as τ is decreased. Note that when τ 1

2, the link function is not invertible everywhere: this results in the loss not being proper composite per our definition.

C.2 PARTIALLY HUBERISED LOSS

Figure 7 illustrates the partially Huberised version of the logistic loss, as well as the base log-loss. Following the proof of Lemma 6, the partially Huberised loss φτ has implicit link function

1 1+τ σ( v) if σ(v) 1

τ τ σ(v) 1+τ σ(v) if σ(v) 1 1

Figure 9 illustrates the link function. The logistic loss has ϕ(u) = log u, and so ϕ (1/2) = 2. For τ = 1, the link function will be non-invertible everywhere, which is expected since the loss here is the linear loss, which is not suitable for class-probability estimation. For τ (1, 2), the link function will be invertible for p / 1 1

τ . Intuitively, the case τ (1, 2) corresponds to the linear regions of the losses on positive and negative instances crossing over. For τ 2, the link function will always be invertible.

It may be observed that partial Huberisation causes the link function to saturate at values [1/(1 + τ), τ/(1+τ)]: this does not affect classification calibration, but does imply that rescaling is necessary in order to intepret the output probabilities.

C.3 ILLUSTRATION OF GENERALISED CROSS-ENTROPY

Figure 8 illustrates the base ϕα loss, and its composition with a sigmoid link function.

Published as a conference paper at ICLR 2020

φτ(v) τ = 1

v φ τ(v) τ = 1

(b) Derivative.

Figure 5: Illustration of Huberised logistic loss. The effect of Huberisation is to linearise the loss beyond a certain point, or equally, to cap the derivative when its magnitude exceeds a certain threshold.

Figure 6: Illustration of link function for the Huberised logistic loss. The effect of Huberisation is to make the link function saturate slower. Note that when τ 1

2, the link function is not invertible everywhere.

0.2 0.4 0.6 0.8

3 τ = 2 τ = 4 τ =

(a) Log-loss.

3 τ = 2 τ = 4 τ =

(b) Log-loss with sigmoid link.

Figure 7: Illustration of partial Huberisation, in terms of the underlying proper loss and its composite form.

Published as a conference paper at ICLR 2020

0.2 0.4 0.6 0.8

α = 0 α = 0.1 α = 0.5 α = 0.9

(a) Base loss.

α = 0 α = 0.1 α = 0.5 α = 0.9

(b) Composite loss with sigmoid link.

Figure 8: Illustration of generalised cross-entropy loss for various choices of α [0, 1).

3 τ = 2 τ = 4 τ =

Figure 9: Illustration of link function for partially Huberised logistic loss. The effect of partial Huberisation is to make the link function saturate slower. When τ < 2, the link function is not invertible everywhere.

Published as a conference paper at ICLR 2020

D ROBUSTNESS TO INSTANCE OUTLIERS

Following Ding (2013, Section 3.2.2), we consider robustness to an arbitrary corruption under a linear scoring model. For margin loss ℓ(x, y; θ) .= φ(y θTx) with empirical risk minimiser ˆθN on a sample {(xn, yn)}N n=1, suppose that the sample is corrupted with an outlier (x , y ). One would like to ensure that ˆθN is not swayed by making x 2 arbitrarily large: that is,

lim x 2 + 1 N + 1

n=1 ℓ(xn, yn; ˆθN) + ℓ(x , y ; ˆθN) = 0,

or equivalently, lim x 2 + ℓ(x , y ; ˆθN) = 0. Fortunately, the saturating behaviour of the partially Huberised loss affords this, as we now show. Proposition 8. Pick any convex, differentiable, proper composite margin loss φ, whose link F satisfies limz z F (z) = 0. For any τ > 0, let ℓτ(x, y; θ) .= φτ(y θTx) be the τ-partially Huberised loss under a linear model. Then, for any (x, y) and θ such that θ 2 < + and θTx = 0,

lim x 2 + ℓτ(x, y; θ) 2 = 0.

Proof of Proposition 8. By definition,

ℓτ(x, y; θ) = y x φ (y θTx).

ℓτ(x, y; θ) 2 = x 2 |φ (y θTx)|

= (y θTx) |φ (y θTx)| 1 y θ 2 cos ψx,θ ,

where ψx,θ denotes the angle between x and θ. If θTx = 0, then cos ψx,θ = 0, and so the third term is finite. Thus, ℓτ(x, y; θ) 2 limz |z| φ τ(z), depending on the sign of y θTx. By definition of φτ, the derivative of the loss asymptotes to either F (v), or φ (v). Now, limz z F (v) = 0 by assumption, and limz + z φ (v) = 0 since φ is convex, the claim is shown.

Published as a conference paper at ICLR 2020

E ADDITIONAL SYNTHETIC EXPERIMENTS

We provide some more details regarding the synthetic data used in the body.

E.1 DETAILS OF LONG & SERVEDIO (2010) DATASET

The problem considered in Long & Servedio (2010) (see Figure 10) comprises a distribution concentrated on six atoms { (1, 0), (γ, 5γ), (γ, γ)} R2 for some γ > 0; we chose γ = 1 24. An instance (x1, x2) is labelled as y = Jx1 0K. The instances are weighted so that the first four atoms have probability mass 1

8, and the last two atoms mass 1

4. We modify this distribution slightly by treating the atoms as means of isotropic Gaussians, and treating the marginal distribution over instances to be a mixture of these Gaussians with mixing weights given by the corresponding probability masses of the atoms.

E.2 DETAILS OF OUTLIER DATASET

For the experiment involving outliers in feature space, the data comprises points on the line. Positively labelled samples are drawn from a unit variance Gaussian centered at (1, 1), with positively labeled outliers drawn from ( 200, 1). Negatively labelled samples comprise the negation of all points. We learn an unregularised linear classifier from this data, which corresponds to a single scalar θ.

Published as a conference paper at ICLR 2020

1 0.5 0.5 1

Figure 10: Distribution proposed in Long & Servedio (2010), which defeats any member of a broad family of convex losses. The data comprises six points, with the blue points labelled positive, and the red points labelled negative. The two fat points have twice as much probability mass as their thin counterparts. While the dataset is trivially linearly separable, minimising a broad range of convex losses with a linear model under any non-zero amount of symmetric label noise results in a predictor that is tantamount to random guessing.