Many adaptive optimization methods have been proposed and used in deep learning, in which Adam is regarded as the default algorithm and widely used in many deep learning frameworks. Recently, many variants of Adam, such as Adabound, RAdam and Adabelief, have been proposed and show better performance than Adam. However, these variants mainly focus on changing the stepsize by making differences on the gradient or the square of it. Motivated by the fact that suitable damping is important for the success of powerful second-order optimizers, we discuss the impact of the constant $\epsilon$ for Adam in this paper. Surprisingly, we can obtain better performance than Adam simply changing the position of $\epsilon$. Based on this finding, we propose a new variant of Adam called EAdam, which doesn't need extra hyper-parameters or computational costs. We also discuss the relationships and differences between our method and Adam. Finally, we conduct extensive experiments on various popular tasks and models. Experimental results show that our method can bring significant improvement compared with Adam. Our code is available at https://github.com/yuanwei2019/EAdam-optimizer.

Esports, despite its expanding interest, lacks fundamental sports analytics resources such as accessible data or proven and reproducible analytical frameworks. Even Counter-Strike: Global Offensive (CSGO), the second most popular esport, suffers from these problems. Thus, quantitative evaluation of CSGO players, a task important to teams, media, bettors and fans, is difficult. To address this, we introduce (1) a data model for CSGO with an open-source implementation; (2) a graph distance measure for defining distances in CSGO; and (3) a context-aware framework to value players' actions based on changes in their team's chances of winning. Using over 70 million in-game CSGO events, we demonstrate our framework's consistency and independence compared to existing valuation frameworks. We also provide use cases demonstrating high-impact play identification and uncertainty estimation.

Learning to quantify (a.k.a.\ quantification) is a task concerned with training unbiased estimators of class prevalence via supervised learning. This task originated with the observation that "Classify and Count" (CC), the trivial method of obtaining class prevalence estimates, is often a biased estimator, and thus delivers suboptimal quantification accuracy; following this observation, several methods for learning to quantify have been proposed that have been shown to outperform CC. In this work we contend that previous works have failed to use properly optimised versions of CC. We thus reassess the real merits of CC (and its variants), and argue that, while still inferior to some cutting-edge methods, they deliver near-state-of-the-art accuracy once (a) hyperparameter optimisation is performed, and (b) this optimisation is performed by using a true quantification loss instead of a standard classification-based loss. Experiments on three publicly available binary sentiment classification datasets support these conclusions.

This paper proposes a new defense against neural network backdooring attacks that are maliciously trained to mispredict in the presence of attacker-chosen triggers. Our defense is based on the intuition that the feature extraction layers of a backdoored network embed new features to detect the presence of a trigger and the subsequent classification layers learn to mispredict when triggers are detected. Therefore, to detect backdoors, the proposed defense uses two synergistic anomaly detectors trained on clean validation data: the first is a novelty detector that checks for anomalous features, while the second detects anomalous mappings from features to outputs by comparing with a separate classifier trained on validation data. The approach is evaluated on a wide range of backdoored networks (with multiple variations of triggers) that successfully evade state-of-the-art defenses. Additionally, we evaluate the robustness of our approach on imperceptible perturbations, scalability on large-scale datasets, and effectiveness under domain shift. This paper also shows that the defense can be further improved using data augmentation.

Digital payment volume has proliferated in recent years with the rapid growth of small businesses and online shops. When processing these digital transactions, recognizing each merchant's real identity (i.e., business type) is vital to ensure the integrity of payment processing systems. Conventionally, this problem is formulated as a time series classification problem solely using the merchant transaction history. However, with the large scale of the data, and changing behaviors of merchants and consumers over time, it is extremely challenging to achieve satisfying performance from off-the-shelf classification methods. In this work, we approach this problem from a multi-modal learning perspective, where we use not only the merchant time series data but also the information of merchant-merchant relationship (i.e., affinity) to verify the self-reported business type (i.e., merchant category) of a given merchant. Specifically, we design two individual encoders, where one is responsible for encoding temporal information and the other is responsible for affinity information, and a mechanism to fuse the outputs of the two encoders to accomplish the identification task. Our experiments on real-world credit card transaction data between 71,668 merchants and 433,772,755 customers have demonstrated the effectiveness and efficiency of the proposed model.

Classical learning theory suggests that the optimal generalization performance of a machine learning model should occur at an intermediate model complexity, with simpler models exhibiting high bias and more complex models exhibiting high variance of the predictive function. However, such a simple trade-off does not adequately describe deep learning models that simultaneously attain low bias and variance in the heavily overparameterized regime. A primary obstacle in explaining this behavior is that deep learning algorithms typically involve multiple sources of randomness whose individual contributions are not visible in the total variance. To enable fine-grained analysis, we describe an interpretable, symmetric decomposition of the variance into terms associated with the randomness from sampling, initialization, and the labels. Moreover, we compute the high-dimensional asymptotic behavior of this decomposition for random feature kernel regression, and analyze the strikingly rich phenomenology that arises. We find that the bias decreases monotonically with the network width, but the variance terms exhibit non-monotonic behavior and can diverge at the interpolation boundary, even in the absence of label noise. The divergence is caused by the \emph{interaction} between sampling and initialization and can therefore be eliminated by marginalizing over samples (i.e. bagging) \emph{or} over the initial parameters (i.e. ensemble learning).

Machine learning has the potential to fuel further advances in data science, but it is greatly hindered by an ad hoc design process, poor data hygiene, and a lack of statistical rigor in model evaluation. Recently, these issues have begun to attract more attention as they have caused public and embarrassing issues in research and development. Drawing from our experience as machine learning researchers, we follow the machine learning process from algorithm design to data collection to model evaluation, drawing attention to common pitfalls and providing practical recommendations for improvements. At each step, case studies are introduced to highlight how these pitfalls occur in practice, and where things could be improved.

The goal in Reinforcement Learning (RL) is to learn agents that maximize long-term environmental rewards. Deep RL, which uses deep neural networks as function approximators for the policy and value-functions, has achieved outstanding results on a wide variety of sequential decision making problems, with the barometer of success usually being the total returns accumulated by the final policy. Due to the intrinsic nature of direct reward maximization, seldom is the focus on how the behavioral characteristics of the trained agent compare with the other possible behaviors in the solution space. For instance, consider the robotic manipulator arm in Figure 1a and the peg-insertion task. Though the task description is simple, for a sufficiently flexible arm, there are numerous ways (positions of the joints and the end-effector) to insert the peg in the hole (Figure 1b).

For a function $\mathsf{K} : \mathbb{R}^{d} \times \mathbb{R}^{d} \to \mathbb{R}_{\geq 0}$, and a set $P = \{ x_1, \ldots, x_n\} \subset \mathbb{R}^d$ of $n$ points, the $\mathsf{K}$ graph $G_P$ of $P$ is the complete graph on $n$ nodes where the weight between nodes $i$ and $j$ is given by $\mathsf{K}(x_i, x_j)$. In this paper, we initiate the study of when efficient spectral graph theory is possible on these graphs. We investigate whether or not it is possible to solve the following problems in $n^{1+o(1)}$ time for a $\mathsf{K}$-graph $G_P$ when $d < n^{o(1)}$: $\bullet$ Multiply a given vector by the adjacency matrix or Laplacian matrix of $G_P$ $\bullet$ Find a spectral sparsifier of $G_P$ $\bullet$ Solve a Laplacian system in $G_P$'s Laplacian matrix For each of these problems, we consider all functions of the form $\mathsf{K}(u,v) = f(\|u-v\|_2^2)$ for a function $f:\mathbb{R} \rightarrow \mathbb{R}$. We provide algorithms and comparable hardness results for many such $\mathsf{K}$, including the Gaussian kernel, Neural tangent kernels, and more. For example, in dimension $d = \Omega(\log n)$, we show that there is a parameter associated with the function $f$ for which low parameter values imply $n^{1+o(1)}$ time algorithms for all three of these problems and high parameter values imply the nonexistence of subquadratic time algorithms assuming Strong Exponential Time Hypothesis ($\mathsf{SETH}$), given natural assumptions on $f$. As part of our results, we also show that the exponential dependence on the dimension $d$ in the celebrated fast multipole method of Greengard and Rokhlin cannot be improved, assuming $\mathsf{SETH}$, for a broad class of functions $f$. To the best of our knowledge, this is the first formal limitation proven about fast multipole methods.

The loss surface of an overparameterized neural network (NN) possesses many global minima of zero training error. We explain how common variants of the standard NN training procedure change the minimizer obtained. First, we make explicit how the size of the initialization of a strongly overparameterized NN affects the minimizer and can deteriorate its final test performance. We propose a strategy to limit this effect. Then, we demonstrate that for adaptive optimization such as AdaGrad, the obtained minimizer generally differs from the gradient descent (GD) minimizer. This adaptive minimizer is changed further by stochastic mini-batch training, even though in the non-adaptive case GD and stochastic GD result in essentially the same minimizer. Lastly, we explain that these effects remain relevant for less overparameterized NNs. While overparameterization has its benefits, our work highlights that it induces sources of error absent from underparameterized models, some of which can be challenging to control.