Machine learning techniques are widely used for estimating causal effects. Double/debiased machine learning (DML) (Chernozhukov et al., 2018) uses a double-robust score function that relies on the prediction of nuisance functions, such as the propensity score, which is the probability of treatment assignment conditional on covariates. Estimators relying on double-robust score functions are highly sensitive to errors in propensity score predictions. Machine learners increase the severity of this problem as they tend to over- or underestimate these probabilities. Several calibration approaches have been proposed to improve probabilistic forecasts of machine learners. This paper investigates the use of probability calibration approaches within the DML framework. Simulation results demonstrate that calibrating propensity scores may significantly reduces the root mean squared error of DML estimates of the average treatment effect in finite samples. We showcase it in an empirical example and provide conditions under which calibration does not alter the asymptotic properties of the DML estimator.

Federated Learning (FL) is becoming a popular paradigm for leveraging distributed data and preserving data privacy. However, due to the distributed characteristic, FL systems are vulnerable to Byzantine attacks that compromised clients attack the global model by uploading malicious model updates. Most existing Byzantine-robust FL systems statistically analyze the weights of whole individual model updates uploaded by clients to defend against Byzantine attacks. With the development of layer-level and parameter-level fine-grained attacks, the attacks' stealthiness and effectiveness have been significantly improved. Due to unawareness or overreaction, the existing model-level defense methods degrade the training efficiency and model performance. To address this problem, we propose SkyMask, a new attack-agnostic robust FL system that leverages fine-grained learnable masks to identify malicious model updates at the parameter-level. Specifically, the FL server applies parameter-level masks to model updates uploaded by clients and trains the masks over a small clean dataset (i.e., root dataset) to learn the subtle difference between benign and malicious model updates in a high-dimension space. Our extensive experiments involve different models on three public datasets under state-of-the-art (SOTA) attacks, where the results show that SkyMask achieves up to 10% higher testing accuracy compared with SOTA defense strategies and successfully defends against attacks with malicious clients of a high fraction up to 80%. In the meantime, the experimental results demonstrate the scalability of our approach and the weak dependence on the data distribution of the root dataset.

We present an empirical study of scaling properties of encoder-decoder Transformer models used in neural machine translation (NMT). We show that cross-entropy loss as a function of model size follows a certain scaling law. Specifically (i) We propose a formula which describes the scaling behavior of cross-entropy loss as a bivariate function of encoder and decoder size, and show that it gives accurate predictions under a variety of scaling approaches and languages; we show that the total number of parameters alone is not sufficient for such purposes. (ii) We observe different power law exponents when scaling the decoder vs scaling the encoder, and provide recommendations for optimal allocation of encoder/decoder capacity based on this observation. (iii) We also report that the scaling behavior of the model is acutely influenced by composition bias of the train/test sets, which we define as any deviation from naturally generated text (either via machine generated or human translated text). We observe that natural text on the target side enjoys scaling, which manifests as successful reduction of the cross-entropy loss. (iv) Finally, we investigate the relationship between the cross-entropy loss and the quality of the generated translations. We find two different behaviors, depending on the nature of the test data. For test sets which were originally translated from target language to source language, both loss and BLEU score improve as model size increases. In contrast, for test sets originally translated from source language to target language, the loss improves, but the BLEU score stops improving after a certain threshold. We release generated text from all models used in this study.