We consider the task of minimizing the sum of smooth and strongly convex functions stored in a decentralized manner across the nodes of a communication network whose links are allowed to change in time. We solve two fundamental problems for this task. First, we establish the first lower bounds on the number of decentralized communication rounds and the number of local computations required to find an ϵ-accurate solution.

SIDER is a dataset for predicting side effect from the small molecule structure. It contains 27 classification tasks, corresponding to the 27 system organ classes following MedDRA classifications [1]. If taking a closer look at the MedDRA classification on the system organ level on its website, we can find a claim of "System Organ Classes (SOCs) which are groupings by aetiology (e.g. In addition, there is a SOC to contain issues pertaining to products and one to contain social circumstances." In fact, the two tasks among the 27 tasks are named "Social circumstances" and "Product issues", that corresponds to the claims above. Predicting such label from molecular structure alone is futile and therefore does not serve the purpose of a benchmarking dataset. The other problematic example in MoleculeNet is the PCBA dataset, originally used in [44]. However, as claimed in the original paper, "It should be noted that we did not perform any preprocessing of our datasets, such as removing potential experimental artifacts". And we have demonstrated the importance of removing the experimental artifacts in the data processing pipeline in the main text. There are more example issues with MoleculeNet that can be found in [52]. For Therapeutics Data Commons (TDC) [24], we used filters in our pipeline on small molecule-related tasks on and found issues with them. The promiscuity filter is not applied due to the long running time.

While deep learning has revolutionized computer-aided drug discovery, the AI community has predominantly focused on model innovation and placed less emphasis on establishing best benchmarking practices.

We present a new concentration of measure inequality for sums of independent bounded random variables, which we name a split-kl inequality. The inequality is particularly well-suited for ternary random variables, which naturally show up in a variety of problems, including analysis of excess losses in classification, analysis of weighted majority votes, and learning with abstention. We demonstrate that for ternary random variables the inequality is simultaneously competitive with the kl inequality, the Empirical Bernstein inequality, and the Unexpected Bernstein inequality, and in certain regimes outperforms all of them. It resolves an open question by Tolstikhin and Seldin [2013] and Mhammedi et al. [2019] on how to match simultaneously the combinatorial power of the kl inequality when the distribution happens to be close to binary and the power of Bersntein inequalities to exploit low variance when the probability mass is concentrated on the middle value.

In his address to a joint session of Congress, the president predicted that "our country is on the verge of a comeback the likes of which the world has never witnessed." That prediction is backed up by his recent announcements of massive new private sector investments in AI infrastructure and new executive orders to ensure that the U.S. leads the world in the industries of the future. In order to fulfill the promise that those actions suggest, however, it's essential that President Donald Trump support steps to shore up America's intellectual property system, the cornerstone of our innovation economy, rooting out malicious foreign interests and installing new leadership to help guide the comeback. To start, we need to address the fact that legal damages for patent infringement are no longer calculated reliably. U.S. courts have strayed from commonsense assessments to the detriment of American innovation.

When facing uncertainty, decision-makers want predictions they can trust. A machine learning provider can convey confidence to decision-makers by guaranteeing their predictions are distribution calibrated -- amongst the inputs that receive a predicted class probabilities vector q, the actual distribution over classes is q. For multi-class prediction problems, however, achieving distribution calibration tends to be infeasible, requiring sample complexity exponential in the number of classes C. In this work, we introduce a new notion--decision calibration--that requires the predicted distribution and true distribution to be "indistinguishable" to a set of downstream decision-makers. When all possible decision makers are under consideration, decision calibration is the same as distribution calibration. However, when we only consider decision makers choosing between a bounded number of actions (e.g.