Yahoo's native advertising (also known as Gemini native) serves billions of ad impressions daily, reaching a yearly run-rate of many hundred of millions USD. Driving the Gemini native models that are used to predict both click probability (pCTR) and conversion probability (pCONV) is OFFSET - a feature enhanced collaborative-filtering (CF) based event prediction algorithm. \offset is a one-pass algorithm that updates its model for every new batch of logged data using a stochastic gradient descent (SGD) based approach. Since OFFSET represents its users by their features (i.e., user-less model) due to sparsity issues, rule based hard frequency capping (HFC) is used to control the number of times a certain user views a certain ad. Moreover, related statistics reveal that user ad fatigue results in a dramatic drop in click through rate (CTR). Therefore, to improve click prediction accuracy, we propose a soft frequency capping (SFC) approach, where the frequency feature is incorporated into the OFFSET model as a user-ad feature and its weight vector is learned via logistic regression as part of OFFSET training. Online evaluation of the soft frequency capping algorithm via bucket testing showed a significant 7.3% revenue lift. Since then, the frequency feature enhanced model has been pushed to production serving all traffic, and is generating a hefty revenue lift for Yahoo Gemini native. We also report related statistics that reveal, among other things, that while users' gender does not affect ad fatigue, the latter seems to increase with users' age.

We consider two-player zero-sum stochastic games and propose a two-timescale $Q$-learning algorithm with function approximation that is payoff-based, convergent, rational, and symmetric between the two players. In two-timescale $Q$-learning, the fast-timescale iterates are updated in spirit to the stochastic gradient descent and the slow-timescale iterates (which we use to compute the policies) are updated by taking a convex combination between its previous iterate and the latest fast-timescale iterate. Introducing the slow timescale as well as its update equation marks as our main algorithmic novelty. In the special case of linear function approximation, we establish, to the best of our knowledge, the first last-iterate finite-sample bound for payoff-based independent learning dynamics of these types. The result implies a polynomial sample complexity to find a Nash equilibrium in such stochastic games. To establish the results, we model our proposed algorithm as a two-timescale stochastic approximation and derive the finite-sample bound through a Lyapunov-based approach. The key novelty lies in constructing a valid Lyapunov function to capture the evolution of the slow-timescale iterates. Specifically, through a change of variable, we show that the update equation of the slow-timescale iterates resembles the classical smoothed best-response dynamics, where the regularized Nash gap serves as a valid Lyapunov function. This insight enables us to construct a valid Lyapunov function via a generalized variant of the Moreau envelope of the regularized Nash gap. The construction of our Lyapunov function might be of broad independent interest in studying the behavior of stochastic approximation algorithms.

Recent data-privacy laws have sparked interest in machine unlearning, which involves removing the effect of specific training samples from a learnt model as if they were never present in the original training dataset. The challenge of machine unlearning is to discard information about the ``forget'' data in the learnt model without altering the knowledge about the remaining dataset and to do so more efficiently than the naive retraining approach. To achieve this, we adopt a projected-gradient based learning method, named as Projected-Gradient Unlearning (PGU), in which the model takes steps in the orthogonal direction to the gradient subspaces deemed unimportant for the retaining dataset, so as to its knowledge is preserved. By utilizing Stochastic Gradient Descent (SGD) to update the model weights, our method can efficiently scale to any model and dataset size. We provide empirically evidence to demonstrate that our unlearning method can produce models that behave similar to models retrained from scratch across various metrics even when the training dataset is no longer accessible. Our code is available at https://github.com/hnanhtuan/projected_gradient_unlearning.

We consider the stochastic gradient method with random reshuffling ($\mathsf{RR}$) for tackling smooth nonconvex optimization problems. $\mathsf{RR}$ finds broad applications in practice, notably in training neural networks. In this work, we first investigate the concentration property of $\mathsf{RR}$'s sampling procedure and establish a new high probability sample complexity guarantee for driving the gradient (without expectation) below $\varepsilon$, which effectively characterizes the efficiency of a single $\mathsf{RR}$ execution. Our derived complexity matches the best existing in-expectation one up to a logarithmic term while imposing no additional assumptions nor changing $\mathsf{RR}$'s updating rule. Furthermore, by leveraging our derived high probability descent property and bound on the stochastic error, we propose a simple and computable stopping criterion for $\mathsf{RR}$ (denoted as $\mathsf{RR}$-$\mathsf{sc}$). This criterion is guaranteed to be triggered after a finite number of iterations, and then $\mathsf{RR}$-$\mathsf{sc}$ returns an iterate with its gradient below $\varepsilon$ with high probability. Moreover, building on the proposed stopping criterion, we design a perturbed random reshuffling method ($\mathsf{p}$-$\mathsf{RR}$) that involves an additional randomized perturbation procedure near stationary points. We derive that $\mathsf{p}$-$\mathsf{RR}$ provably escapes strict saddle points and efficiently returns a second-order stationary point with high probability, without making any sub-Gaussian tail-type assumptions on the stochastic gradient errors. Finally, we conduct numerical experiments on neural network training to support our theoretical findings.

Natural gradient descent has a remarkable property that in the small learning rate limit, it displays an invariance with respect to network reparameterizations, leading to robust training behavior even for highly covariant network parameterizations. We show that optimization algorithms with this property can be viewed as discrete approximations of natural transformations from the functor determining an optimizer's state space from the diffeomorphism group if its configuration manifold, to the functor determining that state space's tangent bundle from this group. Algorithms with this property enjoy greater efficiency when used to train poorly parameterized networks, as the network evolution they generate is approximately invariant to network reparameterizations. More specifically, the flow generated by these algorithms in the limit as the learning rate vanishes is invariant under smooth reparameterizations, the respective flows of the parameters being determined by equivariant maps. By casting this property a natural transformation, we allow for generalizations beyond equivariance with respect to group actions; this framework can account for non-invertible maps such as projections, creating a framework for the direct comparison of training behavior across non-isomorphic network architectures, and the formal examination of limiting behavior as network size increases by considering inverse limits of these projections, should they exist. We introduce a simple method of introducing this naturality more generally and examine a number of popular machine learning training algorithms, finding that most are unnatural.

Designing civil structures such as bridges, dams or buildings is a complex task requiring many synergies from several experts. Each is responsible for different parts of the process. This is often done in a sequential manner, e.g. the structural engineer makes a design under the assumption of certain material properties (e.g. the strength class of the concrete), and then the material engineer optimizes the material with these restrictions. This paper proposes a holistic optimization procedure, which combines the concrete mixture design and structural simulations in a joint, forward workflow that we ultimately seek to invert. In this manner, new mixtures beyond standard ranges can be considered. Any design effort should account for the presence of uncertainties which can be aleatoric or epistemic as when data is used to calibrate physical models or identify models that fill missing links in the workflow. Inverting the causal relations established poses several challenges especially when these involve physics-based models which most often than not do not provide derivatives/sensitivities or when design constraints are present. To this end, we advocate Variational Optimization, with proposed extensions and appropriately chosen heuristics to overcome the aforementioned challenges. The proposed methodology is illustrated using the design of a precast concrete beam with the objective to minimize the global warming potential while satisfying a number of constraints associated with its load-bearing capacity after 28days according to the Eurocode, the demoulding time as computed by a complex nonlinear Finite Element model, and the maximum temperature during the hydration.

While many phenomena in physics and engineering are formally high-dimensional, their long-time dynamics often live on a lower-dimensional manifold. The present work introduces an autoencoder framework that combines implicit regularization with internal linear layers and $L_2$ regularization (weight decay) to automatically estimate the underlying dimensionality of a data set, produce an orthogonal manifold coordinate system, and provide the mapping functions between the ambient space and manifold space, allowing for out-of-sample projections. We validate our framework's ability to estimate the manifold dimension for a series of datasets from dynamical systems of varying complexities and compare to other state-of-the-art estimators. We analyze the training dynamics of the network to glean insight into the mechanism of low-rank learning and find that collectively each of the implicit regularizing layers compound the low-rank representation and even self-correct during training. Analysis of gradient descent dynamics for this architecture in the linear case reveals the role of the internal linear layers in leading to faster decay of a "collective weight variable" incorporating all layers, and the role of weight decay in breaking degeneracies and thus driving convergence along directions in which no decay would occur in its absence. We show that this framework can be naturally extended for applications of state-space modeling and forecasting by generating a data-driven dynamic model of a spatiotemporally chaotic partial differential equation using only the manifold coordinates. Finally, we demonstrate that our framework is robust to hyperparameter choices.

The potential of learned models for fundamental scientific research and discovery is drawing increasing attention worldwide. Physics-informed neural networks (PINNs), where the loss function directly embeds governing equations of scientific phenomena, is one of the key techniques at the forefront of recent advances. PINNs are typically trained using stochastic gradient descent methods, akin to their deep learning counterparts. However, analysis in this paper shows that PINNs' unique loss formulations lead to a high degree of complexity and ruggedness that may not be conducive for gradient descent. Unlike in standard deep learning, PINN training requires globally optimum parameter values that satisfy physical laws as closely as possible. Spurious local optimum, indicative of erroneous physics, must be avoided. Hence, neuroevolution algorithms, with their superior global search capacity, may be a better choice for PINNs relative to gradient descent methods. Here, we propose a set of five benchmark problems, with open-source codes, spanning diverse physical phenomena for novel neuroevolution algorithm development. Using this, we compare two neuroevolution algorithms against the commonly used stochastic gradient descent, and our baseline results support the claim that neuroevolution can surpass gradient descent, ensuring better physics compliance in the predicted outputs. %Furthermore, implementing neuroevolution with JAX leads to orders of magnitude speedup relative to standard implementations.

This work introduces DADAO: the first decentralized, accelerated, asynchronous, primal, first-order algorithm to minimize a sum of $L$-smooth and $\mu$-strongly convex functions distributed over a given network of size $n$. Our key insight is based on modeling the local gradient updates and gossip communication procedures with separate independent Poisson Point Processes. This allows us to decouple the computation and communication steps, which can be run in parallel, while making the whole approach completely asynchronous. This leads to communication acceleration compared to synchronous approaches. Our new method employs primal gradients and does not use a multi-consensus inner loop nor other ad-hoc mechanisms such as Error Feedback, Gradient Tracking, or a Proximal operator. By relating the inverse of the smallest positive eigenvalue of the Laplacian matrix $\chi_1$ and the maximal resistance $\chi_2\leq \chi_1$ of the graph to a sufficient minimal communication rate between the nodes of the network, we show that our algorithm requires $\mathcal{O}(n\sqrt{\frac{L}{\mu}}\log(\frac{1}{\epsilon}))$ local gradients and only $\mathcal{O}(n\sqrt{\chi_1\chi_2}\sqrt{\frac{L}{\mu}}\log(\frac{1}{\epsilon}))$ communications to reach a precision $\epsilon$, up to logarithmic terms. Thus, we simultaneously obtain an accelerated rate for both computations and communications, leading to an improvement over state-of-the-art works, our simulations further validating the strength of our relatively unconstrained method.

Kernel methods are a popular class of nonlinear predictive models in machine learning. Scalable algorithms for learning kernel models need to be iterative in nature, but convergence can be slow due to poor conditioning. Spectral preconditioning is an important tool to speed-up the convergence of such iterative algorithms for training kernel models. However computing and storing a spectral preconditioner can be expensive which can lead to large computational and storage overheads, precluding the application of kernel methods to problems with large datasets. A Nystrom approximation of the spectral preconditioner is often cheaper to compute and store, and has demonstrated success in practical applications. In this paper we analyze the trade-offs of using such an approximated preconditioner. Specifically, we show that a sample of logarithmic size (as a function of the size of the dataset) enables the Nystrom-based approximated preconditioner to accelerate gradient descent nearly as well as the exact preconditioner, while also reducing the computational and storage overheads.