We consider the problem of regret minimization in non-parametric stochastic bandits. When the rewards are known to be bounded from above, there exists asymptotically optimal algorithms, with asymptotic regret depending on an infi-mum of Kullback-Leibler divergences (KL).

This work was done while the author was an intern at Disney Research | Studios 37th Conference on Neural Information Processing Systems (NeurIPS 2023).

The rapid advancement of large-language models (LLMs) has driven extensive research into parameter compression after training has been completed, yet compression during the training phase remains largely unexplored. In this work, we introduce Rate-Constrained Training (BackSlash), a novel training-time compression approach based on rate-distortion optimization (RDO). BackSlash enables a flexible trade-off between model accuracy and complexity, significantly reducing parameter redundancy while preserving performance. Experiments in various architectures and tasks demonstrate that BackSlash can reduce memory usage by 60% - 90% without accuracy loss and provides significant compression gain compared to compression after training. Moreover, BackSlash proves to be highly versatile: it enhances generalization with small Lagrange multipliers, improves model robustness to pruning (maintaining accuracy even at 80% pruning rates), and enables network simplification for accelerated inference on edge devices.

We consider the problem of regret minimization in non-parametric stochastic bandits. When the rewards are known to be bounded from above, there exists asymptotically optimal algorithms, with asymptotic regret depending on an infi-mum of Kullback-Leibler divergences (KL).

For animations and additional results please also check the included videos. We use a similar optimization strategy with DreamFusion, so unless otherwise noted the hyperparam-eters remain the same. DreamFusion we also train on a TPUv4 machine with 4 chips. We increase the number of optimization iterations from 15,000 to 50,000. We did not observe any significant benefits by training for more iterations.

This work was done while the author was an intern at Disney Research | Studios 37th Conference on Neural Information Processing Systems (NeurIPS 2023).

The authors discuss how the problems can be formulated as optimization of objective functions defined on the subgraphs. A straightforward search over the subgraphs is computationally infeasible, so the authors present a highly novel approach that leads to computationally efficient tests. The paper includes proofs that the tests are nearly minimax optimal for the exponential family of distributions and graphs satisfying the polynomial growth property. The paper concludes with an analysis of synthetic and real datasets. Strengths: (1) The paper addresses a problem of growing importance and presents novel approaches for statistical tests.

To infer a multilayer representation of high-dimensional count vectors, we propose the Poisson gamma belief network (PGBN) that factorizes each of its layers into the product of a connection weight matrix and the nonnegative real hidden units of the next layer. The PGBN's hidden layers are jointly trained with an upward-downward Gibbs sampler, each iteration of which upward samples Dirichlet distributed connection weight vectors starting from the first layer (bottom data layer), and then downward samples gamma distributed hidden units starting from the top hidden layer. The gamma-negative binomial process combined with a layer-wise training strategy allows the PGBN to infer the width of each layer given a fixed budget on the width of the first layer. The PGBN with a single hidden layer reduces to Poisson factor analysis. Example results on text analysis illustrate interesting relationships between the width of the first layer and the inferred network structure, and demonstrate that the PGBN, whose hidden units are imposed with correlated gamma priors, can add more layers to increase its performance gains over Poisson factor analysis, given the same limit on the width of the first layer.

Estimating the contact state between a grasped tool and the environment is essential for performing contact tasks such as assembly and object manipulation. Force signals are valuable for estimating the contact state, as they can be utilized even when the contact location is obscured by the tool. Previous studies proposed methods for estimating contact positions using force/torque signals; however, most methods require the geometry of the tool surface to be known. Although several studies have proposed methods that do not require the tool shape, these methods require considerable time for estimation or are limited to tools with low-dimensional shape parameters. Here, we propose a method for simultaneously estimating the contact position and tool shape, where the tool shape is represented by a grid, which is high-dimensional (more than 1000 dimensional). The proposed method uses a particle filter in which each particle has individual tool shape parameters, thereby to avoid directly handling a high-dimensional parameter space. The proposed method is evaluated through simulations and experiments using tools with curved shapes on a plane. Consequently, the proposed method can estimate the shape of the tool simultaneously with the contact positions, making the contact-position estimation more accurate.