SBT model uses randomized sharding and sub-models to improve Bayesian Additive Regression Trees.
problem Improving efficiency and accuracy of Bayesian Additive Regression Trees.
method Randomized sharding, sub-models, intersection tree structure, optimal design.
result Theoretical optimal weights and worst-case complexity of SBT model.
A new framework for clustering high-dimensional data using vertical shards.
problem Clustering high-dimensional data with the curse of dimensionality.
method Vertical Consensus Inference (VCI) that splits data into vertical shards for posterior inference.
result VCI can approximate inference on random partitions for high-dimensional data.
This paper optimizes deep learning training by efficiently sharding weight updates across replicas.
problem Redundant weight update computation on all replicas in data-parallel training.
method Automatic sharding of weight updates using static analysis and transformations on the training graph.
result Substantial speedups achieved on large-scale models using Cloud TPUs.
GShard enables scaling of large neural networks with automatic sharding and lightweight APIs.
problem Scaling neural networks to handle vast training data and compute efficiently.
method GShard uses lightweight annotation APIs and XLA compiler extensions for parallel computation.
result GShard successfully trained a 600 billion parameter model on 2048 TPUs in 4 days.
This work proposes a method to optimize hyperparameters without validation data.
problem Optimizing hyperparameters for neural networks with limited data.
method Neural network partitioning and subnetwork optimization.
result The method optimizes hyperparameters efficiently and without validation data.
We analyze the Hessian spectra of large models up to 100B parameters.
problem Accurate Hessian spectra of large foundation models are difficult to obtain.
method We use shard-local finite-difference Hessian vector products and stochastic Lanczos quadrature.
result We produce the first large-scale spectral density estimates of foundation models.
GraphEraser improves unlearning efficiency for graph data.
problem Improving unlearning efficiency for graph data.
method Two novel graph partition algorithms and a learning-based aggregation method.
result Achieves up to 35.94× unlearning time improvement on large datasets.
In federated distributed learning, the goal is to optimize a global training objective defined over distributed devices, where the data shard at each device is sampled from a possibly different distribution (a.k.a., heterogeneous or non i.i.d. data samples). In this paper, we generalize the local stochastic and full gr…
New method reconstructs hidden structures from noisy data.
problem Resurrecting hidden structures from incomplete or distorted data.
method Integrates Atiyah--Molino framework and Hantjies tensor.
result Exceptional robustness in noisy conditions with error-bounded reconstructions.
mSAM improves generalization by making models flatter.
problem Over-parameterized models can have varying generalization performance.
method mSAM modifies loss function to favor flatter minima through adversarial perturbations.
result mSAM achieves flatter minima than SAM and SGD, improving generalization.
NEST optimizes deep learning training by placing devices efficiently across networks and memory.
problem Inefficient device placement in distributed deep learning leads to high communication and memory overhead.
method NEST uses network-, compute-, and memory-aware dynamic programming to optimize device placement.
result NEST achieves up to 2.43 times higher throughput and better memory efficiency.
New method ensures consistent inference across different tensor parallel sizes for large language models.
problem Non-deterministic inference in large language models due to inconsistent reduction orders across GPUs.
method Tree-Based Invariant Kernels (TBIK) that align intra- and inter-GPU reduction orders through a unified hierarchical binary tree structure.
result Bit-wise identical results across different tensor parallel sizes for RL training.
Adaptive batch size schedules improve language model training efficiency and generalization.
problem Dilemma of choosing batch sizes in large-scale model training.
method General-purpose adaptive batch size schedules compatible with data and model parallelism.
result Adaptive batch size schedules outperform constant batch sizes and heuristic warmup schedules.
Large Sinkhorn couplings improve flow models in data generation tasks.
problem Training flow models with optimal transport couplings.
method Using large batches of source and target points, and applying entropic regularization with a low ε. result Flow models perform better with large Sinkhorn couplings and low regularization.