Multimodal deep learning systems are deployed in dynamic scenarios due to the robustness afforded by multiple sensing modalities. Nevertheless, they struggle with varying compute resource availability (due to multi-tenancy, device heterogeneity, etc.) and fluctuating quality of inputs (from sensor feed corruption, environmental noise, etc.). Current multimodal systems employ static resource provisioning and cannot easily adapt when compute resources change over time. Additionally, their reliance on processing sensor data with fixed feature extractors is ill-equipped to handle variations in modality quality. Consequently, uninformative modalities, such as those with high noise, needlessly consume resources better allocated towards other modalities. We propose ADMN, a layer-wise Adaptive Depth Multimodal Network capable of tackling both challenges - it adjusts the total number of active layers across all modalities to meet compute resource constraints, and continually reallocates layers across input modalities according to their modality quality. Our evaluations showcase ADMN can match the accuracy of state-of-the-art networks while reducing up to 75% of their floating-point operations.

Due to their architecture and how they are trained, artificial neural networks are typically not robust toward pruning, replacing, or shuffling layers at test time. However, such properties would be desirable for different applications, such as distributed neural network architectures where the order of execution cannot be guaranteed or parts of the network can fail during inference. In this work, we address these issues through a number of proposed training approaches for vision transformers whose most important component is randomizing the execution order of attention modules at training time. We show that with our proposed approaches, vision transformers are indeed capable to adapt to arbitrary layer execution orders at test time assuming one tolerates a reduction (about 20\%) in accuracy at the same model size. We also find that our trained models can be randomly merged with each other resulting in functional ("Frankenstein") models without loss of performance compared to the source models. Finally, we layer-prune our models at test time and find that their performance declines gracefully.

Handling the ever-increasing scale of contemporary deep learning and transformer-based models poses a significant challenge. Overparameterized Transformer networks outperform prior art in Natural Language processing and Computer Vision. These models contain hundreds of millions of parameters, demanding significant computational resources and making them prone to overfitting. In this work we present LayerCollapse, a form of structured pruning to reduce the depth of fully connected layers. We develop a novel regularizer allowing for post-training compression without finetuning, while having limited impact on performance. LayerCollapse controls model expressiveness with regularization on the activations between fully connected layers, modulating the linearity of activation functions. A linear activation function reduces the rank of the transformation to the rank of the corresponding linear transformation. We demonstrate the effectiveness of LayerCollapse by showing its compression capabilities in sentimental analysis and image classification benchmarks. Moreover we show LayerCollapse is an effective compression aware regularization method in a language modeling benchmark.

Overparameterized transformer networks have obtained state of the art results in various natural language processing tasks, such as machine translation, language modeling, and question answering. These models contain hundreds of millions of parameters, necessitating a large amount of computation and making them prone to overfitting. In this work, we explore LayerDrop, a form of structured dropout, which has a regularization effect during training and allows for efficient pruning at inference time. In particular, we show that it is possible to select sub-networks of any depth from one large network without having to finetune them and with limited impact on performance. We demonstrate the effectiveness of our approach by improving the state of the art on machine translation, language modeling, summarization, question answering, and language understanding benchmarks. Moreover, we show that our approach leads to small BERT-like models of higher quality compared to training from scratch or using distillation.