Search for: All records

Search spaces hallmark the advancement of Neural Architecture Search (NAS). Large and complex search spaces with versatile building operators and structures provide more opportunities to brew promising architectures, yet pose severe challenges on efficient exploration and exploitation. Subsequently, several search space shrinkage methods optimize by selecting a single sub-region that contains some well-performing networks. Small performance and efficiency gains are observed with these methods but such techniques leave room for significantly improved search performance and are ineffective at retaining architectural diversity. We propose LISSNAS, an automated algorithm that shrinks a large space into a diverse, small search space with SOTA search performance. Our approach leverages locality, the relationship between structural and performance similarity, to efficiently extract many pockets of well-performing networks. We showcase our method on an array of search spaces spanning various sizes and datasets. We accentuate the effectiveness of our shrunk spaces when used in one-shot search by achieving the best Top-1 accuracy in two different search spaces. Our method achieves a SOTA Top-1 accuracy of 77.6% in ImageNet under mobile constraints, best-in-class Kendal-Tau, architectural diversity, and search space size.

 

Due to the often limited communication bandwidth of edge devices, most existing federated learning (FL) methods randomly select only a subset of devices to participate in training at each communication round. Compared with engaging all the available clients, such a random-selection mechanism could lead to significant performance degradation on non-IID (independent and identically distributed) data. In this paper, we present our key observation that the essential reason resulting in such performance degradation is the class-imbalance of the grouped data from randomly selected clients. Based on this observation, we design an efficient heterogeneity-aware client sampling mechanism, namely, Federated Class-balanced Sampling (Fed-CBS), which can effectively reduce class-imbalance of the grouped dataset from the intentionally selected clients. We first propose a measure of class-imbalance which can be derived in a privacy-preserving way. Based on this measure, we design a computationefficient client sampling strategy such that the actively selected clients will generate a more classbalanced grouped dataset with theoretical guarantees. Experimental results show that Fed-CBS outperforms the status quo approaches in terms of test accuracy and the rate of convergence while achieving comparable or even better performance than the ideal setting where all the available clients participate in the FL training. 

Knowledge Distillation (KD) (Hinton et al., 2015) is one of the most effective approaches for deploying large-scale pre-trained language models in low-latency environments by transferring the knowledge contained in the largescale models to smaller student models. Previous KD approaches use the soft labels and intermediate activations generated by the teacher to transfer knowledge to the student model parameters alone. In this paper, we show that having access to non-parametric memory in the form of a knowledge base with the teacher’s soft labels and predictions can further enhance student capacity and improve generalization. To enable the student to retrieve from the knowledge base effectively, we propose a new Retrieval-augmented KD framework with a loss function that aligns the relational knowledge in teacher and student embedding spaces. We show through extensive experiments that our retrieval mechanism can achieve state-of-the-art performance for taskspecific knowledge distillation on the GLUE benchmark (Wang et al., 2018a). 

Deformable Convolutional Networks (DCN) have been proposed as a powerful tool to boost the representation power of Convolutional Neural Networks (CNN) in computer vision tasks via adaptive sampling of the input feature map. Much like vision transformers, DCNs utilize a more flexible inductive bias than standard CNNs and have also been shown to improve performance of particular models. For example, drop-in DCN layers were shown to increase the AP score of Mask RCNN by 10.6 points while introducing only 1% additional parameters and FLOPs, improving the state-of-the art model at the time of publication. However, despite evidence that more DCN layers placed earlier in the network can further improve performance, we have not seen this trend continue with further scaling of deformations in CNNs, unlike for vision transformers. Benchmarking experiments show that a realistically sized DCN layer (64H×64W, 64 in-out channel) incurs a 4× slowdown on a GPU platform, discouraging the more ubiquitous use of deformations in CNNs. These slowdowns are caused by the irregular input-dependent access patterns of the bilinear interpolation operator, which has a disproportionately low arithmetic intensity (AI) compared to the rest of the DCN. To address the disproportionate slowdown of DCNs and enable their expanded use in CNNs, we propose DefT, a series of workload-aware optimizations for DCN kernels. DefT identifies performance bottlenecks in DCNs and fuses specific operators that are observed to limit DCN AI. Our approach also uses statistical information of DCN workloads to adapt the workload tiling to the DCN layer dimensions, minimizing costly out-of-boundary input accesses. Experimental results show that DefT mitigates up to half of DCN slowdown over the current-art PyTorch implementation. This translates to a layerwise speedup of up to 134% and a reduction of normalized training time of 46% on a fully DCN-enabled ResNet model.