Spatial-temporal graph models are prevailing for abstracting and modelling spatial and temporal dependencies. In this work, we ask the following question: whether and to what extent can we localise spatial-temporal graph models? We limit our scope to adaptive spatial-temporal graph neural networks (ASTGNNs), the state-of-the-art model architecture. Our approach to localisation involves sparsifying the spatial graph adjacency matrices. To this end, we propose Adaptive Graph Sparsification (AGS), a graph sparsification algorithm which successfully enables the localisation of ASTGNNs to an extreme extent (fully localisation). We apply AGS to two distinct ASTGNN architectures and nine spatial-temporal datasets. Intriguingly, we observe that spatial graphs in ASTGNNs can be sparsified by over 99.5% without any decline in test accuracy. Furthermore, even when ASTGNNs are fully localised, becoming graph-less and purely temporal, we record no drop in accuracy for the majority of tested datasets, with only minor accuracy deterioration observed in the remaining datasets. However, when the partially or fully localised ASTGNNs are reinitialised and retrained on the same data, there is a considerable and consistent drop in accuracy. Based on these observations, we reckon that (i) in the tested data, the information provided by the spatial dependencies is primarily included in the information provided by the temporal dependencies and, thus, can be essentially ignored for inference; and (ii) although the spatial dependencies provide redundant information, it is vital for the effective training of ASTGNNs and thus cannot be ignored during training. Furthermore, the localisation of ASTGNNs holds the potential to reduce the heavy computation overhead required on large-scale spatial-temporal data and further enable the distributed deployment of ASTGNNs.

Data collected by IoT devices are often private and have a large diversity across users. Therefore, learning requires pre-training a model with available representative data samples, deploying the pre-trained model on IoT devices, and adapting the deployed model on the device with local data. Such an on-device adaption for deep learning empowered applications demands data and memory efficiency. However, existing gradient-based meta learning schemes fail to support memory-efficient adaptation. To this end, we propose p-Meta, a new meta learning method that enforces structure-wise partial parameter updates while ensuring fast generalization to unseen tasks. Evaluations on few-shot image classification and reinforcement learning tasks show that p-Meta not only improves the accuracy but also substantially reduces the peak dynamic memory by a factor of 2.5 on average compared to state-of-the-art few-shot adaptation methods.

Adapting neural networks to unseen tasks with few training samples on resource-constrained devices benefits various Internet-of-Things applications. Such neural networks should learn the new tasks in few shots and be compact in size. Meta-learning enables few-shot learning, yet the meta-trained networks can be over-parameterised. However, naive combination of standard compression techniques like network pruning with meta-learning jeopardises the ability for fast adaptation. In this work, we propose adaptation-aware network pruning (ANP), a novel pruning scheme that works with existing meta-learning methods for a compact network capable of fast adaptation. ANP uses weight importance metric that is based on the sensitivity of the meta-objective rather than the conventional loss function, and adopts approximation of derivatives and layer-wise pruning techniques to reduce the overhead of computing the new importance metric. Evaluations on few-shot classification benchmarks show that ANP can prune meta-trained convolutional and residual networks by 85% without affecting their fast adaptation.

Many mobile applications demand selective execution of multiple correlated deep learning inference tasks on resource-constrained platforms. Given a set of deep neural networks, each pre-trained for a single task, it is desired that executing arbitrary combinations of tasks yields minimal computation cost. Pruning each network separately yields suboptimal computation cost due to task relatedness. A promising remedy is to merge the networks into a multitask network to eliminate redundancy across tasks before network pruning. However, pruning a multitask network combined by existing network merging schemes cannot minimise the computation cost of every task combination because they do not consider such a future pruning. To this end, we theoretically identify the conditions such that pruning a multitask network minimises the computation of all task combinations. On this basis, we propose Pruning-Aware Merging (PAM), a heuristic network merging scheme to construct a multitask network that approximates these conditions. The merged network is then ready to be further pruned by existing network pruning methods. Evaluations with different pruning schemes, datasets, and network architectures show that PAM achieves up to 4.87x less computation against the baseline without network merging, and up to 2.01x less computation against the baseline with a state-of-the-art network merging scheme.

There is a growing trend to deploy deep neural networks at the edge for high-accuracy, real-time data mining and user interaction. Applications such as speech recognition and language understanding often apply a deep neural network to encode an input sequence and then use a decoder to generate the output sequence. A promising technique to accelerate these applications on resource-constrained devices is network pruning, which compresses the size of the deep neural network without severe drop in inference accuracy. However, we observe that although existing network pruning algorithms prove effective to speed up the prior deep neural network, they lead to dramatic slowdown of the subsequent decoding and may not always reduce the overall latency of the entire application. To rectify such drawbacks, we propose entropy-based pruning, a new regularizer that can be seamlessly integrated into existing network pruning algorithms. Our key theoretical insight is that reducing the information entropy of the deep neural network outputs decreases the upper bound of the subsequent decoding search space. We validate our solution with two state-of-the-art network pruning algorithms on two model architectures. Experimental results show that compared with existing network pruning algorithms, our entropy-based pruning method notably suppresses and even eliminates the increase of decoding time, and achieves shorter overall latency with only negligible extra accuracy loss in the applications.