Dynamic DNNs and Runtime Management for Efficient Inference on Mobile/Embedded Devices

Deep neural network (DNN) inference is increasingly being executed on mobile and embedded platforms due to several key advantages in latency, privacy and always-on availability. However, due to limited computing resources, efficient DNN deployment on mobile and embedded platforms is challenging. Although many hardware accelerators and static model compression methods were proposed by previous works, at system runtime, multiple applications are typically executed concurrently and compete for hardware resources. This raises two main challenges: Runtime Hardware Availability and Runtime Application Variability. Previous works have addressed these challenges through either dynamic neural networks that contain sub-networks with different performance trade-offs or runtime hardware resource management. In this thesis, we proposed a combined method, a system was developed for DNN performance trade-off management, combining the runtime trade-off opportunities in both algorithms and hardware to meet dynamically changing application performance targets and hardware constraints in real time. We co-designed novel Dynamic Super-Networks to maximise runtime system-level performance and energy efficiency on heterogeneous hardware platforms. Compared with SOTA, our experimental results using ImageNet on the GPU of Jetson Xavier NX show our model is 2.4x faster for similar ImageNet Top-1 accuracy, or 5.1% higher accuracy at similar latency. We also designed a hierarchical runtime resource manager that tunes both dynamic neural networks and DVFS at runtime. Compared with the Linux DVFS governor schedutil, our runtime approach achieves up to a 19% energy reduction and a 9% latency reduction in single model deployment scenario, and an 89% energy reduction and a 23% latency reduction in a two concurrent model deployment scenario.

Fluid Dynamic DNNs for Reliable and Adaptive Distributed Inference on Edge Devices

Distributed inference is a popular approach for efficient DNN inference at the edge. However, traditional Static and Dynamic DNNs are not distribution-friendly, causing system reliability and adaptability issues. In this paper, we introduce Fluid Dynamic DNNs (Fluid DyDNNs), tailored for distributed inference. Distinct from Static and Dynamic DNNs, Fluid DyDNNs utilize a novel nested incremental training algorithm to enable independent and combined operation of its sub-networks, enhancing system reliability and adaptability. Evaluation on embedded Arm CPUs with a DNN model and the MNIST dataset, shows that in scenarios of single device failure, Fluid DyDNNs ensure continued inference, whereas Static and Dynamic DNNs fail. When devices are fully operational, Fluid DyDNNs can operate in either a High-Accuracy mode and achieve comparable accuracy with Static DNNs, or in a High-Throughput mode and achieve 2.5x and 2x throughput compared with Static and Dynamic DNNs, respectively.

Improving the Robustness of Neural Multiplication Units with Reversible Stochasticity

Multilayer Perceptrons struggle to learn certain simple arithmetic tasks. Specialist neural modules for arithmetic can outperform classical architectures with gains in extrapolation, interpretability and convergence speeds, but are highly sensitive to the training range. In this paper, we show that Neural Multiplication Units (NMUs) are unable to reliably learn tasks as simple as multiplying two inputs when given different training ranges. Causes of failure are linked to inductive and input biases which encourage convergence to solutions in undesirable optima. A solution, the stochastic NMU (sNMU), is proposed to apply reversible stochasticity, encouraging avoidance of such optima whilst converging to the true solution. Empirically, we show that stochasticity provides improved robustness with the potential to improve learned representations of upstream networks for numerical and image tasks.

Physically Embodied Deep Image Optimisation

Physical sketches are created by learning programs to control a drawing robot. A differentiable rasteriser is used to optimise sets of drawing strokes to match an input image, using deep networks to provide an encoding for which we can compute a loss. The optimised drawing primitives can then be translated into G-code commands which command a robot to draw the image using drawing instruments such as pens and pencils on a physical support medium.

Shared Visual Representations of Drawing for Communication: How do different biases affect human interpretability and intent?

We present an investigation into how representational losses can affect the drawings produced by artificial agents playing a communication game. Building upon recent advances, we show that a combination of powerful pretrained encoder networks, with appropriate inductive biases, can lead to agents that draw recognisable sketches, whilst still communicating well. Further, we start to develop an approach to help automatically analyse the semantic content being conveyed by a sketch and demonstrate that current approaches to inducing perceptual biases lead to a notion of objectness being a key feature despite the agent training being self-supervised.

Learning Division with Neural Arithmetic Logic Modules

To achieve systematic generalisation, it first makes sense to master simple tasks such as arithmetic. Of the four fundamental arithmetic operations (+,-,$\times$,$\div$), division is considered the most difficult for both humans and computers. In this paper we show that robustly learning division in a systematic manner remains a challenge even at the simplest level of dividing two numbers. We propose two novel approaches for division which we call the Neural Reciprocal Unit (NRU) and the Neural Multiplicative Reciprocal Unit (NMRU), and present improvements for an existing division module, the Real Neural Power Unit (Real NPU). Experiments in learning division with input redundancy on 225 different training sets, find that our proposed modifications to the Real NPU obtains an average success of 85.3$\%$ improving over the original by 15.1$\%$. In light of the suggestion above, our NMRU approach can further improve the success to 91.6$\%$.

GhostShiftAddNet: More Features from Energy-Efficient Operations

Deep convolutional neural networks (CNNs) are computationally and memory intensive. In CNNs, intensive multiplication can have resource implications that may challenge the ability for effective deployment of inference on resource-constrained edge devices. This paper proposes GhostShiftAddNet, where the motivation is to implement a hardware-efficient deep network: a multiplication-free CNN with fewer redundant features. We introduce a new bottleneck block, GhostSA, that converts all multiplications in the block to cheap operations. The bottleneck uses an appropriate number of bit-shift filters to process intrinsic feature maps, then applies a series of transformations that consist of bit-wise shifts with addition operations to generate more feature maps that fully learn to capture information underlying intrinsic features. We schedule the number of bit-shift and addition operations for different hardware platforms. We conduct extensive experiments and ablation studies with desktop and embedded (Jetson Nano) devices for implementation and measurements. We demonstrate the proposed GhostSA block can replace bottleneck blocks in the backbone of state-of-the-art networks architectures and gives improved performance on image classification benchmarks. Further, our GhostShiftAddNet can achieve higher classification accuracy with fewer FLOPs and parameters (reduced by up to 3x) than GhostNet. When compared to GhostNet, inference latency on the Jetson Nano is improved by 1.3x and 2x on the GPU and CPU respectively.

Dynamic Transformer for Efficient Machine Translation on Embedded Devices

The Transformer architecture is widely used for machine translation tasks. However, its resource-intensive nature makes it challenging to implement on constrained embedded devices, particularly where available hardware resources can vary at run-time. We propose a dynamic machine translation model that scales the Transformer architecture based on the available resources at any particular time. The proposed approach, 'Dynamic-HAT', uses a HAT SuperTransformer as the backbone to search for SubTransformers with different accuracy-latency trade-offs at design time. The optimal SubTransformers are sampled from the SuperTransformer at run-time, depending on latency constraints. The Dynamic-HAT is tested on the Jetson Nano and the approach uses inherited SubTransformers sampled directly from the SuperTransformer with a switching time of <1s. Using inherited SubTransformers results in a BLEU score loss of <1.5% because the SubTransformer configuration is not retrained from scratch after sampling. However, to recover this loss in performance, the dimensions of the design space can be reduced to tailor it to a family of target hardware. The new reduced design space results in a BLEU score increase of approximately 1% for sub-optimal models from the original design space, with a wide range for performance scaling between 0.356s - 1.526s for the GPU and 2.9s - 7.31s for the CPU.

Language Models as Zero-shot Visual Semantic Learners

Visual Semantic Embedding (VSE) models, which map images into a rich semantic embedding space, have been a milestone in object recognition and zero-shot learning. Current approaches to VSE heavily rely on static word em-bedding techniques. In this work, we propose a Visual Se-mantic Embedding Probe (VSEP) designed to probe the semantic information of contextualized word embeddings in visual semantic understanding tasks. We show that the knowledge encoded in transformer language models can be exploited for tasks requiring visual semantic understanding.The VSEP with contextual representations can distinguish word-level object representations in complicated scenes as a compositional zero-shot learner. We further introduce a zero-shot setting with VSEPs to evaluate a model's ability to associate a novel word with a novel visual category. We find that contextual representations in language mod-els outperform static word embeddings, when the compositional chain of object is short. We notice that current visual semantic embedding models lack a mutual exclusivity bias which limits their performance.

What Remains of Visual Semantic Embeddings

Zero shot learning (ZSL) has seen a surge in interest over the decade for its tight links with the mechanism making young children recognize novel objects. Although different paradigms of visual semantic embedding models are designed to align visual features and distributed word representations, it is unclear to what extent current ZSL models encode semantic information from distributed word representations. In this work, we introduce the split of tiered-ImageNet to the ZSL task, in order to avoid the structural flaws in the standard ImageNet benchmark. We build a unified framework for ZSL with contrastive learning as pre-training, which guarantees no semantic information leakage and encourages linearly separable visual features. Our work makes it fair for evaluating visual semantic embedding models on a ZSL setting in which semantic inference is decisive. With this framework, we show that current ZSL models struggle with encoding semantic relationships from word analogy and word hierarchy. Our analyses provide motivation for exploring the role of context language representations in ZSL tasks.