Temporal Decisions: Leveraging Temporal Correlation for Efficient Decisions in Early Exit Neural Networks

Deep Learning is becoming increasingly relevant in Embedded and Internet-of-things applications. However, deploying models on embedded devices poses a challenge due to their resource limitations. This can impact the model's inference accuracy and latency. One potential solution are Early Exit Neural Networks, which adjust model depth dynamically through additional classifiers attached between their hidden layers. However, the real-time termination decision mechanism is critical for the system's efficiency, latency, and sustained accuracy. This paper introduces Difference Detection and Temporal Patience as decision mechanisms for Early Exit Neural Networks. They leverage the temporal correlation present in sensor data streams to efficiently terminate the inference. We evaluate their effectiveness in health monitoring, image classification, and wake-word detection tasks. Our novel contributions were able to reduce the computational footprint compared to established decision mechanisms significantly while maintaining higher accuracy scores. We achieved a reduction of mean operations per inference by up to 80% while maintaining accuracy levels within 5% of the original model. These findings highlight the importance of considering temporal correlation in sensor data to improve the termination decision.

Efficient Post-Training Augmentation for Adaptive Inference in Heterogeneous and Distributed IoT Environments

Early Exit Neural Networks (EENNs) present a solution to enhance the efficiency of neural network deployments. However, creating EENNs is challenging and requires specialized domain knowledge, due to the large amount of additional design choices. To address this issue, we propose an automated augmentation flow that focuses on converting an existing model into an EENN. It performs all required design decisions for the deployment to heterogeneous or distributed hardware targets: Our framework constructs the EENN architecture, maps its subgraphs to the hardware targets, and configures its decision mechanism. To the best of our knowledge, it is the first framework that is able to perform all of these steps. We evaluated our approach on a collection of Internet-of-Things and standard image classification use cases. For a speech command detection task, our solution was able to reduce the mean operations per inference by 59.67%. For an ECG classification task, it was able to terminate all samples early, reducing the mean inference energy by 74.9% and computations by 78.3%. On CIFAR-10, our solution was able to achieve up to a 58.75% reduction in computations. The search on a ResNet-152 base model for CIFAR-10 took less than nine hours on a laptop CPU. Our proposed approach enables the creation of EENN optimized for IoT environments and can reduce the inference cost of Deep Learning applications on embedded and fog platforms, while also significantly reducing the search cost - making it more accessible for scientists and engineers in industry and research. The low search cost improves the accessibility of EENNs, with the potential to improve the efficiency of neural networks in a wide range of practical applications.

Semi-supervised Active Learning for Video Action Detection

In this work, we focus on label efficient learning for video action detection. We develop a novel semi-supervised active learning approach which utilizes both labeled as well as unlabeled data along with informative sample selection for action detection. Video action detection requires spatio-temporal localization along with classification, which poses several challenges for both active learning informative sample selection as well as semi-supervised learning pseudo label generation. First, we propose NoiseAug, a simple augmentation strategy which effectively selects informative samples for video action detection. Next, we propose fft-attention, a novel technique based on high-pass filtering which enables effective utilization of pseudo label for SSL in video action detection by emphasizing on relevant activity region within a video. We evaluate the proposed approach on three different benchmark datasets, UCF-101-24, JHMDB-21, and Youtube-VOS. First, we demonstrate its effectiveness on video action detection where the proposed approach outperforms prior works in semi-supervised and weakly-supervised learning along with several baseline approaches in both UCF101-24 and JHMDB-21. Next, we also show its effectiveness on Youtube-VOS for video object segmentation demonstrating its generalization capability for other dense prediction tasks in videos.

AxOMaP: Designing FPGA-based Approximate Arithmetic Operators using Mathematical Programming

With the increasing application of machine learning (ML) algorithms in embedded systems, there is a rising necessity to design low-cost computer arithmetic for these resource-constrained systems. As a result, emerging models of computation, such as approximate and stochastic computing, that leverage the inherent error-resilience of such algorithms are being actively explored for implementing ML inference on resource-constrained systems. Approximate computing (AxC) aims to provide disproportionate gains in the power, performance, and area (PPA) of an application by allowing some level of reduction in its behavioral accuracy (BEHAV). Using approximate operators (AxOs) for computer arithmetic forms one of the more prevalent methods of implementing AxC. AxOs provide the additional scope for finer granularity of optimization, compared to only precision scaling of computer arithmetic. To this end, designing platform-specific and cost-efficient approximate operators forms an important research goal. Recently, multiple works have reported using AI/ML-based approaches for synthesizing novel FPGA-based AxOs. However, most of such works limit usage of AI/ML to designing ML-based surrogate functions used during iterative optimization processes. To this end, we propose a novel data analysis-driven mathematical programming-based approach to synthesizing approximate operators for FPGAs. Specifically, we formulate mixed integer quadratically constrained programs based on the results of correlation analysis of the characterization data and use the solutions to enable a more directed search approach for evolutionary optimization algorithms. Compared to traditional evolutionary algorithms-based optimization, we report up to 21% improvement in the hypervolume, for joint optimization of PPA and BEHAV, in the design of signed 8-bit multipliers.

AxOCS: Scaling FPGA-based Approximate Operators using Configuration Supersampling

The rising usage of AI and ML-based processing across application domains has exacerbated the need for low-cost ML implementation, specifically for resource-constrained embedded systems. To this end, approximate computing, an approach that explores the power, performance, area (PPA), and behavioral accuracy (BEHAV) trade-offs, has emerged as a possible solution for implementing embedded machine learning. Due to the predominance of MAC operations in ML, designing platform-specific approximate arithmetic operators forms one of the major research problems in approximate computing. Recently there has been a rising usage of AI/ML-based design space exploration techniques for implementing approximate operators. However, most of these approaches are limited to using ML-based surrogate functions for predicting the PPA and BEHAV impact of a set of related design decisions. While this approach leverages the regression capabilities of ML methods, it does not exploit the more advanced approaches in ML. To this end, we propose AxOCS, a methodology for designing approximate arithmetic operators through ML-based supersampling. Specifically, we present a method to leverage the correlation of PPA and BEHAV metrics across operators of varying bit-widths for generating larger bit-width operators. The proposed approach involves traversing the relatively smaller design space of smaller bit-width operators and employing its associated Design-PPA-BEHAV relationship to generate initial solutions for metaheuristics-based optimization for larger operators. The experimental evaluation of AxOCS for FPGA-optimized approximate operators shows that the proposed approach significantly improves the quality-resulting hypervolume for multi-objective optimization-of 8x8 signed approximate multipliers.

Temporal Patience: Efficient Adaptive Deep Learning for Embedded Radar Data Processing

Radar sensors offer power-efficient solutions for always-on smart devices, but processing the data streams on resource-constrained embedded platforms remains challenging. This paper presents novel techniques that leverage the temporal correlation present in streaming radar data to enhance the efficiency of Early Exit Neural Networks for Deep Learning inference on embedded devices. These networks add additional classifier branches between the architecture's hidden layers that allow for an early termination of the inference if their result is deemed sufficient enough by an at-runtime decision mechanism. Our methods enable more informed decisions on when to terminate the inference, reducing computational costs while maintaining a minimal loss of accuracy. Our results demonstrate that our techniques save up to 26% of operations per inference over a Single Exit Network and 12% over a confidence-based Early Exit version. Our proposed techniques work on commodity hardware and can be combined with traditional optimizations, making them accessible for resource-constrained embedded platforms commonly used in smart devices. Such efficiency gains enable real-time radar data processing on resource-constrained platforms, allowing for new applications in the context of smart homes, Internet-of-Things, and human-computer interaction.

Benchmarking self-supervised video representation learning

Self-supervised learning is an effective way for label-free model pre-training, especially in the video domain where labeling is expensive. Existing self-supervised works in the video domain use varying experimental setups to demonstrate their effectiveness and comparison across approaches becomes challenging with no standard benchmark. In this work, we first provide a benchmark that enables a comparison of existing approaches on the same ground. Next, we study five different aspects of self-supervised learning important for videos; 1) dataset size, 2) complexity, 3) data distribution, 4) data noise, and, 5)feature analysis. To facilitate this study, we focus on seven different methods along with seven different network architectures and perform an extensive set of experiments on 5 different datasets with an evaluation of two different downstream tasks. We present several interesting insights from this study which span across different properties of pretraining and target datasets, pretext-tasks, and model architectures among others. We further put some of these insights to the real test and propose an approach that requires a limited amount of training data and outperforms existing state-of-the-art approaches which use 10x pretraining data. We believe this work will pave the way for researchers to a better understanding of self-supervised pretext tasks in video representation learning.

Robust Empirical Risk Minimization with Tolerance

Developing simple, sample-efficient learning algorithms for robust classification is a pressing issue in today's tech-dominated world, and current theoretical techniques requiring exponential sample complexity and complicated improper learning rules fall far from answering the need. In this work we study the fundamental paradigm of (robust) $\textit{empirical risk minimization}$ (RERM), a simple process in which the learner outputs any hypothesis minimizing its training error. RERM famously fails to robustly learn VC classes (Montasser et al., 2019a), a bound we show extends even to `nice' settings such as (bounded) halfspaces. As such, we study a recent relaxation of the robust model called $\textit{tolerant}$ robust learning (Ashtiani et al., 2022) where the output classifier is compared to the best achievable error over slightly larger perturbation sets. We show that under geometric niceness conditions, a natural tolerant variant of RERM is indeed sufficient for $\gamma$-tolerant robust learning VC classes over $\mathbb{R}^d$, and requires only $\tilde{O}\left( \frac{VC(H)d\log \frac{D}{\gamma\delta}}{\epsilon^2}\right)$ samples for robustness regions of (maximum) diameter $D$.

Combining Gradients and Probabilities for Heterogeneous Approximation of Neural Networks

This work explores the search for heterogeneous approximate multiplier configurations for neural networks that produce high accuracy and low energy consumption. We discuss the validity of additive Gaussian noise added to accurate neural network computations as a surrogate model for behavioral simulation of approximate multipliers. The continuous and differentiable properties of the solution space spanned by the additive Gaussian noise model are used as a heuristic that generates meaningful estimates of layer robustness without the need for combinatorial optimization techniques. Instead, the amount of noise injected into the accurate computations is learned during network training using backpropagation. A probabilistic model of the multiplier error is presented to bridge the gap between the domains; the model estimates the standard deviation of the approximate multiplier error, connecting solutions in the additive Gaussian noise space to actual hardware instances. Our experiments show that the combination of heterogeneous approximation and neural network retraining reduces the energy consumption for multiplications by 70% to 79% for different ResNet variants on the CIFAR-10 dataset with a Top-1 accuracy loss below one percentage point. For the more complex Tiny ImageNet task, our VGG16 model achieves a 53 % reduction in energy consumption with a drop in Top-5 accuracy of 0.5 percentage points. We further demonstrate that our error model can predict the parameters of an approximate multiplier in the context of the commonly used additive Gaussian noise (AGN) model with high accuracy. Our software implementation is available under https://github.com/etrommer/agn-approx.

Video Action Detection: Analysing Limitations and Challenges

Beyond possessing large enough size to feed data hungry machines (eg, transformers), what attributes measure the quality of a dataset? Assuming that the definitions of such attributes do exist, how do we quantify among their relative existences? Our work attempts to explore these questions for video action detection. The task aims to spatio-temporally localize an actor and assign a relevant action class. We first analyze the existing datasets on video action detection and discuss their limitations. Next, we propose a new dataset, Multi Actor Multi Action (MAMA) which overcomes these limitations and is more suitable for real world applications. In addition, we perform a biasness study which analyzes a key property differentiating videos from static images: the temporal aspect. This reveals if the actions in these datasets really need the motion information of an actor, or whether they predict the occurrence of an action even by looking at a single frame. Finally, we investigate the widely held assumptions on the importance of temporal ordering: is temporal ordering important for detecting these actions? Such extreme experiments show existence of biases which have managed to creep into existing methods inspite of careful modeling.