The development of containers as a tool for scalable computing has led to the increased use of the serverless edge computing paradigm. In particular, containers enable fast deployment of services to edge networks, allowing users to access spatially closer servers and resulting in lower latency. As instantiating containers can cause extra delay called a cold start delay, container caching, which keeps used containers alive in memory for reuse by the next user, has been proposed to mitigate the delay. However, such edge servers are constrained in their memory capacity, and there is a need for an efficient caching strategy. In this paper, we demonstrate that Multi-Agent Reinforcement Learning (MARL) can effectively learn a cache replacement policy and outperform traditional heuristic and centralized Deep Reinforcement Learning (DRL) algorithms. Our results also show that the proposed MARL’s smaller action space has a significant advantage over DRL, which requires full information about the network. In particular, our research demonstrates that MARL is able to scale to larger networks without significant sacrifices to performance.

Deep learning models have shown great performances in natural language processing tasks. While much attention has been paid to improvements in utility, privacy leakage and social bias are two major concerns arising in trained models. In order to tackle these problems, we protect individuals' sensitive information and mitigate gender bias simultaneously. First, we propose a selective privacy-preserving method that only obscures individuals' sensitive information. Then we propose a negative multi-task learning framework to mitigate the gender bias which contains a main task and a gender prediction task. We analyze two existing word embeddings and evaluate them on sentiment analysis and a medical text classification task. Our experimental results show that our negative multi-task learning framework can mitigate the gender bias while keeping models’ utility.

Crowdsourcing is a popular method for crowd workers to collaborate on tasks. However, workers coordinate and share answers during the crowdsourcing process. The term for this is "collusion". Copies from others and repeated submissions are detrimental to the quality of the assignments. The majority of the existing research on collusion detection is limited to ground truth problems (e.g., labeling tasks) and requires a predetermined threshold to be established in advance. In this paper, we aim to detect collusion behavior of workers in an adaptive way, and propose an Adaptive Clustering Based Collusion Detection approach (ACCD) for a broad range of task types and data types solved via crowdsourcing (e.g., continuous rating with or without distributions). Extensive experiments on both real-world and synthetic datasets show the superiority of ACCD over state-of-the-art approaches.

We study conversational group detection in varied social scenes using a message-passing Graph Neural Network (GNN) in combination with the Dominant Sets clustering algorithm. Our approach first describes a scene as an interaction graph, where nodes encode individual features and edges encode pairwise relationship data. Then, it uses a GNN to predict pairwise affinity values that represent the likelihood of two people interacting together, and computes non-overlapping group assignments based on these affinities. We evaluate the proposed approach on the Cocktail Party and MatchNMingle datasets. Our results suggest that using GNNs to leverage both individual and relationship features when computing groups is beneficial, especially when more features are available for each individual.

Many past studies have been conducted on speech/music discrimination due to the potential applications for broadcast and other media; however, it remains possible to expand the experimental scope to include samples of speech with varying amounts of background music. This paper focuses on the development and evaluation of two measures of the ratio between speech energy and music energy: a reference measure called speech-to-music ratio (SMR), which is known objectively only prior to mixing, and a feature called the stereo-input mix-to-peripheral level feature (SIMPL), which is computed from the stereo mixed signal as an imprecise estimate of SMR. SIMPL is an objective signal measure calculated by taking advantage of broadcast mixing techniques in which vocals are typically placed at stereo center, unlike most instruments. Conversely, SMR is a hidden variable defined by the relationship between the powers of portions of audio attributed to speech and music. It is shown that SIMPL is predictive of SMR and can be combined with state-of-the-art features in order to improve performance. For evaluation, this new metric is applied in speech/music (binary) classification, speech/music/mixed (trinary) classification, and a new speech-to-music ratio estimation problem. Promising results are achieved, including 93.06% accuracy for trinary classification and 3.86 dB RMSE for estimation of the SMR.