#1 Researchers Achieve 6.7ms on Mobile with over 78% ImageNet Accuracy
With the increasing demand to efficiently deploy DNNs on mobile edge devices, it becomes much more critical to reduce unnecessary computation and increase the execution speed. Conventional methods to achieve this, including model compression and network architecture search (NAS), are primarily performed independently and do not fully consider compiler-level optimizations, a must-do for mobile acceleration.
To address the challenge, a group of researchers has proposed
- A general category of fine-grained structured pruning applicable to various DNN layers
- A comprehensive, compiler automatic code generation framework supporting different DNNs and different pruning schemes bridging the gap of model compression and NAS.
Further, they propose NPAS, a compiler-aware unified network pruning, and architecture search. Also, to deal with large search space, they present a meta-modeling procedure based on reinforcement learning with fast evaluation and Bayesian optimization, ensuring the total number of training epochs comparable with representative NAS frameworks.
Proposed framework achieves 6.7ms, 5.9ms, 3.9ms ImageNet inference times with 78.2%, 75% (MobileNet-V3 level), and 71% (MobileNet-V2 level) Top-1 accuracy respectively on an off-the-shelf mobile phone, consistently outperforming previous work.
#2 Data Augmentation for Graph Neural Network
Data augmentation has been widely used to improve generalizability of machine learning models. However, data augmentation for facilitating GNN training has unique challenges due to graph irregularity.
In this work, researchers tackle this problem by utilizing neural edge predictors as a means of exposing GNNs to likely (but nonexistent) edges and limiting exposure to unlikely (but existent) ones. They show that such edge predictors can encode class-homophily to promote intra-class edges and inter-class edges.
They propose the GAUG graph data augmentation framework which uses these insights to improve node classification performance in two inference settings. Extensive experiments show the proposed GAUG-O and GAUG-M achieves up to 17% (9%) absolute F1 performance improvements across architectures and datasets, and 15% (8%) over augmentation baselines.
#3 The NetHack Reinforcement Learning Environment
Progress in Reinforcement Learning (RL) algorithms goes hand-in-hand with the development of challenging environments that test current methods’ limits. While existing RL environments are either sufficiently complex or based on fast simulation, they are rarely both.
This paper presents the NetHack Learning Environment (NLE), a scalable, procedurally generated, stochastic, rich, and challenging environment for RL research based on the popular single-player terminal-based roguelike game, NetHack. Researchers argue that NetHack is sufficiently complex to drive long-term research on problems such as exploration, planning, skill acquisition, and language-conditioned RL, while dramatically reducing the computational resources required to gather a large amount of experience.
In the study, researchers compare NLE and its task suite to existing alternatives and discuss why it is an ideal medium for testing RL agents’ robustness and systematic generalization. They demonstrate empirical success for the early stages of the game using a distributed Deep RL baseline and Random Network Distillation exploration, alongside qualitative analysis of various agents trained in the environment. NLE is open source at Github
#4 Researchers Release the First of it Kind Benchmark for Generic Multiple Object Tracking
In this paper, researchers with Microsoft, Temple University, and Stony Brook University propose the first generic multiple object tracking (GMOT) benchmark named GMOT-40.
By thoroughly considering major MOT factors and carefully annotating all tracking objects, GMOT-40 contains 40 sequences evenly distributed in 10 object categories. Associated with GMOT are two GMOT evaluation Protocols, one focusing on target association and the other on one-shot tracking.
Several new baseline algorithms dedicated to one-shot GMOT are developed and evaluated together with relevant MOT trackers to provide references for future study. The evaluation shows that there is still considerable room to improve for GMOT, and further studies are desired. Overall, the researchers expect the benchmark to largely facilitate future research on GMOT, a critical yet under-explored computer vision problem.
#5 FairFace Challenge 2020 Results and Winning Methods
This work presents the design and results of the FairFace Recognition Challenge at ECCV’2020. The challenge attracted 151 participants. Their submissions were evaluated on a reannotated version of IJBC database enriched by newly collected 12,549 public domain images.
Participants were ranked using a new evaluation protocol where both accuracy and bias scores were considered. Top winning solutions obtained high performance in terms of accuracy (≥ 0.999 AUC-ROC) and bias scores. Common strategies by the participants were face pre-processing, homogenization of data distributions, the use of bias aware loss functions and ensemble models. The analysis of top-10 teams shows higher false positive rates (and lower false negative rates) for females with dark skin tone as well as the potential of eyeglasses and young age to increase the false positive rates too. See more results : FairFace Challenge at ECCV 2020