Class-Balanced Softmax: A Bayes Theory-Based Method for Long-Tailed Recognition

Deep learning models using traditional softmax classifiers have achieved remarkable success in various classification tasks. However, their performance degrades significantly on imbalanced datasets. Although Balanced Softmax is widely adopted as a state-of-the-art rebalancing method, it possesses inherent limitations, such as yielding disproportionately lower testing accuracy for tail classes. To mitigate these shortcomings, we propose the Class-Balanced Softmax (CBS). Rooted in a theoretical Bayesian framework and a heuristic power-law assumption, the CBS is a simple logit adjustment that is computationally inexpensive and easily integrated into existing pipelines. Furthermore, we characterise a fundamental phenomenon in models trained on imbalanced data, termed the preference issue, wherein models exhibit higher training error and a larger generalisation gap for classes with limited data. To quantify this issue, we introduce a novel metric and demonstrate that CBS effectively mitigates the preference issue. Extensive experiments on large-scale benchmarks show that CBS is highly scalable and outperforms existing methods, including Balanced Softmax.

Array Placement in Distributed Massive MIMO for Power Saving considering Environment Information

Distributed massive MIMO (D-mMIMO) has been considered for future networks as it holds the potential to offer superior capacity while enabling energy savings in the network. A D-mMIMO system has multiple arrays. Optimizing the locations of the arrays is essential for the energy efficiency of the system. In existing works, array placement has been optimized mostly based on common channel models, which rely on a given statistical distribution and Euclidean distance between user locations and arrays. These models are justified if applied to sufficiently large cells, where the statistical description of the channel is expected to fit its empirical condition. However, with the advent of small cells, this is no longer the case. The channel propagation condition becomes highly environment-specific. This paper investigates array placement optimization with different ways of modeling the propagation conditions taking the environment information (e.g., buildings) into account. We capture the environment information via a graph. Two shortest path-based propagation models are introduced based on the graph. We validate the performance of the models with a ray-tracing simulator considering different signal coverage levels. The simulation results demonstrate that higher energy efficiency is achieved by using the array placements found via the proposed models compared to a Euclidean distance-based propagation model in small cells. The average power saving at 96% signal coverage, for example, reaches more than 5 dB.

Array Placement in Distributed Massive MIMO for power saving considering Radiation Pattern

A distributed antenna system (DAS) consists of several interconnected access points (APs) which are distributed over an area. Each AP has an antenna array. In previous studies, the DAS has been demonstrated great potential to improve capacity and power efficiency compared to a centralized antenna system (CAS) which has all the antennas located in one place. The existing research also has shown that the placement of the APs is essential for the performance of the DAS. However, most research on AP placement does not take into account realistic constraints. For instance, they assume that APs can be placed at any location in a region or the array radiation pattern of each AP is isotropic. This paper focuses on optimizing the AP placement for the DAS with massive MIMO (D-mMIMO) in order to reduce the transmit power. A square topology for the AP placement is applied, which is reasonable for deploying the D-mMIMO in urban areas while also offering theoretically interesting insights. We investigate the impact of the radiation pattern, signal coherence, and region size on the D-mMIMO's performance. Our results suggest that i) among the three factors, the array radiation pattern of each AP is the most important one in determining the optimal AP placement for the D-mMIMO; ii) the performance of the D-mMIMO is highly impacted by the placement and array radiation pattern of each AP, the D-mMIMO with unoptimized placement may perform even worse than the CAS with massive MIMO (C-mMIMO); iii) with the consideration of patch antennas and the mutual coupling effect, the optimized D-mMIMO can potentially save more than 7 dB transmit power compared to the C-mMIMO. Furthermore, our analytical results also provide an intuition for determining an adequate AP placement for the D-mMIMO in practice.