Weakly-supervised temporal action localization aims to recognize and localize action segments in untrimmed videos given only video-level action labels for training. Without the boundary information of action segments, existing methods mostly rely on multiple instance learning (MIL), where the predictions of unlabeled instances (i.e., video snippets) are supervised by classifying labeled bags (i.e., untrimmed videos). However, this formulation typically treats snippets in a video as independent instances, ignoring the underlying temporal structures within and across action segments. To address this problem, we propose ASM-Loc, a novel WTAL framework that enables explicit, action-aware segment modeling beyond standard MIL-based methods. Our framework entails three segment-centric components: (i) dynamic segment sampling for compensating the contribution of short actions; (ii) intra- and inter-segment attention for modeling action dynamics and capturing temporal dependencies; (iii) pseudo instance-level supervision for improving action boundary prediction. Furthermore, a multi-step refinement strategy is proposed to progressively improve action proposals along the model training process. Extensive experiments on THUMOS-14 and ActivityNet-v1.3 demonstrate the effectiveness of our approach, establishing new state of the art on both datasets. The code and models are publicly available at https://github.com/boheumd/ASM-Loc.

Tensor decomposition has been proved to be effective for solving many problems in signal processing and machine learning[1]. Recently, tensor decomposition finds its advantage for compressing deep neural networks. In many applications of deep neural networks, it is critical to reduce the number of parameters and computation workload to accelerate inference speed in deployment of the network. Modern deep neural network consists of multiple layers with multi-array weights where tensor decomposition is a natural way to perform compression. It is achieved by decomposing the weight tensors in convolutional layers or fully-connected layers with specified tensor ranks (e.g. canonical ranks, tensor train ranks). Conventional tensor decomposition in compressing deep neural networks selects the ranks manually that requires tedious human efforts to finetune the performance. To overcome this issue, we propose a novel rank selection scheme, which is inspired by reinforcement learning, to automatically select ranks in recently studied tensor ring decomposition in each convolutional layer. Experimental results validate that our learning based rank selection significantly outperforms hand-crafted rank selection heuristics on a number of benchmark datasets, for the purpose of effectively compressing deep neural networks while maintaining comparable accuracy.

In a full-duplex bidirectional interference network with 2K transceivers, there are K communication pairs: each user transmits a message to and receives a message from one intended user and interferes with and experiences interference from all other users. All nodes may interact, or adapt inputs to past received signals, and may thus co-operate with each other. We derive a new outer bound, and use interference alignment to demonstrate that the optimal degrees of freedom (DoF, also known as the multiplexing gain) is K: full-duplex operation doubles the DoF, but interaction and co-operation does not further increase the DoF. We next characterize the DoF of a full-duplex bidirectional interference network with a MIMO full-duplex relay. If the relay is noncausal/instantaneous (at time k forwards a function of its received signals up to time k) and has 2K antennas, we demonstrate a one-shot scheme where the relay mitigates all interference to achieve the interference-free 2K DoF. In contrast, if the relay is causal (at time k forwards a function of its received signals up to time k - 1), we show that a full-duplex MIMO relay cannot increase the DoF of the full-duplex bidirectional interference network beyond K, as if no relay or interaction is present.

In a K-pair-user two-way interference channel (TWIC), 2K messages and 2K transmitters/receivers form a K-user IC in the forward direction (K messages) and another K-user IC in the backward direction which operate in full-duplex. All nodes may interact, or adapt inputs to past received signals. The optimal degrees of freedom (DoF, also known as the multiplexing gain) is known to be K [1]: full-duplex operation doubles the DoF, but interaction does not further increase the DoF. In this paper, we characterize the DoF of the K-pair-user TWIC with a MIMO, full-duplex relay. If the relay is noncausal/ instantaneous (at time k forwards a function of its received signals up to time k) and has 2K antennas, we demonstrate a one-shot scheme where the relay mitigates all interference to achieve the interference-free 2K DoF. In contrast, if the relay is causal (at time k forwards a function of its received signals up to time k - 1), we show that a full-duplex MIMO relay cannot increase the DoF of the K-pair-user TWIC beyond K, as if no relay or interaction is present.

Most wireless communication networks are two-way, where nodes act as both sources and destinations of messages. This allows for adaptation at or interaction between the nodes-a node's channel inputs may be functions of its message(s) and previously received signals allowing for potentially larger rates than those achievable in feedback-free one-way channels where inputs are functions of messages only. However, examples exist of channels where adaptation is not beneficial from a capacity perspective. We ask whether analogous results hold for several multiuser two-way networks. We first consider deterministic two-way channel models: the binary modulo-2 addition channel and a generalization of this, and the linear deterministic channel, which models Gaussian channels at high SNR. For these deterministic models, we obtain the capacity region for the two-way multiple access/broadcast channel (MAC/BC), the two-way Z channel, and the two-way interference channel (under certain partial adaptation constraints in some regimes). We permit all nodes to adapt their channel inputs to past outputs (except for portions of the linear high-SNR two-way interference channel where we only permit two of the four nodes to fully adapt). However, we show that the two-way fully or partially adaptive capacity region consists of two parallel one-way regions operating simultaneously in opposite directions, i.e., adaptation is useless. We next consider two noisy channel models: 1) the Gaussian two-way MAC/BC, where we show that adaptation can at most increase the sum-rate by (1/2) bit in each direction and 2) the two-way interference channel, where partial adaptation is shown to be useless when the interference is very strong. In the strong and weak interference regimes, we show that the nonadaptive Han and Kobayashi scheme utilized in parallel in both directions achieves to within a constant gap for the symmetric rate of the fully (for some regimes) or partially (for the remaining regimes) adaptive models. The central technical contribution is the derivation of new, computable outer bounds which allow for adaptation.

We characterize the degrees of freedom (DoF) of the full-duplex two-way interference channel with a non causal multi-antenna relay and time-varying channels. In a two-way interference channel (IC), 4 messages and 4 transmitters/receivers form an IC in the forward direction (2 messages) and another IC in the backward direction (2 messages) which operate simultaneously in full-duplex mode. Furthermore, all nodes are permitted to interact, i.e. adapt current channel inputs to past received signals. We propose a novel block Markov coding scheme combining the ideas of interference alignment and successive decoding to achieve the full, maximal, 4 degrees of freedom asymptotically. All source/destination nodes have a single antenna while 4 antennas at the relay are sufficient to achieve the full DoF. Interestingly, this implies that the non causal relay is able to effectively mitigate interference in this two way setting - i.e. each user in the two-way interference channel is able to exchange information with its desired user at interference-free rates thanks to the relay.

In multi-user two-way channels nodes are both sources and destinations of messages. This allows for “adaptation” at or “interaction” between the nodes - the next channel inputs may be a function of the past received signals at a particular node. How to best adapt is key to two-way communication problems, rendering them complex and challenging. However, examples exist of channels where adaptation is not beneficial from a capacity perspective; it is known that for the point-to-point two-way modulo 2 adder and Gaussian channels, adaptation does not increase capacity. Recently, it was shown that the two-way modulo-2 additive versions of the multiple-access / broadcast (MAC/BC respectively, in the two directions), the Z channel and the interference channel also have capacity regions equal to two parallel one-way versions of the channels. In this work we show that the same is true for the linear deterministic multi-user two-way channels which approximate their Gaussian counterparts at high SNR, which include the two-way MAC/BC channel, the two-way Z channel, and the two-way interference channel under some adaptation constraints. For all three channel models we obtain the capacity region, which is that of two one-way channels in each direction, which may be achieved without the use of adaptation.