This work addresses the linear consensus problem in multi-agent systems under adversarial attacks. We examine scenarios where legitimate agents utilize stochastic inter-agent trust observations to assess the likelihood of neighboring agents acting maliciously in order to mitigate their impact. Malicious agents, in turn, aim to strategically choose what values to send to maximize the disagreement among legitimate agents in a finite horizon. In contrast to prior studies that assume static adversarial behavior and focus on asymptotic consensus, this work investigates the impact of strategic attacks within a finite number of iterations. Specifically, we characterize the maximum disagreement that malicious agents can induce in finite time and the computational complexity of computing the best attack strategy. We compare the effectiveness of different attack strategies through numerical experiments.
We propose a unified multitasking deep learning method to identify and segment structural regions and detect immune cells of non-small cell lung cancer tissue. This approach combines multitask learning and deep supervision to significantly reduce training time compared with single-task models, while achieving competitive performance with 50% fewer parameters. We balance tasks using homoscedastic uncertainty and address class imbalance through a combination of focal loss and class weight adjustment. Tested on multiplexed WSI, our method overcomes challenges related to variability and limited annotations, while extracting robust spatial information from the tumor microenvironment (TME). This data can be leveraged to quantify tumor-immune dynamics and intercellular communication, thereby providing objective insights into the TME's structural and functional complexity.
Deep Reinforcement Learning (DRL) has achieved remarkable success across various domains, yet its real-world applicability remains challenging due to training instability, particularly in complex environments with sparse rewards. Imitation learning mitigates this issue by leveraging expert demonstrations, allowing learning in scenarios where standard RL struggles. Experts are typically humans or teacher agents, often trained with privileged information (PI)-data available during training but not at inference. In this work, we extend beyond standard action replication from teacher to student by also transferring knowledge of latent representations. First, a privileged teacher is trained to use ground-truth information to accelerate learning. Then, a student agent, which lacks access to PI, is trained to align both its actions and the geometry of its intermediate representations with those of the teacher-effectively cloning both decision-making and latent behavior. We evaluate our method on two robotic manipulation tasks and demonstrate that latent behavior cloning significantly improves reward performance and convergence time compared to standard behavior cloning.
Musical instrument classification in the symbolic domain is a challenging task due to the inherent differences between MIDI representations and their corresponding acoustic signals. In this study, we propose a contrastive learning-based framework, Contrastive Audio-MIDI Learning (CAMIL), to improve MIDI-domain instrument classification by leveraging both symbolic and audio information. In our approach, we use MIDI embeddings as anchors, pairing them with their corresponding audio embeddings as positive samples and audio embeddings from different instruments as negative samples. By optimizing a contrastive loss function, our model learns to align MIDI embeddings with their corresponding audio representations while pushing apart embeddings of different instruments. We evaluate our method on the Lakh MIDI dataset and demonstrate that it improves instrument classification performance in the symbolic domain. Our results highlight the potential of contrastive learning in bridging the gap between audio and MIDI representations for more robust musical instrument recognition.
Identification of linear time-varying acoustic systems with multiple inputs and outputs is required in signal processing tasks like echo cancellation or crosstalk cancellation. When all inputs are excited simultaneously, identification is difficult because each output is a superposition of the influence of all inputs. If the inputs are correlated, identification is even more difficult due to the so-called non-uniqueness problem. A recent approach uses an extended Kalman filter to identify acoustic systems on nonlinear lower-dimensional manifolds. We extend this approach to MIMO systems. Instead of simply increasing the size of the neural networks, we propose architectural variants to control the number of parameters. We show that restricting the size of the network in exchange for its flexibility is beneficial for online system identification.
