Adversarial patch attacks critically endanger object detection systems by causing severe mispredictions with small, easily realizable perturbations in both digital and physical environments. Existing defenses such as certified methods or patch detection suffer from high latency, while conventional adversarial training often overfits to specific patches and lacks generalization, particularly in multi-object scenarios. To overcome high latency and poor generalization, we introduce TRACE (Tuning Robustness by Adversarial-patch Confounder Elimination), an adversarial fine-tuning framework that leverages Instrumental Variable Regression in the feature space. TRACE treats patch-related variations—including location, rotation, and brightness—as confounders, thereby eliminating spurious correlations and guiding the model toward causal features that sustain robust detection. Evaluations on YOLOv5 and YOLOv8 show that TRACE consistently outperforms conventional defense methods in both efficiency and robustness under adaptive and unseen patch attacks. Moreover, physical testbed experiments confirm its effectiveness beyond digital settings, highlighting TRACE as a practical solution for achieving generalized robustness in object detection.

We report work in progress from interdisciplinary research on Assisted Living Technology in smart homes for older adults with mild cognitive impairments or dementia. We present our field trial, the set-up for collecting and storing data from real homes, and preliminary results on action recognition using low resolution depth video cameras. The data have been collected from seven apartments with one resident each over a period of two weeks. We propose a pre-processing of the depth videos by applying an Infinite Response Filter (IIR) for extracting the movements in the frames prior to classification. In this work we classify four actions: TV interaction (turn it on/ off and switch over), standing up, sitting down, and no movement. Our first results indicate that using the IIR filter for movement information extraction improves accuracy and can be an efficient method for recognizing actions. Our current implementation uses a convolutional long short-term memory (ConvLSTM) neural network, and achieved an average peak accuracy of 86%.

Stock market prediction was once considered to be infeasible. Recent studies on using text contents of information reporting platforms has opened up new ways of analyzing the stock market with machine learning. we propose using the Securities and Exchange Committee (SEC) mandated 10-Q form as a possible source of data for stock predictions. Using the 10-Q reports of S&P 500 companies, we create our corpus by extracting bag-of-words (BOW) of any additions made to the 10-Q documents. Then, we create feed-forward multilayer neural network on stock price ratios of different target prediction periods and achieve positive prediction rates. We demonstrate that text contents of 10-Q form may have information value to stock price prediction models.

Conventional cameras capture images with limited depth of field and no depth information. Camera systems have been proposed that enable additional depth information to be captured with the image. These systems reduce the resolution of the captured image or result in reduced sensitivity of the lens. We demonstrate a camera that is able to capture extended depth of field images together with depth information at each single frame while requiring minimal impact on the physical design of the camera or its performance. In this paper we show results with a camera for mobile devices, but this technology (named dual aperture to recall the major change in the camera model) can be applied with even greater effect in larger form factor cameras.

Unlike traditional trained channel equalizers, not much work has been done to theoretically characterize the convergence properties of blind channel equalizers due to their inherent nonlinearity. It is only recently that convergence properties of some well-known algorithms such as the generalised Sato algorithm (GSA) and the constant modulus algorithm (CMA) have been analytically derived. In this paper, the convergence properties of the stop-and-go algorithm proposed by Picchi and Prati (1987) are analyzed. The derived mean squared error and the coefficient trajectories are compared with simulation results to verify the validity of the analytical results.