Search for: All records

We study deep neural networks for the multi-label classification (MLab) task through the lens of neural collapse (NC). Previous works have been restricted to the multi-class classification setting and discovered a prevalent NC phenomenon comprising of the following properties for the last-layer features: (i) the variability of features within every class collapses to zero, (ii) the set of feature means form an equi-angular tight frame (ETF), and (iii) the last layer classifiers collapse to the feature mean upon some scaling. We generalize the study to multi-label learning, and prove for the first time that a generalized NC phenomenon holds with the "pick-all-label'' formulation, which we term as MLab NC. While the ETF geometry remains consistent for features with a single label, multi-label scenarios introduce a unique combinatorial aspect we term the "tag-wise average" property, where the means of features with multiple labels are the scaled averages of means for single-label instances. Theoretically, under proper assumptions on the features, we establish that the only global optimizer of the pick-all-label cross-entropy loss satisfy the multi-label NC. In practice, we demonstrate that our findings can lead to better test performance with more efficient training techniques for MLab learning. 

This work proposes a deep learning (DL)-based framework, namely Sim2Real, for spectral signal reconstruction in reconstructive spectroscopy, focusing on efficient data sampling and fast inference time. The work focuses on the challenge of reconstructing real-world spectral signals in an extreme setting where only device-informed simulated data are available for training. Such device-informed simulated data are much easier to collect than real-world data but exhibit large distribution shifts from their real-world counterparts. To leverage such simulated data effectively, a hierarchical data augmentation strategy is introduced to mitigate the adverse effects of this domain shift, and a corresponding neural network for the spectral signal reconstruction with our augmented data is designed. Experiments using a real dataset measured from our spectrometer device demonstrate that Sim2Real achieves significant speed-up during the inference while attaining on-par performance with the state-of-the-art optimization-based methods. 

Classification and topic modeling are popular techniques in machine learning that extract information from large-scale datasets. By incorporating a priori information such as labels or important features, methods have been developed to perform classification and topic modeling tasks; however, most methods that can perform both do not allow for guidance of the topics or features. In this paper, we propose a novel method, namely Guided Semi-Supervised Non-negative Matrix Factorization (GSSNMF), that performs both classification and topic modeling by incorporating supervision from both pre-assigned document class labels and user-designed seed words. We test the performance of this method on legal documents provided by the California Innocence Project and the 20 Newsgroups dataset. Our results show that the proposed method improves both classification accuracy and topic coherence in comparison to past methods such as Semi-Supervised Non-negative Matrix Factorization (SSNMF), Guided Non-negative Matrix Factorization (Guided NMF), and Topic Supervised NMF. 

Electrical muscle stimulation (EMS) is an emergent technique that miniaturizes force feedback, especially popular for untethered haptic devices, such as mobile gaming, VR, or AR. However, the actuation displayed by interactive systems based on EMS is coarse and imprecise. EMS systems mostly focus on inducing movements in large muscle groups such as legs, arms, and wrists; whereas individual finger poses, which would be required, for example, to actuate a user's fingers to fingerspell even the simplest letters in sign language, are not possible. The lack of dexterity in EMS stems from two fundamental limitations: (1) lack of independence: when a particular finger is actuated by EMS, the current runs through nearby muscles, causing unwanted actuation of adjacent fingers; and, (2) unwanted oscillations: while it is relatively easy for EMS to start moving a finger, it is very hard for EMS to stop and hold that finger at a precise angle; because, to stop a finger, virtually all EMS systems contract the opposing muscle, typically achieved via controllers (e.g., PID)—unfortunately, even with the best controller tuning, this often results in unwanted oscillations. To tackle these limitations, we propose dextrEMS, an EMS-based haptic device featuring mechanical brakes attached to each finger joint. The key idea behind dextrEMS is that while the EMS actuates the fingers, it is our mechanical brake that stops the finger in a precise position. Moreover, it is also the brakes that allow dextrEMS to select which fingers are moved by EMS, eliminating unwanted movements by preventing adjacent fingers from moving. We implemented dextrEMS as an untethered haptic device, weighing only 68g, that actuates eight finger joints independently (metacarpophalangeal and proximal interphalangeal joints for four fingers), which we demonstrate in a wide range of haptic applications, such as assisted fingerspelling, a piano tutorial, guitar tutorial, and a VR game. Finally, in our technical evaluation, we found that dextrEMS outperformed EMS alone by doubling its independence and reducing unwanted oscillations.