News

Enhancing image feature extraction to boost image classification accuracy has been a significant research focus at the MCL lab. Initially, PixelHop++ was developed to efficiently extract image features and perform accurate image classification. Subsequently, the Least-Squares Normal Transform (LNT) was introduced to further enhance image features, improving classification results with PixelHop++ on standard image databases such as MNIST and FMNIST. Despite achieving commendable performance, further refinements remain desirable to push accuracy limits even higher.

To address this, we propose a novel pipeline of four distinct experimental setups involving different pooling strategies—absolute maximum pooling and variance pooling—at hops 1 and 2. We extract LNT features specifically from hops 1, 3, and 4 for each experiment. At hop-1, pooling (either max or variance) generates 10 LNT features per channel, resulting in a total of 250 features. Hop-3 involves transforming the (N, 3, 3, Feature) tensor to produce 90 LNT features. From hop-4, 10 additional LNT features are acquired following a DFT-based feature selection. These 350 LNT features from hops 1, 3, and 4 are concatenated alongside selected hop-4 features. Finally, features aggregated from all four experimental setups are combined, and a 10-class classifier is trained on these comprehensive feature sets, demonstrating an improvement in classification performance.

Image dehazing plays a crucial role in digital imaging by removing atmospheric distortions such as haze and fog, thereby enhancing scene clarity for applications ranging from photography to autonomous driving. Traditionally, methods like the Dark Channel Prior (DCP) have been used to estimate haze effects, leveraging the observation that most non-hazy images contain very dark pixels in at least one color channel.

In a significant advancement, researchers have now introduced a novel approach that combines the strength of DCP with the efficiency of the GUSL pipeline. In this new method, DCP serves as the foundational technique to provide an initial estimate of the haze, while the GUSL pipeline is employed to predict and correct the residual errors left by the DCP. This two-step process refines the dehazing process by capturing subtle details that DCP alone might miss.

The GUSL pipeline utilizes unsupervised representation learning for robust feature extraction, followed by supervised feature learning to enhance computational efficiency and output quality. This approach not only improves the overall dehazing performance but also maintains a lightweight design suitable for real-time applications on resource-constrained devices.

By integrating DCP with residue prediction through GUSL, the new method delivers superior image clarity with reduced computational overhead, making it an attractive solution for modern imaging challenges in mobile and edge computing environments.

Demosaicing is a critical process in digital imaging. Since each pixel on a typical sensor captures only one color channel, red, green, or blue, the complete color image must be reconstructed from incomplete data. Conventional approaches, including deep learning models, have made impressive strides in quality, yet their significant computational requirements often limit their deployment on resource-constrained edge devices.Mahtab Movahhedrad of the MCL Lab has introduced an innovative approach to digital imaging that promises to transform how devices reconstruct full-color images from partial sensor data. The new method, dubbed green U-shaped image demosaicing (GUSID), leverages green learning (GL) principles to offer a lightweight, transparent, and efficient alternative to traditional, computationally heavy deep learning techniques. GUSID takes a distinct path. Instead of relying on deep neural networks, it uses unsupervised representation learning for robust feature extraction, followed by supervised feature learning to enhance computational efficiency and maintain high-quality output. This dual-stage process allows GUSID to minimize computational overhead while delivering competitive accuracy. Its compact design and support for parallelized training make it particularly well-suited for real-time vision applications on devices with limited processing capabilities. As digital imaging continues to evolve, breakthroughs like GUSID not only enhance performance but also pave the way for future innovations in edge computing and real-time processing. The MCL Lab is poised to lead this exciting frontier, proving that sometimes, a smarter, leaner approach can make all the difference.

The IR-drop (voltage drop) analysis in integrated circuits is essential as the circuit designs become more complex and compact. Due to the complicated and tiny designs, the power delivery network (PDN) cannot deliver a target voltage to each cell, causing reliability issues and performance drops. Therefore, IR-drop estimation is a crucial step to ensure the functionality of the circuits.

Traditionally, IR-drop estimation relies on solving linear equations based on Kirchhoff’s current and voltage laws. Nevertheless, as the designs become more complex, the computational cost and simulation time increase significantly. In this research, an energy-efficient and high-performance static IR-drop estimation called GIRD (Green IR-Drop) is proposed.  GIRD processes the IC design input in three steps. First, the input netlist data are converted to multi-channel maps. Their joint spatial-spectral representations are determined with PixelHop. Next, discriminant features are selected using the relevant feature test (RFT). Finally, the selected features are fed to the XGBoost (eXtreme Gradient Boosting trees) regressor. Both PixelHop and RFT are green learning tools. GIRD yields a low carbon footprint due to its smaller model sizes and lower computational complexity. Besides, its performance scales well with small training datasets. Experiments on synthetic and real circuits are given to demonstrate the superior performance of GIRD. The model size and the complexity, measured by the Floating Point Operations (FLOPs) of GIRD, are only 1/1000 and 1/100 of deep-learning methods, respectively.

We are so happy to welcome a new MCL member, Cynthia Huang joining MCL this semester. Here is a quick interview with Cynthia:

1. Could you briefly introduce yourself and your research interests?

I am Cynthia Huang, a junior undergraduate student majoring in Computer Engineering and Computer Science at USC. My research interests include machine learning, AutoML, and computer vision. Previously, I have worked on AutoML, specifically on simultaneous optimization of neural network architecture and weights. I hope to continue deepening my exploration of these fields in the future.

2. What is your impression about MCL and USC?

My impression of MCL is that it is a collaborative and inspiring group, where many talented individuals come together with exciting ideas. I truly appreciate the team spirit and supportive environment, where everyone is open to sharing insights and learning from one another. I find great joy in engaging in research discussions, exchanging perspectives, and brainstorming solutions to challenging problems in MCL. I am very excited about the opportunity to collaborate with everyone and look forward to contributing to meaningful research in the future!

3. What are your future expectations and plans in MCL?

Currently, I am working on semantic segmentation using green learning. In the future, I look forward to continuing research discussions and collaborations with everyone at MCL. I hope to further explore new ideas, exchange knowledge, and contribute to exciting research. Additionally, I am excited to connect with more members of the MCL community.