Better intelligence.
Smaller footprint.

FairTabGen is a large language model–based framework for generating synthetic tabular data that integrates both counterfactual and causal fairness into the generation process. By combining fairness-aware prompting, in-context learning, and iterative refinement, it creates data that is realistic, equitable, and useful for downstream tasks. The framework improves fairness metrics by up to 10% while preserving strong predictive performance, even when working with only a fraction of the original data—making it highly practical for privacy-sensitive and data-scarce applications.

Efficient deployment of AI models in autonomous and edge applications requires balancing computational demands with performance. This paper proposes three approximation schemes—approximate multipliers, low-multiplicative convolution, and variational inference with quantization—that collectively address this challenge across CNNs, DNNs, and especially Vision Transformers. By integrating these techniques, the study achieves substantial reductions in energy consumption and model size while maintaining accuracy close to baseline levels. The results highlight the potential of approximate computing as a viable strategy for enabling high-performance Vision Transformers and other models in resource-constrained environments.

Optimization and approximation techniques are commonly used to improve the efficiency of DNNs, but they often come at the cost of reduced performance. This paper introduces a novel approach that achieves both efficiency and performance improvement simultaneously through approximation. By leveraging Information Bottleneck theory, the work provides the first formal connection between efficiency techniques and information-theoretic analysis, using the information plane to study how approximation affects the learning behavior of DNNs.

This paper presents a comprehensive overview of Approximate Computing (AxC) techniques in Machine Learning, with a focus on energy-efficient Deep Learning. It explores four main strategies—quantization, approximate multiplication, in-memory computing, and input-dependent approximation—and discusses their impact on reducing energy consumption while maintaining reliable performance. The work highlights both software-based implementations for general-purpose systems and hardware-integrated solutions for custom SoC/SiP designs, aiming to support robust and efficient AI acceleration in mobile and edge applications.

CNNs are powerful but resource-intensive, making them challenging to deploy on edge and power-constrained devices. This paper introduces a novel family of approximate multipliers that reduce energy and delay while maintaining strong machine learning performance. It presents the first comprehensive study to demonstrate improvements in both computational efficiency and model accuracy, enabling practical, high-performance CNNs for resource-limited environments.

This survey explores the landscape of runtime adaptive Approximate Computing (AxC) techniques in Deep Learning, with a focus on energy-efficient deployment in domains such as computer vision and medical applications. It reviews methods like adaptive pruning, quantization, approximate multipliers, memory-aware optimizations, and reinforcement learning-based control strategies, particularly for CNNs. The work highlights the role of resource constraints and application-specific needs in selecting suitable AxC techniques and provides insights into their benefits and limitations, offering a valuable reference for researchers and practitioners aiming to balance energy efficiency with model accuracy.

This paper introduces the first Persian musical instrument recognition system designed to identify instruments in polyphonic audio recordings. The system operates in two stages: blind source separation using FastICA, followed by feature extraction and a two-step classification process based on Mel-frequency cepstral coefficients and spectral features. It first recognizes the instrument family, then the specific instrument. The model achieves high accuracy, demonstrating its effectiveness for analyzing complex Persian music signals.