Armin Gerami

🔬 Research

Auditing Algorithmic Bias in Transformer-Based Stock Trading
Armin Gerami, Ramani Duraiswami; Neurips'25
Summary: This study introduces a metric based on Partial Information Decomposition (PID) to audit how transformer models used in finance rely on volatile data and are affected by price movement frequency. The analysis reveals that the model entirely disregards data volatility and shows a significant bias toward assets with lower-frequency price movements.

Optimized Linear Attention GPU Kernel Implementation
Armin Gerami, Ramani Duraiswami; TMLR'25
Summary: Linear attention (Kernel Separation) is a promising approach for calculating the attention in transformers with linear time complexity. However, it introduces a significant data movement overhead. We present a CUDA library that achieves a 3.3× speedup and 3.6× reduction in memory consumption for linear attention implementation compared to the state-of-the-art.

GUST: Graph Edge-Coloring Utilization for Accelerating Sparse Matrix Vector Multiplication
Armin Gerami, Bahar Asgari; ASPLOS'25
Summary: A software/hardware codesign to accelerate sparse matrix-vector multiplication. The harware enables resource sharing while the graph-edge-coloring schedules the input stream to prevent collisions.

Efficient Spatial Audio Rendering Via Differentiable FIR To IIR Estimation
Armin Gerami, Bowen Zhi, Dmitry N. Zotkin, Ramani Duraiswami; ICASSP'25
Summary: We introduce a novel convex optimization algorithm for estimating FIR filters with an IIR filter, and achieve 2× speedup with an accuracy of %0.01 through differential programming.

Towards Efficient Implementation of Differentiable Sparse Matrix Multiplication
Armin Gerami, Ramani Duraiswami; Preprint
Summary: Sparse Matrix-Matrix Multiplication (SpMM) is a crucial operation for improving the efficiency of large deep neural networks, but creating a high-performance, differentiable version compatible with gradient-based training is challenging. This paper introduces a novel differentiable SpMM implementation that achieves a 2.6x speedup over the state-of-the-art, enabling more efficient and scalable training for sparse neural network models.

Room Impulse Response Synthesis via Differentiable Feedback Delay Networks
Armin Gerami, Ramani Duraiswami; Preprint
Summary: We introduce a computationally efficient and tunable feedback-delay-network (FDN) architecture for real-time room-impulse-response (RIR) rendering, addressing the computational and latency challenges inherent in traditional convolution and Fourier transform-based methods.

🚀 Invention Disclosures

Differentiable FIR To IIR Filter Estimation

Rapid Energy and Emission Auditor

💻 Coding Languages

CUDA, C++, Python, Verilog