Debabrata Mahapatra

Debabrata Mahapatra, Chaosheng Dong, Yetian Chen et al.

Learning to Rank (LTR) technique is ubiquitous in the Information Retrieval system nowadays, especially in the Search Ranking application. The query-item relevance labels typically used to train the ranking model are often noisy measurements of human behavior, e.g., product rating for product search. The coarse measurements make the ground truth ranking non-unique with respect to a single relevance criterion. To resolve ambiguity, it is desirable to train a model using many relevance criteria, giving rise to Multi-Label LTR (MLLTR). Moreover, it formulates multiple goals that may be conflicting yet important to optimize for simultaneously, e.g., in product search, a ranking model can be trained based on product quality and purchase likelihood to increase revenue. In this research, we leverage the Multi-Objective Optimization (MOO) aspect of the MLLTR problem and employ recently developed MOO algorithms to solve it. Specifically, we propose a general framework where the information from labels can be combined in a variety of ways to meaningfully characterize the trade-off among the goals. Our framework allows for any gradient based MOO algorithm to be used for solving the MLLTR problem. We test the proposed framework on two publicly available LTR datasets and one e-commerce dataset to show its efficacy.

Shubham Agarwal, Sai Sundaresan, Subrata Mitra et al.

Retrieval-Augmented Generation (RAG) is often used with Large Language Models (LLMs) to infuse domain knowledge or user-specific information. In RAG, given a user query, a retriever extracts chunks of relevant text from a knowledge base. These chunks are sent to an LLM as part of the input prompt. Typically, any given chunk is repeatedly retrieved across user questions. However, currently, for every question, attention-layers in LLMs fully compute the key values (KVs) repeatedly for the input chunks, as state-of-the-art methods cannot reuse KV-caches when chunks appear at arbitrary locations with arbitrary contexts. Naive reuse leads to output quality degradation. This leads to potentially redundant computations on expensive GPUs and increases latency. In this work, we propose Cache-Craft, a system for managing and reusing precomputed KVs corresponding to the text chunks (we call chunk-caches) in RAG-based systems. We present how to identify chunk-caches that are reusable, how to efficiently perform a small fraction of recomputation to fix the cache to maintain output quality, and how to efficiently store and evict chunk-caches in the hardware for maximizing reuse while masking any overheads. With real production workloads as well as synthetic datasets, we show that Cache-Craft reduces redundant computation by 51% over SOTA prefix-caching and 75% over full recomputation. Additionally, with continuous batching on a real production workload, we get a 1.6X speed up in throughput and a 2X reduction in end-to-end response latency over prefix-caching while maintaining quality, for both the LLaMA-3-8B and LLaMA-3-70B models.

Debabrata Mahapatra, Shubham Agarwal, Apoorv Saxena et al.

Prior work on input-token importance in auto-regressive transformers has relied on Softmax-normalized attention weights, which obscure the richer structure of pre-Softmax query-key logits. We introduce RCStat, a statistical framework that harnesses raw attention logits via Relative Contextualization (RC), a random variable measuring contextual alignment between token segments, and derive an efficient upper bound for RC. We demonstrate two applications: (i) Key-Value compression, where RC-based thresholds drive adaptive key-value eviction for substantial cache reduction with minimal quality loss; and (ii) Attribution, where RC yields higher-fidelity token-, sentence-, and chunk-level explanations than post-Softmax methods. Across question answering, summarization, and attribution benchmarks, RCStat achieves significant empirical gains, delivering state-of-the-art compression and attribution performance without any model retraining.

Debabrata Mahapatra, Vaibhav Rajan

Given multiple non-convex objective functions and objective-specific weights, Chebyshev scalarization (CS) is a well-known approach to obtain an Exact Pareto Optimal (EPO), i.e., a solution on the Pareto front (PF) that intersects the ray defined by the inverse of the weights. First-order optimizers that use the CS formulation to find EPO solutions encounter practical problems of oscillations and stagnation that affect convergence. Moreover, when initialized with a PO solution, they do not guarantee a controlled trajectory that lies completely on the PF. These shortcomings lead to modeling limitations and computational inefficiency in multi-task learning (MTL) and multi-criteria decision-making (MCDM) methods that utilize CS for their underlying non-convex multi-objective optimization (MOO). To address these shortcomings, we design a new MOO method, EPO Search. We prove that EPO Search converges to an EPO solution and empirically illustrate its computational efficiency and robustness to initialization. When initialized on the PF, EPO Search can trace the PF and converge to the required EPO solution at a linear rate of convergence. Using EPO Search we develop new algorithms: PESA-EPO for approximating the PF in a posteriori MCDM, and GP-EPO for preference elicitation in interactive MCDM; experiments on benchmark datasets confirm their advantages over competing alternatives. EPO Search scales linearly with the number of decision variables which enables its use for training deep networks. Empirical results on real data from personalized medicine, e-commerce and hydrometeorology demonstrate the efficacy of EPO Search for deep MTL.

Debabrata Mahapatra, Subhadip Mukherjee, Chandra Sekhar Seelamantula

We address the problem of reconstructing sparse signals from noisy and compressive measurements using a feed-forward deep neural network (DNN) with an architecture motivated by the iterative shrinkage-thresholding algorithm (ISTA). We maintain the weights and biases of the network links as prescribed by ISTA and model the nonlinear activation function using a linear expansion of thresholds (LET), which has been very successful in image denoising and deconvolution. The optimal set of coefficients of the parametrized activation is learned over a training dataset containing measurement-sparse signal pairs, corresponding to a fixed sensing matrix. For training, we develop an efficient second-order algorithm, which requires only matrix-vector product computations in every training epoch (Hessian-free optimization) and offers superior convergence performance than gradient-descent optimization. Subsequently, we derive an improved network architecture inspired by FISTA, a faster version of ISTA, to achieve similar signal estimation performance with about 50% of the number of layers. The resulting architecture turns out to be a deep residual network, which has recently been shown to exhibit superior performance in several visual recognition tasks. Numerical experiments demonstrate that the proposed DNN architectures lead to 3 to 4 dB improvement in the reconstruction signal-to-noise ratio (SNR), compared with the state-of-the-art sparse coding algorithms.