Franco Maria Nardini

Vincenzo Paparella, Vito Walter Anelli, Franco Maria Nardini et al.

Information Retrieval (IR) and Recommender Systems (RS) tasks are moving from computing a ranking of final results based on a single metric to multi-objective problems. Solving these problems leads to a set of Pareto-optimal solutions, known as Pareto frontier, in which no objective can be further improved without hurting the others. In principle, all the points on the Pareto frontier are potential candidates to represent the best model selected with respect to the combination of two, or more, metrics. To our knowledge, there are no well-recognized strategies to decide which point should be selected on the frontier. In this paper, we propose a novel, post-hoc, theoretically-justified technique, named "Population Distance from Utopia" (PDU), to identify and select the one-best Pareto-optimal solution from the frontier. In detail, PDU analyzes the distribution of the points by investigating how far each point is from its utopia point (the ideal performance for the objectives). The possibility of considering fine-grained utopia points allows PDU to select solutions tailored to individual user preferences, a novel feature we call "calibration". We compare PDU against existing state-of-the-art strategies through extensive experiments on tasks from both IR and RS. Experimental results show that PDU and combined with calibration notably impact the solution selection. Furthermore, the results show that the proposed framework selects a solution in a principled way, irrespective of its position on the frontier, thus overcoming the limits of other strategies.

Franco Maria Nardini, Cosimo Rulli, Salvatore Trani et al.

Quantization and pruning are two effective Deep Neural Networks model compression methods. In this paper, we propose Automatic Prune Binarization (APB), a novel compression technique combining quantization with pruning. APB enhances the representational capability of binary networks using a few full-precision weights. Our technique jointly maximizes the accuracy of the network while minimizing its memory impact by deciding whether each weight should be binarized or kept in full precision. We show how to efficiently perform a forward pass through layers compressed using APB by decomposing it into a binary and a sparse-dense matrix multiplication. Moreover, we design two novel efficient algorithms for extremely quantized matrix multiplication on CPU, leveraging highly efficient bitwise operations. The proposed algorithms are 6.9x and 1.5x faster than available state-of-the-art solutions. We extensively evaluate APB on two widely adopted model compression datasets, namely CIFAR10 and ImageNet. APB delivers better accuracy/memory trade-off compared to state-of-the-art methods based on i) quantization, ii) pruning, and iii) combination of pruning and quantization. APB outperforms quantization in the accuracy/efficiency trade-off, being up to 2x faster than the 2-bit quantized model with no loss in accuracy.

Thong Nguyen, Cosimo Rulli, Franco Maria Nardini et al.

State-of-the-art Learned Sparse Retrieval (LSR) models, such as Splade, typically employ a Language Modeling (LM) head to project latent hidden states into a lexically-anchored logit matrix. This intermediate matrix is subsequently transformed into a sparse lexical representation through element-wise operations (ReLU, Log1P) and max-pooling over the sequence dimension. Despite its effectiveness, the LM head creates a massive memory bottleneck due to the sheer size of the vocabulary (V), which can range from 30,000 to over 250,000 tokens in recent models. Materializing this matrix creates a significant memory bottleneck, limiting model scaling. The resulting I/O overhead between operators further throttles throughput and runtime performance. In this paper, we propose Sparton, a fast memory-efficient Triton kernel tailored for the LM head in LSR models. Sparton utilizes a fused approach that integrates the tiled matrix multiplication, ReLU, Log1P, and max-reduction into a single GPU kernel. By performing an early online reduction directly on raw logit tiles, Sparton avoids materializing the full logit matrix in memory. Our experiments demonstrate that the Sparton kernel, in isolation, achieves up to a 4.8x speedup and an order-of-magnitude reduction in peak memory usage compared to PyTorch baselines. Integrated into Splade (|V| ~ 30k), Sparton enables a 33% larger batch size and 14% faster training with no effectiveness loss. On a multilingual backbone (|V| ~ 250k), these gains jump to a 26x larger batch size and 2.5x faster training.

Silvio Martinico, Franco Maria Nardini, Cosimo Rulli et al.

Multivector retrieval models achieve state-of-the-art effectiveness through fine-grained token-level representations, but their deployment incurs substantial computational and memory costs. Current solutions, based on the well-known k-means clustering algorithm, group similar vectors together to enable both effective compression and efficient retrieval. However, standard k-means scales poorly with the number of clusters and dataset size, and favours frequent tokens during training while underrepresenting rare, discriminative ones. In this work, we introduce TACHIOM, a multivector retrieval system that exploits token-level structure to significantly accelerate both clustering and retrieval. By accounting for tokens' distribution during centroid allocation, TACHIOM easily scales to millions of centroids, enabling highly accurate document scoring using only centroids, avoiding expensive token-level computation. TACHIOM combines a graph-based index over centroids with an optimized Product Quantization layout for efficient final scoring. Experiments on MS-MARCOv1 and LoTTE show that TACHIOM achieves up to $247\times$ faster clustering than k-means and up to $9.8\times$ retrieval speedup over state-of-the-art systems while maintaining comparable or superior effectiveness.

Sebastian Bruch, Aditya Krishnan, Franco Maria Nardini

Clustering-based nearest neighbor search is an effective method in which points are partitioned into geometric shards to form an index, with only a few shards searched during query processing to find a set of top-$k$ vectors. Even though the search efficacy is heavily influenced by the algorithm that identifies the shards to probe, it has received little attention in the literature. This work bridges that gap by studying routing in clustering-based maximum inner product search. We unpack existing routers and notice the surprising contribution of optimism. We then take a page from the sequential decision making literature and formalize that insight following the principle of ``optimism in the face of uncertainty.'' In particular, we present a framework that incorporates the moments of the distribution of inner products within each shard to estimate the maximum inner product. We then present an instance of our algorithm that uses only the first two moments to reach the same accuracy as state-of-the-art routers such as ScaNN by probing up to $50\%$ fewer points on benchmark datasets. Our algorithm is also space-efficient: we design a sketch of the second moment whose size is independent of the number of points and requires $\mathcal{O}(1)$ vectors per shard.

Franco Maria Nardini, Raffaele Perego, Nicola Tonellotto et al.

We investigate the exploitation of both lexical and neural relevance signals for ad-hoc passage retrieval. Our exploration involves a large-scale training dataset in which dense neural representations of MS-MARCO queries and passages are complemented and integrated with 253 hand-crafted lexical features extracted from the same corpus. Blending of the relevance signals from the two different groups of features is learned by a classical Learning-to-Rank (LTR) model based on a forest of decision trees. To evaluate our solution, we employ a pipelined architecture where a dense neural retriever serves as the first stage and performs a nearest-neighbor search over the neural representations of the documents. Our LTR model acts instead as the second stage that re-ranks the set of candidates retrieved by the first stage to enhance effectiveness. The results of reproducible experiments conducted with state-of-the-art dense retrievers on publicly available resources show that the proposed solution significantly enhances the end-to-end ranking performance while relatively minimally impacting efficiency. Specifically, we achieve a boost in nDCG@10 of up to 11% with an increase in average query latency of only 4.3%. This confirms the advantage of seamlessly combining two distinct families of signals that mutually contribute to retrieval effectiveness.

Sebastian Bruch, Franco Maria Nardini, Cosimo Rulli et al.

Sparse embeddings of data form an attractive class due to their inherent interpretability: Every dimension is tied to a term in some vocabulary, making it easy to visually decipher the latent space. Sparsity, however, poses unique challenges for Approximate Nearest Neighbor Search (ANNS) which finds, from a collection of vectors, the k vectors closest to a query. To encourage research on this underexplored topic, sparse ANNS featured prominently in a BigANN Challenge at NeurIPS 2023, where approximate algorithms were evaluated on large benchmark datasets by throughput and accuracy. In this work, we introduce a set of novel data structures and algorithmic methods, a combination of which leads to an elegant, effective, and highly efficient solution to sparse ANNS. Our contributions range from a theoretically-grounded sketching algorithm for sparse vectors to reduce their effective dimensionality while preserving inner product-induced ranks; a geometric organization of the inverted index; and the blending of local and global information to improve the efficiency and efficacy of ANNS. Empirically, our final algorithm, dubbed Seismic, reaches sub-millisecond per-query latency with high accuracy on a large-scale benchmark dataset using a single CPU.

Andrea Giuseppe Di Francesco, Maria Sofia Bucarelli, Franco Maria Nardini et al.

Early-exit mechanisms allow deep neural networks to halt inference as soon as classification confidence is high enough, adaptively trading depth for confidence, and thereby cutting latency and energy on easy inputs while retaining full-depth accuracy for harder ones. Similarly, adding early exit mechanisms to Graph Neural Networks (GNNs), the go-to models for graph-structured data, allows for dynamic trading depth for confidence on simple graphs while maintaining full-depth accuracy on harder and more complex graphs to capture intricate relationships. Although early exits have proven effective across various deep learning domains, their potential within GNNs in scenarios that require deep architectures while resisting over-smoothing and over-squashing remains largely unexplored. We unlock that potential by first introducing Symmetric-Anti-Symmetric Graph Neural Networks (SAS-GNN), whose symmetry-based inductive biases mitigate these issues and yield stable intermediate representations that can be useful to allow early exiting in GNNs. Building on this backbone, we present Early-Exit Graph Neural Networks (EEGNNs), which append confidence-aware exit heads that allow on-the-fly termination of propagation based on each node or the entire graph. Experiments show that EEGNNs preserve robust performance as depth grows and deliver competitive accuracy on heterophilic and long-range benchmarks, matching attention-based and asynchronous message-passing models while substantially reducing computation and latency. We plan to release the code to reproduce our experiments.

Francesco Lettich, Emanuele Carlini, Franco Maria Nardini et al.

The rise of Artificial Intelligence and Large Language Models is driving increased GPU usage in data centers for complex training and inference tasks, impacting operational costs, energy demands, and the environmental footprint of large-scale computing infrastructures. This work addresses the online scheduling problem in GPU datacenters, which involves scheduling tasks without knowledge of their future arrivals. We focus on two objectives: minimizing GPU fragmentation and reducing power consumption. GPU fragmentation occurs when partial GPU allocations hinder the efficient use of remaining resources, especially as the datacenter nears full capacity. A recent scheduling policy, Fragmentation Gradient Descent (FGD), leverages a fragmentation metric to address this issue. Reducing power consumption is also crucial due to the significant power demands of GPUs. To this end, we propose PWR, a novel scheduling policy to minimize power usage by selecting power-efficient GPU and CPU combinations. This involves a simplified model for measuring power consumption integrated into a Kubernetes score plugin. Through an extensive experimental evaluation in a simulated cluster, we show how PWR, when combined with FGD, achieves a balanced trade-off between reducing power consumption and minimizing GPU fragmentation.

Lorenzo Beretta, Franco Maria Nardini, Roberto Trani et al.

A tournament graph is a complete directed graph, which can be used to model a round-robin tournament between $n$ players. In this paper, we address the problem of finding a champion of the tournament, also known as Copeland winner, which is a player that wins the highest number of matches. In detail, we aim to investigate algorithms that find the champion by playing a low number of matches. Solving this problem allows us to speed up several Information Retrieval and Recommender System applications, including question answering, conversational search, etc. Indeed, these applications often search for the champion inducing a round-robin tournament among the players by employing a machine learning model to estimate who wins each pairwise comparison. Our contribution, thus, allows finding the champion by performing a low number of model inferences. We prove that any deterministic or randomized algorithm finding a champion with constant success probability requires $Ω(\ell n)$ comparisons, where $\ell$ is the number of matches lost by the champion. We then present an asymptotically-optimal deterministic algorithm matching this lower bound without knowing $\ell$, and we extend our analysis to three variants of the problem. Lastly, we conduct a comprehensive experimental assessment of the proposed algorithms on a question answering task on public data. Results show that our proposed algorithms speed up the retrieval of the champion up to $13\times$ with respect to the state-of-the-art algorithm that perform the full tournament.

Francesco Busolin, Claudio Lucchese, Franco Maria Nardini et al.

Modern search engine ranking pipelines are commonly based on large machine-learned ensembles of regression trees. We propose LEAR, a novel - learned - technique aimed to reduce the average number of trees traversed by documents to accumulate the scores, thus reducing the overall query response time. LEAR exploits a classifier that predicts whether a document can early exit the ensemble because it is unlikely to be ranked among the final top-k results. The early exit decision occurs at a sentinel point, i.e., after having evaluated a limited number of trees, and the partial scores are exploited to filter out non-promising documents. We evaluate LEAR by deploying it in a production-like setting, adopting a state-of-the-art algorithm for ensembles traversal. We provide a comprehensive experimental evaluation on two public datasets. The experiments show that LEAR has a significant impact on the efficiency of the query processing without hindering its ranking quality. In detail, on a first dataset, LEAR is able to achieve a speedup of 3x without any loss in NDCG1@0, while on a second dataset the speedup is larger than 5x with a negligible NDCG@10 loss (< 0.05%).

Lorenzo Valerio, Franco Maria Nardini, Andrea Passarella et al.

Neural Networks (NN), although successfully applied to several Artificial Intelligence tasks, are often unnecessarily over-parametrised. In edge/fog computing, this might make their training prohibitive on resource-constrained devices, contrasting with the current trend of decentralising intelligence from remote data centres to local constrained devices. Therefore, we investigate the problem of training effective NN models on constrained devices having a fixed, potentially small, memory budget. We target techniques that are both resource-efficient and performance effective while enabling significant network compression. Our Dynamic Hard Pruning (DynHP) technique incrementally prunes the network during training, identifying neurons that marginally contribute to the model accuracy. DynHP enables a tunable size reduction of the final neural network and reduces the NN memory occupancy during training. Freed memory is reused by a \emph{dynamic batch sizing} approach to counterbalance the accuracy degradation caused by the hard pruning strategy, improving its convergence and effectiveness. We assess the performance of DynHP through reproducible experiments on three public datasets, comparing them against reference competitors. Results show that DynHP compresses a NN up to $10$ times without significant performance drops (up to $3.5\%$ additional error w.r.t. the competitors), reducing up to $80\%$ the training memory occupancy.

Claudio Lucchese, Franco Maria Nardini, Salvatore Orlando et al.

Search engine ranking pipelines are commonly based on large ensembles of machine-learned decision trees. The tight constraints on query response time recently motivated researchers to investigate algorithms to make faster the traversal of the additive ensemble or to early terminate the evaluation of documents that are unlikely to be ranked among the top-k. In this paper, we investigate the novel problem of \textit{query-level early exiting}, aimed at deciding the profitability of early stopping the traversal of the ranking ensemble for all the candidate documents to be scored for a query, by simply returning a ranking based on the additive scores computed by a limited portion of the ensemble. Besides the obvious advantage on query latency and throughput, we address the possible positive impact of query-level early exiting on ranking effectiveness. To this end, we study the actual contribution of incremental portions of the tree ensemble to the ranking of the top-k documents scored for a given query. Our main finding is that queries exhibit different behaviors as scores are accumulated during the traversal of the ensemble and that query-level early stopping can remarkably improve ranking quality. We present a reproducible and comprehensive experimental evaluation, conducted on two public datasets, showing that query-level early exiting achieves an overall gain of up to 7.5% in terms of NDCG@10 with a speedup of the scoring process of up to 2.2x.

Sean MacAvaney, Franco Maria Nardini, Raffaele Perego et al.

In precision-oriented tasks like answer ranking, it is more important to rank many relevant answers highly than to retrieve all relevant answers. It follows that a good ranking strategy would be to learn how to identify the easiest correct answers first (i.e., assign a high ranking score to answers that have characteristics that usually indicate relevance, and a low ranking score to those with characteristics that do not), before incorporating more complex logic to handle difficult cases (e.g., semantic matching or reasoning). In this work, we apply this idea to the training of neural answer rankers using curriculum learning. We propose several heuristics to estimate the difficulty of a given training sample. We show that the proposed heuristics can be used to build a training curriculum that down-weights difficult samples early in the training process. As the training process progresses, our approach gradually shifts to weighting all samples equally, regardless of difficulty. We present a comprehensive evaluation of our proposed idea on three answer ranking datasets. Results show that our approach leads to superior performance of two leading neural ranking architectures, namely BERT and ConvKNRM, using both pointwise and pairwise losses. When applied to a BERT-based ranker, our method yields up to a 4% improvement in MRR and a 9% improvement in P@1 (compared to the model trained without a curriculum). This results in models that can achieve comparable performance to more expensive state-of-the-art techniques.

Sean MacAvaney, Franco Maria Nardini, Raffaele Perego et al.

Deep pretrained transformer networks are effective at various ranking tasks, such as question answering and ad-hoc document ranking. However, their computational expenses deem them cost-prohibitive in practice. Our proposed approach, called PreTTR (Precomputing Transformer Term Representations), considerably reduces the query-time latency of deep transformer networks (up to a 42x speedup on web document ranking) making these networks more practical to use in a real-time ranking scenario. Specifically, we precompute part of the document term representations at indexing time (without a query), and merge them with the query representation at query time to compute the final ranking score. Due to the large size of the token representations, we also propose an effective approach to reduce the storage requirement by training a compression layer to match attention scores. Our compression technique reduces the storage required up to 95% and it can be applied without a substantial degradation in ranking performance.

Sean MacAvaney, Franco Maria Nardini, Raffaele Perego et al.

The identification of relevance with little textual context is a primary challenge in passage retrieval. We address this problem with a representation-based ranking approach that: (1) explicitly models the importance of each term using a contextualized language model; (2) performs passage expansion by propagating the importance to similar terms; and (3) grounds the representations in the lexicon, making them interpretable. Passage representations can be pre-computed at index time to reduce query-time latency. We call our approach EPIC (Expansion via Prediction of Importance with Contextualization). We show that EPIC significantly outperforms prior importance-modeling and document expansion approaches. We also observe that the performance is additive with the current leading first-stage retrieval methods, further narrowing the gap between inexpensive and cost-prohibitive passage ranking approaches. Specifically, EPIC achieves a MRR@10 of 0.304 on the MS-MARCO passage ranking dataset with 78ms average query latency on commodity hardware. We also find that the latency is further reduced to 68ms by pruning document representations, with virtually no difference in effectiveness.