arXiv subjects

Erik Vee

Publications and source records attributed to Erik Vee.

15 recordsLinked to original sources

arXiv2025

GUIDE: Guided Initialization and Distillation of Embeddings

Algorithmic efficiency techniques such as distillation (\cite{hinton2015distillation}) are useful in improving model quality without increasing serving costs, provided a larger teacher model is available for a smaller student model to learn from during training. Standard distillation methods are limited to only forcing the student to match the teacher's outputs. Given the costs associated with training a large model, we believe we should be extracting more useful information from a teacher model than by just making the student match the teacher's outputs. In this paper, we introduce \guide (Guided Initialization and Distillation of Embeddings). \guide can be considered a distillation technique that forces the student to match the teacher in the parameter space. Using \guide we show 25-26\% reduction in the teacher-student quality gap when using large student models (400M - 1B parameters) trained on $\approx$ 20B tokens. We also present a thorough analysis demonstrating that \guide can be combined with knowledge distillation with near additive improvements. Furthermore, we show that applying \guide alone leads to substantially better model quality than applying knowledge distillation by itself. Most importantly, \guide introduces no training or inference overhead and hence any model quality gains from our method are virtually free.

cs.LG↗

arXiv2024

Linear Projections of Teacher Embeddings for Few-Class Distillation

Knowledge Distillation (KD) has emerged as a promising approach for transferring knowledge from a larger, more complex teacher model to a smaller student model. Traditionally, KD involves training the student to mimic the teacher's output probabilities, while more advanced techniques have explored guiding the student to adopt the teacher's internal representations. Despite its widespread success, the performance of KD in binary classification and few-class problems has been less satisfactory. This is because the information about the teacher model's generalization patterns scales directly with the number of classes. Moreover, several sophisticated distillation methods may not be universally applicable or effective for data types beyond Computer Vision. Consequently, effective distillation techniques remain elusive for a range of key real-world applications, such as sentiment analysis, search query understanding, and advertisement-query relevance assessment. Taking these observations into account, we introduce a novel method for distilling knowledge from the teacher's model representations, which we term Learning Embedding Linear Projections (LELP). Inspired by recent findings about the structure of final-layer representations, LELP works by identifying informative linear subspaces in the teacher's embedding space, and splitting them into pseudo-subclasses. The student model is then trained to replicate these pseudo-classes. Our experimental evaluation on large-scale NLP benchmarks like Amazon Reviews and Sentiment140 demonstrate the LELP is consistently competitive with, and typically superior to, existing state-of-the-art distillation algorithms for binary and few-class problems, where most KD methods suffer.

cs.LG↗

arXiv2023

New Tools for Peak Memory Scheduling

We study scheduling of computation graphs to minimize peak memory consumption, an increasingly critical task due to the surge in popularity of large deep-learning models. This problem corresponds to the weighted version of the classical one-shot black pebbling game. We propose the notion of a dominant schedule to capture the idea of finding the ``best'' schedule for a subgraph and introduce new tools to compute and utilize dominant schedules. Surprisingly, we show that despite the strong requirements, a dominant schedule exists for any computation graph; and, moreover, that it is possible to compute the dominant schedule efficiently whenever we can find optimal schedules efficiently for a particular class of graphs (under mild technical conditions). We apply these new tools to analyze trees and series-parallel graphs. We show that the weighted one-shot black pebbling game is strongly NP-complete even when the graph is an out-tree -- or simpler still, a pumpkin, one of the simplest series-parallel graphs. On the positive side, we design a fixed-parameter tractable algorithm to find a dominant schedule (hence also a peak memory minimizing schedule) for series-parallel graphs when parameterized by the out-degree. This algorithm runs in time $2^{O(d \log d)} \cdot poly(n)$ for series-parallel graphs with $n$ nodes and maximum out-degree $d$; for pumpkins, we can improve the dependence on $d$ to $O(2^d \cdot poly(n))$.

cs.DS↗

arXiv2023

Efficient Caching with Reserves via Marking

Online caching is among the most fundamental and well-studied problems in the area of online algorithms. Innovative algorithmic ideas and analysis -- including potential functions and primal-dual techniques -- give insight into this still-growing area. Here, we introduce a new analysis technique that first uses a potential function to upper bound the cost of an online algorithm and then pairs that with a new dual-fitting strategy to lower bound the cost of an offline optimal algorithm. We apply these techniques to the Caching with Reserves problem recently introduced by Ibrahimpur et al. [10] and give an O(log k)-competitive fractional online algorithm via a marking strategy, where k denotes the size of the cache. We also design a new online rounding algorithm that runs in polynomial time to obtain an O(log k)-competitive randomized integral algorithm. Additionally, we provide a new, simple proof for randomized marking for the classical unweighted paging problem.

cs.DS↗

arXiv2023

LayerNAS: Neural Architecture Search in Polynomial Complexity

Neural Architecture Search (NAS) has become a popular method for discovering effective model architectures, especially for target hardware. As such, NAS methods that find optimal architectures under constraints are essential. In our paper, we propose LayerNAS to address the challenge of multi-objective NAS by transforming it into a combinatorial optimization problem, which effectively constrains the search complexity to be polynomial. For a model architecture with $L$ layers, we perform layerwise-search for each layer, selecting from a set of search options $\mathbb{S}$. LayerNAS groups model candidates based on one objective, such as model size or latency, and searches for the optimal model based on another objective, thereby splitting the cost and reward elements of the search. This approach limits the search complexity to $ O(H \cdot |\mathbb{S}| \cdot L) $, where $H$ is a constant set in LayerNAS. Our experiments show that LayerNAS is able to consistently discover superior models across a variety of search spaces in comparison to strong baselines, including search spaces derived from NATS-Bench, MobileNetV2 and MobileNetV3.

cs.LG↗

arXiv2023

SLaM: Student-Label Mixing for Distillation with Unlabeled Examples

Knowledge distillation with unlabeled examples is a powerful training paradigm for generating compact and lightweight student models in applications where the amount of labeled data is limited but one has access to a large pool of unlabeled data. In this setting, a large teacher model generates ``soft'' pseudo-labels for the unlabeled dataset which are then used for training the student model. Despite its success in a wide variety of applications, a shortcoming of this approach is that the teacher's pseudo-labels are often noisy, leading to impaired student performance. In this paper, we present a principled method for knowledge distillation with unlabeled examples that we call Student-Label Mixing (SLaM) and we show that it consistently improves over prior approaches by evaluating it on several standard benchmarks. Finally, we show that SLaM comes with theoretical guarantees; along the way we give an algorithm improving the best-known sample complexity for learning halfspaces with margin under random classification noise, and provide the first convergence analysis for so-called ``forward loss-adjustment" methods.

cs.LG↗

arXiv2023

Robust Active Distillation

Distilling knowledge from a large teacher model to a lightweight one is a widely successful approach for generating compact, powerful models in the semi-supervised learning setting where a limited amount of labeled data is available. In large-scale applications, however, the teacher tends to provide a large number of incorrect soft-labels that impairs student performance. The sheer size of the teacher additionally constrains the number of soft-labels that can be queried due to prohibitive computational and/or financial costs. The difficulty in achieving simultaneous \emph{efficiency} (i.e., minimizing soft-label queries) and \emph{robustness} (i.e., avoiding student inaccuracies due to incorrect labels) hurts the widespread application of knowledge distillation to many modern tasks. In this paper, we present a parameter-free approach with provable guarantees to query the soft-labels of points that are simultaneously informative and correctly labeled by the teacher. At the core of our work lies a game-theoretic formulation that explicitly considers the inherent trade-off between the informativeness and correctness of input instances. We establish bounds on the expected performance of our approach that hold even in worst-case distillation instances. We present empirical evaluations on popular benchmarks that demonstrate the improved distillation performance enabled by our work relative to that of state-of-the-art active learning and active distillation methods.

cs.LG↗

arXiv2022

Weighted Distillation with Unlabeled Examples

Distillation with unlabeled examples is a popular and powerful method for training deep neural networks in settings where the amount of labeled data is limited: A large ''teacher'' neural network is trained on the labeled data available, and then it is used to generate labels on an unlabeled dataset (typically much larger in size). These labels are then utilized to train the smaller ''student'' model which will actually be deployed. Naturally, the success of the approach depends on the quality of the teacher's labels, since the student could be confused if trained on inaccurate data. This paper proposes a principled approach for addressing this issue based on a ''debiasing'' reweighting of the student's loss function tailored to the distillation training paradigm. Our method is hyper-parameter free, data-agnostic, and simple to implement. We demonstrate significant improvements on popular academic datasets and we accompany our results with a theoretical analysis which rigorously justifies the performance of our method in certain settings.

cs.LG↗

arXiv2022

Caching with Reserves

Caching is a crucial component of many computer systems, so naturally it is a well-studied topic in algorithm design. Much of traditional caching research studies cache management for a single-user or single-processor environment. In this paper, we propose two related generalizations of the classical caching problem that capture issues that arise in a multi-user or multi-processor environment. In the caching with reserves problem, a caching algorithm is required to maintain at least $k_i$ pages belonging to user $i$ in the cache at any time, for some given reserve capacities $k_i$. In the public-private caching problem, the cache of total size $k$ is partitioned into subcaches, a private cache of size $k_i$ for each user $i$ and a shared public cache usable by any user. In both of these models, as in the classical caching framework, the objective of the algorithm is to dynamically maintain the cache so as to minimize the total number of cache misses. We show that caching with reserves and public-private caching models are equivalent up to constant factors, and thus focus on the former. Unlike classical caching, both of these models turn out to be NP-hard even in the offline setting, where the page sequence is known in advance. For the offline setting, we design a 2-approximation algorithm, whose analysis carefully keeps track of a potential function to bound the cost. In the online setting, we first design an $O(\ln k)$-competitive fractional algorithm using the primal-dual framework, and then show how to convert it online to a randomized integral algorithm with the same guarantee.

cs.DS↗

arXiv2022

Scheduling with Communication Delay in Near-Linear Time

We consider the problem of efficiently scheduling jobs with precedence constraints on a set of identical machines in the presence of a uniform communication delay. Such precedence-constrained jobs can be modeled as a directed acyclic graph, $G = (V, E)$. In this setting, if two precedence-constrained jobs $u$ and $v$, with $v$ dependent on $u$ ($u \prec v$), are scheduled on different machines, then $v$ must start at least $ρ$ time units after $u$ completes. The scheduling objective is to minimize makespan, i.e. the total time from when the first job starts to when the last job finishes. The focus of this paper is to provide an efficient approximation algorithm with near-linear running time. We build on the algorithm of Lepere and Rapine [STACS 2002] for this problem to give an $O\left(\frac{\ln ρ}{\ln \ln ρ} \right)$-approximation algorithm that runs in $\tilde{O}(|V| + |E|)$ time.

cs.DS↗

arXiv2021

Learning-Augmented Weighted Paging

We consider a natural semi-online model for weighted paging, where at any time the algorithm is given predictions, possibly with errors, about the next arrival of each page. The model is inspired by Belady's classic optimal offline algorithm for unweighted paging, and extends the recently studied model for learning-augmented paging (Lykouris and Vassilvitskii, 2018) to the weighted setting. For the case of perfect predictions, we provide an $\ell$-competitive deterministic and an $O(\log \ell)$-competitive randomized algorithm, where $\ell$ is the number of distinct weight classes. Both these bounds are tight, and imply an $O(\log W)$- and $O(\log \log W)$-competitive ratio, respectively, when the page weights lie between $1$ and $W$. Previously, it was not known how to use these predictions in the weighted setting and only bounds of $k$ and $O(\log k)$ were known, where $k$ is the cache size. Our results also generalize to the interleaved paging setting and to the case of imperfect predictions, with the competitive ratios degrading smoothly from $O(\ell)$ and $O(\log \ell)$ to $O(k)$ and $O(\log k)$, respectively, as the prediction error increases. Our results are based on several insights on structural properties of Belady's algorithm and the sequence of page arrival predictions, and novel potential functions that incorporate these predictions. For the case of unweighted paging, the results imply a very simple potential function based proof of the optimality of Belady's algorithm, which may be of independent interest.

cs.DS↗

arXiv2019

Semi-Online Bipartite Matching

In this paper we introduce the \emph{semi-online} model that generalizes the classical online computational model. The semi-online model postulates that the unknown future has a predictable part and an adversarial part; these parts can be arbitrarily interleaved. An algorithm in this model operates as in the standard online model, i.e., makes an irrevocable decision at each step. We consider bipartite matching in the semi-online model, for both integral and fractional cases. Our main contributions are competitive algorithms for this problem that are close to or match a hardness bound. The competitive ratio of the algorithms nicely interpolates between the truly offline setting (no adversarial part) and the truly online setting (no predictable part).

cs.DS↗

arXiv2012

Ad Serving Using a Compact Allocation Plan

A large fraction of online display advertising is sold via guaranteed contracts: a publisher guarantees to the advertiser a certain number of user visits satisfying the targeting predicates of the contract. The publisher is then tasked with solving the ad serving problem - given a user visit, which of the thousands of matching contracts should be displayed, so that by the expiration time every contract has obtained the requisite number of user visits. The challenges of the problem come from (1) the sheer size of the problem being solved, with tens of thousands of contracts and billions of user visits, (2) the unpredictability of user behavior, since these contracts are sold months ahead of time, when only a forecast of user visits is available and (3) the minute amount of resources available online, as an ad server must respond with a matching contract in a fraction of a second. We present a solution to the guaranteed delivery ad serving problem using {\em compact allocation plans}. These plans, computed offline, can be efficiently queried by the ad server during an ad call; they are small, using only O(1) space for contract; and are stateless, allowing for distributed serving without any central coordination. We evaluate this approach on a real set of user visits and guaranteed contracts and show that the compact allocation plans are an effective way of solving the guaranteed delivery ad serving problem.

cs.DS↗

arXiv2012

SHALE: An Efficient Algorithm for Allocation of Guaranteed Display Advertising

Motivated by the problem of optimizing allocation in guaranteed display advertising, we develop an efficient, lightweight method of generating a compact {\em allocation plan} that can be used to guide ad server decisions. The plan itself uses just O(1) state per guaranteed contract, is robust to noise, and allows us to serve (provably) nearly optimally. The optimization method we develop is scalable, with a small in-memory footprint, and working in linear time per iteration. It is also "stop-anytime", meaning that time-critical applications can stop early and still get a good serving solution. Thus, it is particularly useful for optimizing the large problems arising in the context of display advertising. We demonstrate the effectiveness of our algorithm using actual Yahoo! data.

cs.DS↗

arXiv2010

Inventory Allocation for Online Graphical Display Advertising

We discuss a multi-objective/goal programming model for the allocation of inventory of graphical advertisements. The model considers two types of campaigns: guaranteed delivery (GD), which are sold months in advance, and non-guaranteed delivery (NGD), which are sold using real-time auctions. We investigate various advertiser and publisher objectives such as (a) revenue from the sale of impressions, clicks and conversions, (b) future revenue from the sale of NGD inventory, and (c) "fairness" of allocation. While the first two objectives are monetary, the third is not. This combination of demand types and objectives leads to potentially many variations of our model, which we delineate and evaluate. Our experimental results, which are based on optimization runs using real data sets, demonstrate the effectiveness and flexibility of the proposed model.

cs.CE↗