Searcharxiv⌕ Search

arXiv subjects

Yani Ioannou

Publications and source records attributed to Yani Ioannou.

At least 19 recordsLinked to original sources

The Surprising Effectiveness of Deleting Weights in LLM Post-Training

How little of a pretrained large language model (LLM) has to change for it to acquire new capabilities? We find that zeroing fewer than 0.05% of its weights, with no other modification, matches full fine-tuning (FFT) at scale across three LLM post-training regimes: reinforcement learning with verifiable rewards, supervised fine-tuning, and on-policy distillation. Our method, Deletion Tuning, learns a binary mask over the gate projection of transformer blocks. Two practical advantages follow: the trained mask is hundreds to thousands of times smaller than the corresponding FFT checkpoint, and Deletion Tuning forgets far less, which addresses a central challenge in continual learning. Under reinforcement learning, its prior-task accuracy stays near the base level while FFT, LoRA, and random sparse fine-tuning lose up to 5.4 points. To understand why masking alone suffices, we analyze the geometry of the learned weight changes under reinforcement learning using stable rank, spectrum drift, and principal-weight overlap. Compared with LoRA and random sparse fine-tuning, Deletion Tuning is closest to FFT across the three measures.

cs.LG↗

BMuon: Budgeting Muon's Spectral Geometry for Structured Compressibility

Muon has emerged as a strong optimizer for neural network training and admits an interpretation as steepest descent under a spectral-norm geometry. Yet we find that Muon-trained Transformers can degrade sharply when their MLP hidden channels are pruned after training. To reconcile Muon's optimization performance with structured compressibility, we introduce Budget Muon (BMuon), which modifies Muon's spectral geometry so that hidden channels across blocks compete for a shared allocation budget. The resulting oracle is exactly normalized steepest descent under a family of norms that interpolate between Muon's geometry and structured sparse geometries. Our principal sparsity control is an interpretable global channel-allocation budget, rather than a setting-dependent sparsity-penalty coefficient. For linear factorization models, we show that Muon admits minimum-cost exact representations with redundant hidden width, whereas under the BMuon oracle geometry and a sufficiently scarce budget, every minimum-cost representation is width-minimal. For practical training, we derive a minorization-maximization surrogate for the exact allocation problem. On ImageNet-1K with ViT-S/16, BMuon improves over Muon under gradual structured MLP pruning, and its learned allocations provide a more effective pruning criterion than weight magnitude. On a 1B-parameter LLaMA-style model trained on FineWeb, BMuon improves over Muon under post-training MLP channel pruning.

cs.LG↗

Stress-Testing Efficient Responsible-AI Evaluation: When Compute Savings Change Benchmark Conclusions

Efficient evaluation changes the protocol used to support claims about model behavior, yet it is rarely tested whether those claims remain stable after the evaluation itself is made cheaper. We stress-test conclusion robustness in responsible-AI benchmarking by evaluating three dense and mixture-of-experts models on BBQ and BBQ-V under seven conditions spanning batching, quantization, benchmark reduction, and their combinations. Rather than treating preserved aggregate accuracy as sufficient, we compare accuracy, bias severity and prevalence, reasoning quality, subgroup behavior, subset-membership stability, runtime, and measured GPU energy against a full-benchmark BF16 baseline. Larger batching keeps accuracy within 0.35 percentage points of baseline and produces comparatively small subgroup changes, while reducing energy in five of six model--dataset settings. INT8 largely preserves quality but uses 1.79--4.26$\times$ baseline energy. INT4 causes larger, model- and context-dependent changes. Reduced benchmarks provide the most consistent savings, but very small subsets are substantially more sensitive to which items are retained. Efficient evaluation should therefore be treated as a measurement intervention whose validity must be checked across the conclusions the benchmark is intended to support. Our project website is https://vectorinstitute.github.io/sustainable-rai-evaluation/ and the code is available at https://github.com/VectorInstitute/sustainable-rai-evaluation.

cs.LG↗

Learning Fine-grained Parameter Sharing via Sparse Tensor Decomposition

Large neural networks achieve state-of-the-art performance on many tasks, yet their sheer size hinders deployment on resource-constrained devices. Among existing compression approaches, cross-layer parameter sharing remains relatively unexplored for transformer models. In this paper, we introduce Fine-grained Parameter Sharing (FiPS), a unified framework for compressing transformer Multi-Layer Perceptrons (MLPs) that combines cross-block parameter sharing, low-rank factorization, and sparsity in a single optimization. FiPS concatenates MLP weight matrices across a group of transformer blocks and factorizes them into a shared basis and sparse, layer-specific projection matrices. Both factors are initialized via singular value decomposition (SVD) and jointly optimized by block-wise reconstruction error minimization. FiPS compresses Vision Transformers (ViTs) by up to 33% with less than 1% top-1 accuracy loss on ImageNet-1k, and by up to 57% when combined with fine-tuning. It also compresses Large Language Models (LLMs) by up to 20% while outperforming existing SVD-based methods in perplexity and downstream benchmarks at matched compression. Combined with Quantization-Aware Training (QAT), 3-bit FiPS on Gemma-2-2B achieves lower perplexity than 2-bit QAT alone while matching the same 8x compression. These results establish fine-grained parameter sharing as a practical and effective approach for transformer MLP compression.

cs.LG↗

SparseOpt: Addressing Normalization-induced Gradient Skew in Sparse Training

Dynamic Sparse Training (DST) methods train neural networks by maintaining sparsity while dynamically adapting the network topology. Despite the promise of reduced computation, DST methods converge significantly slower than dense training, often requiring comparable training time to achieve similar accuracy. We demonstrate both analytically and empirically that Batch Normalization (BN) adversely affects sparse training, and propose SparseOpt, a sparsity-aware optimizer, to address this. Experiments on ResNet models across CIFAR-100 and ImageNet demonstrate consistently faster convergence and improved generalization with our proposed method. Our work highlights the limitations of current normalization layers in sparse training and provides the first systematic study of the interaction between Batch Normalization, sparse layers, and DST, taking a significant step toward making DST practically competitive with dense training.

cs.LG↗

REAP the Experts: Why Pruning Prevails for One-Shot MoE compression

Sparsely-activated Mixture-of-Experts (SMoE) models offer efficient pre-training and low latency but their large parameter counts create significant memory overhead, motivating research into expert compression. Contrary to recent findings favouring expert merging on discriminative benchmarks, we find that expert pruning is a superior strategy for generative tasks. We demonstrate that existing merging techniques introduce an irreducible error due to the loss of fine-grained routing control over experts. Leveraging this insight, we propose Router-weighted Expert Activation Pruning (REAP), a novel pruning criterion that considers both router gate-values and expert activation norms to minimize the reconstruction error bound. Across a diverse set of SMoE models ranging from 20B to 1T parameters, REAP consistently outperforms merging and other pruning methods on generative benchmarks, especially at 50% compression. Notably, our method achieves near-lossless compression on code generation tasks with Qwen3-Coder-480B and Kimi-K2, even after pruning 50% of experts.

cs.LG↗

Reducing Hallucinations in LLMs via Factuality-Aware Preference Learning

Preference alignment methods such as RLHF and Direct Preference Optimization (DPO) improve instruction following, but they can also reinforce hallucinations when preference judgments reward fluency and confidence over factual correctness. We introduce F-DPO (Factuality-aware Direct Preference Optimization), a simple extension of DPO that uses only binary factuality labels. F-DPO (i) applies a label-flipping transformation that corrects misordered preference pairs so the chosen response is never less factual than the rejected one, and (ii) adds a factuality-aware margin that emphasizes pairs with clear correctness differences, while reducing to standard DPO when both responses share the same factuality. We construct factuality-aware preference data by augmenting DPO pairs with binary factuality indicators and synthetic hallucinated variants. Across seven open-weight LLMs (1B-14B), F-DPO consistently improves factuality and reduces hallucination rates relative to both base models and standard DPO. On Qwen3-8B, F-DPO reduces hallucination rates by 5x(from 0.424 to 0.084) while improving factuality scores by 50% (from 5.26 to 7.90). F-DPO also generalizes to out-of-distribution benchmarks: on TruthfulQA, Qwen2.5-14B achieves +17% MC1 accuracy (0.500 to 0.585) and +49% MC2 accuracy (0.357 to 0.531). F-DPO requires no auxiliary reward model, token-level annotations, or multi-stage training.

cs.CL↗

Sparse Training from Random Initialization: Aligning Lottery Ticket Masks using Weight Symmetry

The Lottery Ticket Hypothesis (LTH) suggests there exists a sparse LTH mask and weights that achieve the same generalization performance as the dense model while using significantly fewer parameters. However, finding a LTH solution is computationally expensive, and a LTH sparsity mask does not generalize to other random weight initializations. Recent work has suggested that neural networks trained from random initialization find solutions within the same basin modulo permutation, and proposes a method to align trained models within the same loss basin. We hypothesize that misalignment of basins is the reason why LTH masks do not generalize to new random initializations and propose permuting the LTH mask to align with the new optimization basin when performing sparse training from a different random init. We empirically show a significant increase in generalization when sparse training from random initialization with the permuted mask as compared to using the non-permuted LTH mask, on multiple datasets (CIFAR-10, CIFAR-100 and ImageNet) and models (VGG11, ResNet20 and ResNet50).

cs.LG↗

SD$^2$: Self-Distilled Sparse Drafters

Speculative decoding is a powerful technique for reducing the latency of Large Language Models (LLMs), offering a fault-tolerant framework that enables the use of highly compressed draft models. In this work, we introduce Self-Distilled Sparse Drafters (SD$^2$), a novel methodology that leverages self-data distillation and fine-grained weight sparsity to produce highly efficient and well-aligned draft models. SD$^2$ systematically enhances draft token acceptance rates while significantly reducing Multiply-Accumulate operations (MACs), even in the Universal Assisted Generation (UAG) setting, where draft and target models originate from different model families. On a Llama-3.1-70B target model, SD$^2$ provides a 1.59$\times$ higher Mean Accepted Length (MAL) compared to layer-pruned draft models and reduces MACs by over 43.87% with a 8.36% reduction in MAL compared to a dense draft models. Our 1.5B and 3B unstructured sparse drafters outperform both dense and layer-pruned models in terms of end-to-end latency improvements; highlighting the potential of sparsity-aware fine-tuning and compression strategies to improve LLM inference efficiency while maintaining alignment with target models.

cs.CL↗

What is Left After Distillation? How Knowledge Transfer Impacts Fairness and Bias

Knowledge Distillation is a commonly used Deep Neural Network (DNN) compression method, which often maintains overall generalization performance. However, we show that even for balanced image classification datasets, such as CIFAR-100, Tiny ImageNet and ImageNet, as many as 41% of the classes are statistically significantly affected by distillation when comparing class-wise accuracy (i.e. class bias) between a teacher/distilled student or distilled student/non-distilled student model. Changes in class bias are not necessarily an undesirable outcome when considered outside of the context of a model's usage. Using two common fairness metrics, Demographic Parity Difference (DPD) and Equalized Odds Difference (EOD) on models trained with the CelebA, Trifeature, and HateXplain datasets, our results suggest that increasing the distillation temperature improves the distilled student model's fairness, and the distilled student fairness can even surpass the fairness of the teacher model at high temperatures. Additionally, we examine individual fairness, ensuring similar instances receive similar predictions. Our results confirm that higher temperatures also improve the distilled student model's individual fairness. This study highlights the uneven effects of distillation on certain classes and its potentially significant role in fairness, emphasizing that caution is warranted when using distilled models for sensitive application domains.

cs.LG↗

Navigating Extremes: Dynamic Sparsity in Large Output Spaces

In recent years, Dynamic Sparse Training (DST) has emerged as an alternative to post-training pruning for generating efficient models. In principle, DST allows for a more memory efficient training process, as it maintains sparsity throughout the entire training run. However, current DST implementations fail to capitalize on this in practice. Because sparse matrix multiplication is much less efficient than dense matrix multiplication on GPUs, most implementations simulate sparsity by masking weights. In this paper, we leverage recent advances in semi-structured sparse training to apply DST in the domain of classification with large output spaces, where memory-efficiency is paramount. With a label space of possibly millions of candidates, the classification layer alone will consume several gigabytes of memory. Switching from a dense to a fixed fan-in sparse layer updated with sparse evolutionary training (SET); however, severely hampers training convergence, especially at the largest label spaces. We find that poor gradient flow from the sparse classifier to the dense text encoder make it difficult to learn good input representations. By employing an intermediate layer or adding an auxiliary training objective, we recover most of the generalisation performance of the dense model. Overall, we demonstrate the applicability and practical benefits of DST in a challenging domain -- characterized by a highly skewed label distribution that differs substantially from typical DST benchmark datasets -- which enables end-to-end training with millions of labels on commodity hardware.

cs.LG↗

Trustworthy and Responsible AI for Human-Centric Autonomous Decision-Making Systems

Artificial Intelligence (AI) has paved the way for revolutionary decision-making processes, which if harnessed appropriately, can contribute to advancements in various sectors, from healthcare to economics. However, its black box nature presents significant ethical challenges related to bias and transparency. AI applications are hugely impacted by biases, presenting inconsistent and unreliable findings, leading to significant costs and consequences, highlighting and perpetuating inequalities and unequal access to resources. Hence, developing safe, reliable, ethical, and Trustworthy AI systems is essential. Our team of researchers working with Trustworthy and Responsible AI, part of the Transdisciplinary Scholarship Initiative within the University of Calgary, conducts research on Trustworthy and Responsible AI, including fairness, bias mitigation, reproducibility, generalization, interpretability, and authenticity. In this paper, we review and discuss the intricacies of AI biases, definitions, methods of detection and mitigation, and metrics for evaluating bias. We also discuss open challenges with regard to the trustworthiness and widespread application of AI across diverse domains of human-centric decision making, as well as guidelines to foster Responsible and Trustworthy AI models.

cs.AI↗

Meta-GCN: A Dynamically Weighted Loss Minimization Method for Dealing with the Data Imbalance in Graph Neural Networks

Although many real-world applications, such as disease prediction, and fault detection suffer from class imbalance, most existing graph-based classification methods ignore the skewness of the distribution of classes; therefore, tend to be biased towards the majority class(es). Conventional methods typically tackle this problem through the assignment of weights to each one of the class samples based on a function of their loss, which can lead to over-fitting on outliers. In this paper, we propose a meta-learning algorithm, named Meta-GCN, for adaptively learning the example weights by simultaneously minimizing the unbiased meta-data set loss and optimizing the model weights through the use of a small unbiased meta-data set. Through experiments, we have shown that Meta-GCN outperforms state-of-the-art frameworks and other baselines in terms of accuracy, the area under the receiver operating characteristic (AUC-ROC) curve, and macro F1-Score for classification tasks on two different datasets.

cs.LG↗

Dynamic Sparse Training with Structured Sparsity

Dynamic Sparse Training (DST) methods achieve state-of-the-art results in sparse neural network training, matching the generalization of dense models while enabling sparse training and inference. Although the resulting models are highly sparse and theoretically less computationally expensive, achieving speedups with unstructured sparsity on real-world hardware is challenging. In this work, we propose a sparse-to-sparse DST method, Structured RigL (SRigL), to learn a variant of fine-grained structured N:M sparsity by imposing a constant fan-in constraint. Using our empirical analysis of existing DST methods at high sparsity, we additionally employ a neuron ablation method which enables SRigL to achieve state-of-the-art sparse-to-sparse structured DST performance on a variety of Neural Network (NN) architectures. Using a 90% sparse linear layer, we demonstrate a real-world acceleration of 3.4x/2.5x on CPU for online inference and 1.7x/13.0x on GPU for inference with a batch size of 256 when compared to equivalent dense/unstructured (CSR) sparse layers, respectively.

cs.LG↗

Bounding generalization error with input compression: An empirical study with infinite-width networks

Estimating the Generalization Error (GE) of Deep Neural Networks (DNNs) is an important task that often relies on availability of held-out data. The ability to better predict GE based on a single training set may yield overarching DNN design principles to reduce a reliance on trial-and-error, along with other performance assessment advantages. In search of a quantity relevant to GE, we investigate the Mutual Information (MI) between the input and final layer representations, using the infinite-width DNN limit to bound MI. An existing input compression-based GE bound is used to link MI and GE. To the best of our knowledge, this represents the first empirical study of this bound. In our attempt to empirically falsify the theoretical bound, we find that it is often tight for best-performing models. Furthermore, it detects randomization of training labels in many cases, reflects test-time perturbation robustness, and works well given only few training samples. These results are promising given that input compression is broadly applicable where MI can be estimated with confidence.

cs.LG↗

Monitoring Shortcut Learning using Mutual Information

The failure of deep neural networks to generalize to out-of-distribution data is a well-known problem and raises concerns about the deployment of trained networks in safety-critical domains such as healthcare, finance and autonomous vehicles. We study a particular kind of distribution shift $\unicode{x2013}$ shortcuts or spurious correlations in the training data. Shortcut learning is often only exposed when models are evaluated on real-world data that does not contain the same spurious correlations, posing a serious dilemma for AI practitioners to properly assess the effectiveness of a trained model for real-world applications. In this work, we propose to use the mutual information (MI) between the learned representation and the input as a metric to find where in training, the network latches onto shortcuts. Experiments demonstrate that MI can be used as a domain-agnostic metric for monitoring shortcut learning.

cs.LG↗

Rapid Classification of TESS Planet Candidates with Convolutional Neural Networks

Accurately and rapidly classifying exoplanet candidates from transit surveys is a goal of growing importance as the data rates from space-based survey missions increases. This is especially true for NASA's TESS mission which generates thousands of new candidates each month. Here we created the first deep learning model capable of classifying TESS planet candidates. We adapted the neural network model of Ansdell et al. (2018) to TESS data. We then trained and tested this updated model on 4 sectors of high-fidelity, pixel-level simulations data created using the Lilith simulator and processed using the full TESS SPOC pipeline. We find our model performs very well on our simulated data, with 97% average precision and 92% accuracy on planets in the 2-class model. This accuracy is also boosted by another ~4% if planets found at the wrong periods are included. We also performed 3- and 4-class classification of planets, blended & target eclipsing binaries, and non-astrophysical false positives, which have slightly lower average precision and planet accuracies, but are useful for follow-up decisions. When applied to real TESS data, 61% of TCEs coincident with currently published TOIs are recovered as planets, 4% more are suggested to be EBs, and we propose a further 200 TCEs as planet candidates.

astro-ph.EP↗

Scientific Domain Knowledge Improves Exoplanet Transit Classification with Deep Learning

Space-based missions such as Kepler, and soon TESS, provide large datasets that must be analyzed efficiently and systematically. Recent work by Shallue & Vanderburg (2018) successfully used state-of-the-art deep learning models to automatically classify Kepler transit signals as either exoplanets or false positives; our application of their model yielded 95.8% accuracy and 95.5% average precision. Here we expand upon that work by including additional scientific domain knowledge into the network architecture and input representations to significantly increase overall model performance to 97.5% accuracy and 98.0% average precision. Notably, we achieve 15-20% gains in recall for the lowest signal-to-noise transits that can correspond to rocky planets in the habitable zone. We input into the network centroid time-series information derived from Kepler data plus key stellar parameters taken from the Kepler DR25 catalogue. We also implement data augmentation techniques to alleviate model over-fitting. These improvements allow us to drastically reduce the size of the model, while still maintaining improved performance; smaller models are better for generalization, for example from Kepler to TESS data. This work illustrates the importance of including expert domain knowledge in even state-of-the-art deep learning models when applying them to scientific research problems that seek to identify weak signals in noisy data. This classification tool will be especially useful for upcoming space-based photometry missions focused on finding small planets, such as TESS and PLATO.

astro-ph.EP↗