arXiv subjects

Ali Kayyam

Publications and source records attributed to Ali Kayyam.

3 recordsLinked to original sources

arXiv2026

Sticky Routing: Training MoE Models for Memory-Efficient Inference

Mixture-of-Experts (MoE) models activate only a sparse subset of experts per token, yet consecutive tokens frequently activate different experts -- causing constant weight swapping between slow storage and fast memory on edge devices. Existing remedies are either system-level (caching heuristics) or post-hoc (router fine-tuning), leaving the root cause unchanged during pretraining. We propose StickyMoE, a differentiable routing consistency loss that penalises abrupt expert switches between adjacent tokens, encouraging the router to maintain the same expert assignment across semantically coherent spans. StickyMoE requires no architectural changes, adds a single hyperparameter lambda, and unlike post-hoc methods, allows expert representations and routing decisions to co-adapt from the first training step. Experiments on small-scale MoE language models show that StickyMoE reduces the expert switch rate by up to 60% with less than 4% perplexity degradation, Pareto-dominating post-hoc fine-tuning on the quality-locality frontier. Routing temporal locality is most efficiently instilled at training time.

cs.LG↗

arXiv2026

Do Transformers Need Three Projections? Systematic Study of QKV Variants

Transformers have become the standard solution for various AI tasks, with the query, key, and value (QKV) attention formulation playing a central role. However, the individual contribution of these three projections and the impact of omitting some remain poorly understood. We systematically evaluate three projection sharing constraints: a) Q-K=V (shared key-value), b) Q=K-V (shared query-key), and c) Q=K=V (single projection). The last two variants produce symmetric attention maps; to address this, we also explore asymmetric attention via 2D positional encodings. Through experiments spanning synthetic tasks, vision (MNIST, CIFAR, TinyImageNet, anomaly), and language modeling (300M and 1.2B parameter models on 10B tokens), we discovered that our transformers perform on par or occasionally better than the QKV transformer. In language modeling, Q-K=V projection sharing achieves 50% KV cache reduction with only 3.1% perplexity degradation. Crucially, projection sharing is complementary to head sharing (GQA/MQA): combining Q-K=V with GQA-4 yields 87.5% cache reduction, while Q-K=V + MQA achieves 96.9%, enabling practical on-device inference. We show that Q-K=V preserves quality because keys and values can occupy similar representational spaces and attention operates in a low-rank regime, whereas Q=K-V breaks attention directionality. Our results systematically characterize projection sharing as an underexplored instance of weight tying in attention, with direct, quantifiable inference memory benefits, particularly valuable for edge deployment. The code is publicly available at https://github.com/Brainchip-Inc/Do-Transformers-Need-3-Projections

cs.LG↗

arXiv2025

Provenance Networks: End-to-End Exemplar-Based Explainability

We introduce provenance networks, a novel class of neural models designed to provide end-to-end, training-data-driven explainability. Unlike conventional post-hoc methods, provenance networks learn to link each prediction directly to its supporting training examples as part of the model's normal operation, embedding interpretability into the architecture itself. Conceptually, the model operates similarly to a learned KNN, where each output is justified by concrete exemplars weighted by relevance in the feature space. This approach facilitates systematic investigations of the trade-off between memorization and generalization, enables verification of whether a given input was included in the training set, aids in the detection of mislabeled or anomalous data points, enhances resilience to input perturbations, and supports the identification of similar inputs contributing to the generation of a new data point. By jointly optimizing the primary task and the explainability objective, provenance networks offer insights into model behavior that traditional deep networks cannot provide. While the model introduces additional computational cost and currently scales to moderately sized datasets, it provides a complementary approach to existing explainability techniques. In particular, it addresses critical challenges in modern deep learning, including model opaqueness, hallucination, and the assignment of credit to data contributors, thereby improving transparency, robustness, and trustworthiness in neural models.

cs.CV↗