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

Cost-effective Machine Learning Inference Offload for Edge Computing

Computing at the edge is increasingly important since a massive amount of data is generated. This poses challenges in transporting all that data to the remote data centers and cloud, where they can be processed and analyzed. On the other hand, harnessing the edge data is essential for offering data-driven and machine learning-based applications, if the challenges, such as device capabilities, connectivity, and heterogeneity can be mitigated. Machine learning applications are very compute-intensive and require processing of large amount of data. However, edge devices are often resources-constrained, in terms of compute resources, power, storage, and network connectivity. Hence, limiting their potential to run efficiently and accurately state-of-the art deep neural network (DNN) models, which are becoming larger and more complex. This paper proposes a novel offloading mechanism by leveraging installed-base on-premises (edge) computational resources. The proposed mechanism allows the edge devices to offload heavy and compute-intensive workloads to edge nodes instead of using remote cloud. Our offloading mechanism has been prototyped and tested with state-of-the art person and object detection DNN models for mobile robots and video surveillance applications. The performance shows a significant gain compared to cloud-based offloading strategies in terms of accuracy and latency.