Hidden Winning Tickets in Transformer Attention

Kari Jaaskelainen

08 Nov 2025 — 1 min read

Ever heard of the “lottery ticket” idea in AI? It says big neural nets hide small subnetworks that can perform just as well. This paper proves a strong version of that for the heart of Transformers: multi-head attention (MHA).

The big claim: Inside a randomly initialized MHA, there exists a small subnetwork that can closely imitate any MHA with the same input size—provided the hidden size is large enough (grows with input dimension and number of heads).
Beyond attention: Using this result, the authors extend the strong lottery ticket theory to entire Transformers without normalization layers.
Evidence: Experiments show the approximation error shrinks exponentially as the hidden size increases.

Why it matters: if “winning tickets” are guaranteed to exist in attention, we can prune or sparsely train large models more confidently, aiming for smaller, faster, and cheaper Transformers without sacrificing accuracy.

Paper: http://arxiv.org/abs/2511.04217v1

Paper: http://arxiv.org/abs/2511.04217v1

Register: https://www.AiFeta.com

AI MachineLearning Transformers DeepLearning NeuralNetworks EfficientAI ModelCompression Research

Automating GDPR Compliance: A Roadmap for Companies and Law Firms

GDPR compliance is more than checkboxes. A new roadmap from the Privatech project shows how automation and machine learning can help companies and law firms assess—and even generate—privacy compliance. * Shift the focus to data processors’ real workflows: drafting policies, mapping data uses, documenting decisions. * Break compliance into machine-ready

FPGAs for Faster, Leaner Deep Learning: A Review of CNN Accelerators

Deep learning drives image search, robots, and medical scans. Most systems lean on CPUs and GPUs. This review asks: what if we run convolutional neural networks (CNNs) on FPGAs—reconfigurable chips you can tailor to the model? * Why FPGAs: custom dataflows, low latency, and strong energy efficiency—great for cameras,

Dynamic-K: Recommendations That Know When to Stop

Most apps show a fixed number of “top” items—say 10 movies or 20 products—assuming there are always enough good options. But that’s not always true: sometimes there are few relevant items, or some users are extra picky. The result? Filler recommendations. Dynamic-K flips the script. Instead of

Teaching chatbots to stop contradicting themselves (DECODE)

Teaching chatbots to stop contradicting themselves Ever had a bot say one thing, then the opposite a few turns later? This study introduces DECODE—a new task and dataset for spotting contradictions in everyday conversations, drawn from both human-human and human-bot chats. * New data beats existing natural language inference (NLI)

Read more

Automating GDPR Compliance: A Roadmap for Companies and Law Firms

FPGAs for Faster, Leaner Deep Learning: A Review of CNN Accelerators

Dynamic-K: Recommendations That Know When to Stop

Teaching chatbots to stop contradicting themselves (DECODE)