Learning Compressed Embeddings for On-Device Inference ↗
Coauthored MLSys paper published in March 2022 introducing MEmCom, or Multi-Embedding Compression. The method constructs entity-specific embeddings on demand from a jointly trained table of shared vectors and a table containing one learned weight per entity. It addresses the memory requirements that make large recommendation models difficult to deploy on smartphones. The paper reports embedding compression of 4× to 40× across four recommendation datasets with a 4% relative loss in nDCG, plus a separate 32× RankNet compression result with a 1% relative loss in nDCG.
machinelearning.apple.com
Lightweight Safety Guardrails via Synthetic Data and RL-guided Adversarial Training ↗
Coauthored research paper submitted to arXiv on July 11, 2025. The arXiv record credits Haluk Noyan Tokgozoglu, while Apple's ICML 2025 overview credits Noyan Tokgozoglu on the same titled work and lists it among accepted workshop papers. The research combines curated synthetic data with reinforcement-learning-guided generation of adversarial examples to train small safety classifiers. The authors report content-moderation performance matching or surpassing larger models with lower computational overhead.
arxiv.org · Source ↗

Bio, Work & Ideas

Noyan Tokgozoglu

Conference affiliation: Optiver

On this page

Noyan Tokgozoglu has worked at Apple on personalization and machine-learning systems. His research contributions address practical constraints on deploying models: fitting recommendation systems onto devices, personalizing services without exposing users’ queries, and training small models to detect harmful content.

He recruited for Apple Media Products’ personalization ML platform, emphasizing machine-learning systems that directly affect users at scale. His subsequent coauthored work explored how to make those systems smaller and more private.

Smaller recommendation models, more private queries

Compressing embeddings without losing item identity. Tokgozoglu coauthored the 2022 paper introducing MEmCom, or Multi-Embedding Compression. Recommendation models commonly represent each app, movie, or other item with its own numerical vector. These lookup tables can occupy gigabytes, making them difficult to deploy on smartphones. MEmCom constructs each item’s vector on demand from two jointly trained tables: one uses hashing to make multiple items share a vector; the other stores a learned weight for each item so the model can distinguish items sharing that vector. Across four recommendation datasets, the researchers reported embedding compression of 4× to 40× with a 4% relative loss in nDCG, a measure of ranking quality. In a separate experiment involving millions of user interactions with games, they compressed a RankNet model by 32× with a 1% relative loss in nDCG. Tokgozoglu shared the paper as work from his team.

Personalization without revealing the query. In 2024, he contributed to an Apple team’s Private Information Retrieval for Personalization tutorial. The tutorial connected recommendation systems with two cryptographic techniques: private information retrieval lets a user retrieve data without revealing the query to the server, while homomorphic encryption enables computation on encrypted data. It explored how these techniques can support personalized services while preserving the confidentiality of user information.

Training lightweight safety classifiers

Tokgozoglu also coauthored Lightweight Safety Guardrails via Synthetic Data and RL-guided Adversarial Training, submitted in July 2025. The research explored whether small language models could provide effective content moderation at lower computational cost.

The method starts with human-curated examples, expands them through query augmentation and paraphrasing, and subjects the resulting synthetic data to repeated curation. Reinforcement learning then guides a generator toward challenging examples, which are used to further train the safety classifier. This creates an iterative way to expose weaknesses and improve the classifier’s resistance to adversarial inputs. The paper reports that small models trained through this process can match or exceed larger models on content-moderation tasks; it presents a research framework without establishing product deployment.

New York event photos