Behind the Screen: Inside YouTube's Game-Changing Ranking Algorithm

Written by Anne Schulze

Google’s dominance in Artificial Intelligence and Machine Learning is built on a foundation of stability, optimization, and quality. Its search and video ranking algorithms rely on these principles to consistently deliver accurate and relevant results to users worldwide. A key contributor to this success is Nikhil Khani, a Senior Software Engineer at YouTube, whose work has significantly impacted both Google’s internal operations and the digital experience of billions of YouTube users.

Optimizing YouTube’s Ranking Models

Khani has made significant contributions to Google by optimizing YouTube’s resource-intensive ranking models. These models are computationally costly and their slow loading times can negatively impact user experience. To address this, Khani and his team have focused on optimizing the scoring system within strict resource and latency budgets. Their improvements have successfully removed bottlenecks and cleared the path for crucial model scaling projects. The result? A remarkable USD 2.5 million in savings across various initiatives between January and August 2024.

One of the strategies Khani employed was model quantization, a state-of-the-art technique commonly used by large language models to reduce computational costs. His work on model quantization involved utilizing a lower precision data format (bfloat16) for specific calculations within the model, minimizing memory and processing power requirements without losing any prediction accuracy. This approach significantly reduced the overhead associated with data exchange, resulting in faster processing and enhanced efficiency. Remarkably, this has decreased training costs down by 30% and serving costs by 3% – a substantial saving considering the model’s high traffic volume.

In addition to quantization, Khani also streamlined the models by meticulously identifying and eliminating redundancies. This included streamlining YouTube's multi-objective ranking model by removing low-importance objectives. Khani's work also tackled inefficiencies in the ranking model that had significant consequences. One such example is the unnecessary variable assignments between Tensor Processing Units (TPUs, specialized hardware for running machine learning workloads) and Central Processing Units (CPUs, used in case of spillovers from TPUs) which created performance bottlenecks during training. By diligently profiling and subsequently eliminating these operations, Khani achieved a 4% reduction in memory bandwidth usage and a 10% improvement in average response time. These optimizations translated into substantial cost savings, totaling nearly $250,000. In addition to these optimization efforts, Khani has also made significant strides in improving the quality and stability of the Hydra ranking model.

Improving Hydra Quality and Stability

Khani’s relentless pursuit of quality and stability within the Hydra ranking model represents a critical contribution to YouTube’s success. To understand its significance, consider the sheer scale of YouTube: millions of users across the globe accessing the platform simultaneously, with ranking models constantly being updated and rolled out behind the scenes. Even minor glitches or testing errors can cascade into significant disruptions, potentially leading to outages that degrade recommendations and, in the worst cases, prevent the homepage from loading altogether. Such incidents disrupt the user experience and underscore the critical importance of model stability. In early 2023, YouTube’s homepage experienced two such outages, prompting Khani to lead the post-mortem analysis, identify the root causes, and implement preventative measures.

To address the issues, Khani adopted a two-pronged approach: increasing visibility into model stability and improving the model’s robustness to rapid changes in data. A key achievement was the development of a model rollback dashboard that provides deep visibility into unstable models throughout their lifecycle. This dashboard proved so valuable that it was quickly incorporated into YouTube’s launch process, enabling engineers to identify and address stability issues sooner.

Further bolstering the main ranking model's stability, Khani led two high-impact projects. First, he led the implementation of the Clippy Optimizer on the Homepage. Clippy is an innovative optimization algorithm developed at Google Research. Optimizers are algorithms used to adjust the parameters of a machine learning model during training, and a technique called gradient clipping is often employed to prevent instability. Clippy distinguishes itself by effectively fusing these two approaches, leading to a more stable training procedure. Given its significant practical implications and importance to the industry, the research on Clippy was recognized with the prestigious KDD 2023 Best Paper Award. Second, Khani integrated advanced model architectures chosen for their robustness and resilience. Specifically, this involved incorporating proven architectures like Residual Networks (ResNets) and Transformer Networks, known for their ability to handle complex patterns and prevent instability during training. Modifying the architecture of the core ranking model called over 200,000 times every second is a high-stakes undertaking requiring meticulous planning, extensive testing, and careful coordination to avoid any disruptions to YouTube’s service. Khani orchestrated rigorous A/B testing, staged rollouts, and continuous monitoring to ensure seamless integration and validate performance improvements. Ultimately, this work yielded exceptional results: the model was shown to be 5 times more robust to higher variance while improving its overall prediction accuracy which contributed to an increase of a million new active users on the platform.

In addition to these stability improvements, Khani has been instrumental in integrating attention-based architectures into the Hydra model. Attention mechanisms allow AI models to focus on the most relevant parts of the input data, much like humans pay attention to specific details when processing information. Inspired by the success of transformers in natural language processing, Khani and his team replaced the older Deep & Cross Network (DCN) with a pointwise attention mechanism. This change has enabled a more nuanced understanding of feature interactions at scale, leading to significant improvements in offline and online metrics. In fact, the combination of the attention mechanism and increasing the model size by 3 times (another launch led by Khani) resulted in the largest single improvement in YouTube ranking accuracy observed in the past five years.

This shift towards attention-based architectures has not only enhanced the model’s performance but has also positioned YouTube at the forefront of advancements in AI and recommendation systems globally. By demonstrating the effectiveness of attention mechanisms in ranking models, Khani’s work has inspired further research and development within the broader AI community.

Khani’s Lasting Impact on Google and the AI Industry

Khani’s impact, particularly in ranking model optimization, has profoundly impacted YouTube and Google's efficiency and user experience. His dedication to improving model quality and stability has not only created a more seamless and engaging experience for billions of users worldwide, but also established new standards for performance and reliability in the industry. 

With contributions extending beyond Google and influencing the broader AI and ML landscape his constant pursuit of efficient and robust AI systems, has led to cost savings worth millions of dollars and has made a positive impact on sustainability and responsible AI efforts across various sectors. By consistently pushing the boundaries of what’s possible in AI, Khani has solidified his position as a leading figure in the field, shaping the future of digital content consumption, and inspiring others to pursue excellence in machine learning.

The post Behind the Screen: Inside YouTube’s Game-Changing Ranking Algorithm appeared first on New York Tech Media.