Low-Precision Hardware Architectures Meet Recommendation Model Inference at Scale

Tremendous success of machine learning (ML) and the unabated growth in ML model complexity motivated many ML-specific designs in both CPU and accelerator architectures to speed up the model inference. While these architectures are diverse, highly optimized low-precision arithmetic is a component shared by most. Impressive compute throughputs are indeed often exhibited by these architectures on benchmark ML models. Nevertheless, production models such as recommendation systems important to Facebook's personalization services are demanding and complex: These systems must serve billions of users per month responsively with low latency while maintaining high prediction accuracy, notwithstanding computations with many tens of billions parameters per inference. Do these low-precision architectures work well with our production recommendation systems? They do. But not without significant effort. We share in this paper our search strategies to adapt reference recommendation models to low-precision hardware, our optimization of low-precision compute kernels, and the design and development of tool chain so as to maintain our models' accuracy throughout their lifespan during which topic trends and users' interests inevitably evolve. Practicing these low-precision technologies helped us save datacenter capacities while deploying models with up to 5X complexity that would otherwise not be deployed on traditional general-purpose CPUs. We believe these lessons from the trenches promote better co-design between hardware architecture and software engineering and advance the state of the art of ML in industry.

High-performance, Distributed Training of Large-scale Deep Learning Recommendation Models

Deep learning recommendation models (DLRMs) are used across many business-critical services at Facebook and are the single largest AI application in terms of infrastructure demand in its data-centers. In this paper we discuss the SW/HW co-designed solution for high-performance distributed training of large-scale DLRMs. We introduce a high-performance scalable software stack based on PyTorch and pair it with the new evolution of \zion platform, namely \zionex. We demonstrate the capability to train very large DLRMs with up to \emph{12 Trillion parameters} and show that we can attain $40\times$ speedup in terms of time to solution over previous systems. We achieve this by (i) designing the \zionex platform with dedicated scale-out network, provisioned with high bandwidth, optimal topology and efficient transport (ii) implementing an optimized PyTorch-based training stack supporting both model and data parallelism (iii) developing sharding algorithms capable of hierarchical partitioning of the embedding tables along row, column dimensions and load balancing them across multiple workers; (iv) adding high-performance core operators while retaining flexibility to support optimizers with fully deterministic updates (v) leveraging reduced precision communications, multi-level memory hierarchy (HBM+DDR+SSD) and pipelining. Furthermore, we develop and briefly comment on distributed data ingestion and other supporting services that are required for the robust and efficient end-to-end training in production environments.

On the Importance of Diversity in Re-Sampling for Imbalanced Data and Rare Events in Mortality Risk Models

Surgical risk increases significantly when patients present with comorbid conditions. This has resulted in the creation of numerous risk stratification tools with the objective of formulating associated surgical risk to assist both surgeons and patients in decision-making. The Surgical Outcome Risk Tool (SORT) is one of the tools developed to predict mortality risk throughout the entire perioperative period for major elective in-patient surgeries in the UK. In this study, we enhance the original SORT prediction model (UK SORT) by addressing the class imbalance within the dataset. Our proposed method investigates the application of diversity-based selection on top of common re-sampling techniques to enhance the classifier's capability in detecting minority (mortality) events. Diversity amongst training datasets is an essential factor in ensuring re-sampled data keeps an accurate depiction of the minority/majority class region, thereby solving the generalization problem of mainstream sampling approaches. We incorporate the use of the Solow-Polasky measure as a drop-in functionality to evaluate diversity, with the addition of greedy algorithms to identify and discard subsets that share the most similarity. Additionally, through empirical experiments, we prove that the performance of the classifier trained over diversity-based dataset outperforms the original classifier over ten external datasets. Our diversity-based re-sampling method elevates the performance of the UK SORT algorithm by 1.4$.

Mixed-Precision Embedding Using a Cache

In recommendation systems, practitioners observed that increase in the number of embedding tables and their sizes often leads to significant improvement in model performances. Given this and the business importance of these models to major internet companies, embedding tables for personalization tasks have grown to terabyte scale and continue to grow at a significant rate. Meanwhile, these large-scale models are often trained with GPUs where high-performance memory is a scarce resource, thus motivating numerous work on embedding table compression during training. We propose a novel change to embedding tables using a cache memory architecture, where the majority of rows in an embedding is trained in low precision, and the most frequently or recently accessed rows cached and trained in full precision. The proposed architectural change works in conjunction with standard precision reduction and computer arithmetic techniques such as quantization and stochastic rounding. For an open source deep learning recommendation model (DLRM) running with Criteo-Kaggle dataset, we achieve 3x memory reduction with INT8 precision embedding tables and full-precision cache whose size are 5% of the embedding tables, while maintaining accuracy. For an industrial scale model and dataset, we achieve even higher >7x memory reduction with INT4 precision and cache size 1% of embedding tables, while maintaining accuracy, and 16% end-to-end training speedup by reducing GPU-to-host data transfers.

Personalized Context-Aware Multi-Modal Transportation Recommendation

This study proposes to find the most appropriate transport modes with awareness of user preferences (e.g., costs, times) and trip characteristics (e.g., purpose, distance). The work was based on real-life trips obtained from a map application. Several methods including gradient boosting tree, learning to rank, multinomial logit model, automated machine learning, random forest, and shallow neural network have been tried. For some methods, feature selection and over-sampling techniques were also tried. The results show that the best performing method is a gradient boosting tree model with synthetic minority over-sampling technique (SMOTE). Also, results of the multinomial logit model show that (1) an increase in travel cost would decrease the utility of all the transportation modes; (2) people are less sensitive to the travel distance for the metro mode or a multi-modal option that containing metro, i.e., compared to other modes, people would be more willing to tolerate long-distance metro trips. This indicates that metro lines might be a good candidate for large cities.

Future Semantic Segmentation with Convolutional LSTM

We consider the problem of predicting semantic segmentation of future frames in a video. Given several observed frames in a video, our goal is to predict the semantic segmentation map of future frames that are not yet observed. A reliable solution to this problem is useful in many applications that require real-time decision making, such as autonomous driving. We propose a novel model that uses convolutional LSTM (ConvLSTM) to encode the spatiotemporal information of observed frames for future prediction. We also extend our model to use bidirectional ConvLSTM to capture temporal information in both directions. Our proposed approach outperforms other state-of-the-art methods on the benchmark dataset.

Tracking Sentiment in Mail: How Genders Differ on Emotional Axes

With the widespread use of email, we now have access to unprecedented amounts of text that we ourselves have written. In this paper, we show how sentiment analysis can be used in tandem with effective visualizations to quantify and track emotions in many types of mail. We create a large word--emotion association lexicon by crowdsourcing, and use it to compare emotions in love letters, hate mail, and suicide notes. We show that there are marked differences across genders in how they use emotion words in work-place email. For example, women use many words from the joy--sadness axis, whereas men prefer terms from the fear--trust axis. Finally, we show visualizations that can help people track emotions in their emails.