Efficient large-scale neural network training and inference on commodity CPU hardware is of immense practical significance in democratizing deep learning (DL) capabilities. Presently, the process of training massive models consisting of hundreds of millions to billions of parameters requires the extensive use of specialized hardware accelerators, such as GPUs, which are only accessible to a limited number of institutions with considerable financial resources. Moreover, there is often an alarming carbon footprint associated with training and deploying these models. In this paper, we take a step towards addressing these challenges by introducing BOLT, a sparse deep learning library for training large-scale search and recommendation models on standard CPU hardware. BOLT provides a flexible, high-level API for constructing models that will be familiar to users of existing popular DL frameworks. By automatically tuning specialized hyperparameters, BOLT also abstracts away the algorithmic details of sparse network training. We evaluate BOLT on a number of information retrieval tasks including product recommendations, text classification, graph neural networks, and personalization. We find that our proposed system achieves competitive performance with state-of-the-art techniques at a fraction of the cost and energy consumption and an order-of-magnitude faster inference time. BOLT has also been successfully deployed by multiple businesses to address critical problems, and we highlight one customer case study in the field of e-commerce.

In the last decade, large-scale deep learning has fundamentally transformed industrial recommendation systems. However, this revolutionary technology remains prohibitively expensive due to the need for costly and scarce specialized hardware, such as Graphics Processing Units (GPUs), to train and serve models. In this talk, we share our multi-year journey at ThirdAI in developing efficient neural recommendation models that can be trained and deployed on commodity CPU machines without the need for costly accelerators like GPUs. In particular, we discuss the limitations of the current GPU-based ecosystem in machine learning, why recommendation systems are amenable to the strengths of CPU devices, and present results from our efforts to translate years of academic research into a deployable system that fundamentally shifts the economics of training and operating large-scale machine learning models.

Representation learning has transformed the problem of information retrieval into one of finding the approximate set of nearest neighbors in a high dimensional vector space. With limited hardware resources and time-critical queries, the retrieval engines face an inherent tension between latency, accuracy, scalability, compactness, and the ability to load balance in distributed settings. To improve the trade-off, we propose a new algorithm, called BaLanced Index for Scalable Search (BLISS), a highly tunable indexing algorithm with enviably small index sizes, making it easy to scale to billions of vectors. It iteratively refines partitions of items by learning the relevant buckets directly from the query-item relevance data. To ensure that the buckets are balanced, BLISS uses the power-of-K choices strategy. We show that BLISS provides superior load balancing with high probability (and under very benign assumptions). Due to its design, BLISS can be employed for both near-neighbor retrieval (ANN problem) and extreme classification (XML problem). For the case of ANN, we train and index 4 datasets with billion vectors each. We compare the recall, inference time, indexing time, and index size for BLISS with the two most popular and well-optimized libraries- Hierarchical Navigable Small World (HNSW) graph and Facebook's FAISS. BLISS requires 100x lesser RAM than HNSW, making it fit in memory on commodity machines while taking a similar inference time as HNSW for the same recall. Against FAISS-IVF, BLISS achieves similar performance with 3-4x less memory requirement. BLISS is both data and model parallel, making it ideal for distributed implementation for training and inference. For the case of XML, BLISS surpasses the best baselines' precision while being 5x faster for inference on popular multi-label datasets with half a million classes.

Softmax classifiers with a very large number of classes naturally occur in many applications such as natural language processing and information retrieval. The calculation of full softmax is costly from the computational and energy perspective. There have been various sampling approaches to overcome this challenge, popularly known as negative sampling (NS). Ideally, NS should sample negative classes from a distribution that is dependent on the input data, the current parameters, and the correct positive class. Unfortunately, due to the dynamically updated parameters and data samples, there is no sampling scheme that is provably adaptive and samples the negative classes efficiently. Therefore, alternative heuristics like random sampling, static frequency-based sampling, or learning-based biased sampling, which primarily trade either the sampling cost or the adaptivity of samples per iteration are adopted. In this paper, we show two classes of distributions where the sampling scheme is truly adaptive and provably generates negative samples in near-constant time. Our implementation in C++ on CPU is significantly superior, both in terms of wall-clock time and accuracy, compared to the most optimized TensorFlow implementations of other popular negative sampling approaches on powerful NVIDIA V100 GPU.

DNA sequencing, especially of microbial genomes and metagenomes, has been at the core of recent research advances in large-scale comparative genomics. The data deluge has resulted in exponential growth in genomic datasets over the past years and has shown no sign of slowing down. Several recent attempts have been made to tame the computational burden of sequence search on these terabyte and petabyte-scale datasets, including raw reads and assembled genomes. However, no known implementation provides both fast query and construction time, keeps the low false-positive requirement, and offers cheap storage of the data structure. We propose a data structure for search called RAMBO (Repeated And Merged BloOm Filter) which is significantly faster in query time than state-of-the-art genome indexing methods- COBS (Compact bit-sliced signature index), Sequence Bloom Trees, HowDeSBT, and SSBT. Furthermore, it supports insertion and query process parallelism, cheap updates for streaming inputs, has a zero false-negative rate, a low false-positive rate, and a small index size. RAMBO converts the search problem into set membership testing among K documents. Interestingly, it is a count-min sketch type arrangement of a membership testing utility (Bloom Filter in our case). The simplicity of the algorithm and embarrassingly parallel architecture allows us to stream and index a 170TB whole-genome sequence dataset in a mere 9 hours on a cluster of 100 nodes while competing methods require weeks.

Zero-Shot Learning (ZSL) is a classification task where some classes referred to as unseen classes have no training images. Instead, we only have side information about seen and unseen classes, often in the form of semantic or descriptive attributes. Lack of training images from a set of classes restricts the use of standard classification techniques and losses, including the widespread cross-entropy loss. We introduce a novel Similarity Distribution Matching Network (SDM-Net) which is a standard fully connected neural network architecture with a non-trainable penultimate layer consisting of class attributes. The output layer of SDM-Net consists of both seen and unseen classes. To enable zero-shot learning, during training, we regularize the model such that the predicted distribution of unseen class is close in KL divergence to the distribution of similarities between the correct seen class and all the unseen classes. We evaluate the proposed model on five benchmark datasets for zero-shot learning, AwA1, AwA2, aPY, SUN, and CUB datasets. We show that, despite the simplicity, our approach achieves competitive performance with state-of-the-art methods in Generalized-ZSL setting for all of these datasets.

In the last decade, it has been shown that many hard AI tasks, especially in NLP, can be naturally modeled as extreme classification problems leading to improved precision. However, such models are prohibitively expensive to train due to the memory bottleneck in the last layer. For example, a reasonable softmax layer for the dataset of interest in this paper can easily reach well beyond 100 billion parameters (> 400 GB memory). To alleviate this problem, we present Merged-Average Classifiers via Hashing (MACH), a generic $K$-classification algorithm where memory provably scales at $O(\log K)$ without any assumption on the relation between classes. MACH is subtly a count-min sketch structure in disguise, which uses universal hashing to reduce classification with a large number of classes to few embarrassingly parallel and independent classification tasks with a small (constant) number of classes. MACH naturally provides a technique for zero communication model parallelism. We experiment with 6 datasets; some multiclass and some multilabel, and show consistent improvement in precision and recall metrics compared to respective baselines. In particular, we train an end-to -end deep classifier on a private product search dataset sampled from Amazon Search Engine with 70 million queries and 49.46 million documents. MACH outperforms, by a significant margin, the state-of-the-art extreme classification models deployed on commercial search engines: Parabel and dense embedding models. Our largest model has 6.4 billion parameters and trains in less than 35 hrs on a single p3.16x machine. Our training times are 7-10x faster, and our memory footprints are 2-4x smaller than the best baselines. This training time is also significantly lower than the one reported by Google’s mixture of experts (MoE) language model on a comparable model size and hardware.