Pushing the Performance Envelope of DNN-based Recommendation Systems Inference on GPUs

Personalized recommendation is a ubiquitous application on the internet, with many industries and hyperscalers extensively leveraging Deep Learning Recommendation Models (DLRMs) for their personalization needs (like ad serving or movie suggestions). With growing model and dataset sizes pushing computation and memory requirements, GPUs are being increasingly preferred for executing DLRM inference. However, serving newer DLRMs, while meeting acceptable latencies, continues to remain challenging, making traditional deployments increasingly more GPU-hungry, resulting in higher inference serving costs. In this paper, we show that the embedding stage continues to be the primary bottleneck in the GPU inference pipeline, leading up to a 3.2x embedding-only performance slowdown. To thoroughly grasp the problem, we conduct a detailed microarchitecture characterization and highlight the presence of low occupancy in the standard embedding kernels. By leveraging direct compiler optimizations, we achieve optimal occupancy, pushing the performance by up to 53%. Yet, long memory latency stalls continue to exist. To tackle this challenge, we propose specialized plug-and-play-based software prefetching and L2 pinning techniques, which help in hiding and decreasing the latencies. Further, we propose combining them, as they complement each other. Experimental evaluations using A100 GPUs with large models and datasets show that our proposed techniques improve performance by up to 103% for the embedding stage, and up to 77% for the overall DLRM inference pipeline.

A Perspective on Foundation Models for the Electric Power Grid

Foundation models (FMs) currently dominate news headlines. They employ advanced deep learning architectures to extract structural information autonomously from vast datasets through self-supervision. The resulting rich representations of complex systems and dynamics can be applied to many downstream applications. Therefore, FMs can find uses in electric power grids, challenged by the energy transition and climate change. In this paper, we call for the development of, and state why we believe in, the potential of FMs for electric grids. We highlight their strengths and weaknesses amidst the challenges of a changing grid. We argue that an FM learning from diverse grid data and topologies could unlock transformative capabilities, pioneering a new approach in leveraging AI to redefine how we manage complexity and uncertainty in the electric grid. Finally, we discuss a power grid FM concept, namely GridFM, based on graph neural networks and show how different downstream tasks benefit.

Shotluck Holmes: A Family of Efficient Small-Scale Large Language Vision Models For Video Captioning and Summarization

Video is an increasingly prominent and information-dense medium, yet it poses substantial challenges for language models. A typical video consists of a sequence of shorter segments, or shots, that collectively form a coherent narrative. Each shot is analogous to a word in a sentence where multiple data streams of information (such as visual and auditory data) must be processed simultaneously. Comprehension of the entire video requires not only understanding the visual-audio information of each shot but also requires that the model links the ideas between each shot to generate a larger, all-encompassing story. Despite significant progress in the field, current works often overlook videos' more granular shot-by-shot semantic information. In this project, we propose a family of efficient large language vision models (LLVMs) to boost video summarization and captioning called Shotluck Holmes. By leveraging better pretraining and data collection strategies, we extend the abilities of existing small LLVMs from being able to understand a picture to being able to understand a sequence of frames. Specifically, we show that Shotluck Holmes achieves better performance than state-of-the-art results on the Shot2Story video captioning and summary task with significantly smaller and more computationally efficient models.

Data Center Audio/Video Intelligence on Device (DAVID) -- An Edge-AI Platform for Smart-Toys

An overview is given of the DAVID Smart-Toy platform, one of the first Edge AI platform designs to incorporate advanced low-power data processing by neural inference models co-located with the relevant image or audio sensors. There is also on-board capability for in-device text-to-speech generation. Two alternative embodiments are presented: a smart Teddy-bear, and a roving dog-like robot. The platform offers a speech-driven user interface and can observe and interpret user actions and facial expressions via its computer vision sensor node. A particular benefit of this design is that no personally identifiable information passes beyond the neural inference nodes thus providing inbuilt compliance with data protection regulations.

Synthetic Speaking Children -- Why We Need Them and How to Make Them

Contemporary Human Computer Interaction (HCI) research relies primarily on neural network models for machine vision and speech understanding of a system user. Such models require extensively annotated training datasets for optimal performance and when building interfaces for users from a vulnerable population such as young children, GDPR introduces significant complexities in data collection, management, and processing. Motivated by the training needs of an Edge AI smart toy platform this research explores the latest advances in generative neural technologies and provides a working proof of concept of a controllable data generation pipeline for speech driven facial training data at scale. In this context, we demonstrate how StyleGAN2 can be finetuned to create a gender balanced dataset of children's faces. This dataset includes a variety of controllable factors such as facial expressions, age variations, facial poses, and even speech-driven animations with realistic lip synchronization. By combining generative text to speech models for child voice synthesis and a 3D landmark based talking heads pipeline, we can generate highly realistic, entirely synthetic, talking child video clips. These video clips can provide valuable, and controllable, synthetic training data for neural network models, bridging the gap when real data is scarce or restricted due to privacy regulations.

A comparative analysis between Conformer-Transducer, Whisper, and wav2vec2 for improving the child speech recognition

Automatic Speech Recognition (ASR) systems have progressed significantly in their performance on adult speech data; however, transcribing child speech remains challenging due to the acoustic differences in the characteristics of child and adult voices. This work aims to explore the potential of adapting state-of-the-art Conformer-transducer models to child speech to improve child speech recognition performance. Furthermore, the results are compared with those of self-supervised wav2vec2 models and semi-supervised multi-domain Whisper models that were previously finetuned on the same data. We demonstrate that finetuning Conformer-transducer models on child speech yields significant improvements in ASR performance on child speech, compared to the non-finetuned models. We also show Whisper and wav2vec2 adaptation on different child speech datasets. Our detailed comparative analysis shows that wav2vec2 provides the most consistent performance improvements among the three methods studied.

Improved Child Text-to-Speech Synthesis through Fastpitch-based Transfer Learning

Speech synthesis technology has witnessed significant advancements in recent years, enabling the creation of natural and expressive synthetic speech. One area of particular interest is the generation of synthetic child speech, which presents unique challenges due to children's distinct vocal characteristics and developmental stages. This paper presents a novel approach that leverages the Fastpitch text-to-speech (TTS) model for generating high-quality synthetic child speech. This study uses the transfer learning training pipeline. The approach involved finetuning a multi-speaker TTS model to work with child speech. We use the cleaned version of the publicly available MyST dataset (55 hours) for our finetuning experiments. We also release a prototype dataset of synthetic speech samples generated from this research together with model code to support further research. By using a pretrained MOSNet, we conducted an objective assessment that showed a significant correlation between real and synthetic child voices. Additionally, to validate the intelligibility of the generated speech, we employed an automatic speech recognition (ASR) model to compare the word error rates (WER) of real and synthetic child voices. The speaker similarity between the real and generated speech is also measured using a pretrained speaker encoder.

MAGIC-TBR: Multiview Attention Fusion for Transformer-based Bodily Behavior Recognition in Group Settings

Bodily behavioral language is an important social cue, and its automated analysis helps in enhancing the understanding of artificial intelligence systems. Furthermore, behavioral language cues are essential for active engagement in social agent-based user interactions. Despite the progress made in computer vision for tasks like head and body pose estimation, there is still a need to explore the detection of finer behaviors such as gesturing, grooming, or fumbling. This paper proposes a multiview attention fusion method named MAGIC-TBR that combines features extracted from videos and their corresponding Discrete Cosine Transform coefficients via a transformer-based approach. The experiments are conducted on the BBSI dataset and the results demonstrate the effectiveness of the proposed feature fusion with multiview attention. The code is available at: https://github.com/surbhimadan92/MAGIC-TBR

Automatic Concept Embedding Model (ACEM): No train-time concepts, No issue!

Interpretability and explainability of neural networks is continuously increasing in importance, especially within safety-critical domains and to provide the social right to explanation. Concept based explanations align well with how humans reason, proving to be a good way to explain models. Concept Embedding Models (CEMs) are one such concept based explanation architectures. These have shown to overcome the trade-off between explainability and performance. However, they have a key limitation -- they require concept annotations for all their training data. For large datasets, this can be expensive and infeasible. Motivated by this, we propose Automatic Concept Embedding Models (ACEMs), which learn the concept annotations automatically.

Adaptation of Whisper models to child speech recognition

Automatic Speech Recognition (ASR) systems often struggle with transcribing child speech due to the lack of large child speech datasets required to accurately train child-friendly ASR models. However, there are huge amounts of annotated adult speech datasets which were used to create multilingual ASR models, such as Whisper. Our work aims to explore whether such models can be adapted to child speech to improve ASR for children. In addition, we compare Whisper child-adaptations with finetuned self-supervised models, such as wav2vec2. We demonstrate that finetuning Whisper on child speech yields significant improvements in ASR performance on child speech, compared to non finetuned Whisper models. Additionally, utilizing self-supervised Wav2vec2 models that have been finetuned on child speech outperforms Whisper finetuning.