Rahul Sharma

Rahul Sharma, Anil Ramakrishna, Ansel MacLaughlin et al. · amazon-science

Machine Learning (ML) systems are getting increasingly popular, and drive more and more applications and services in our daily life. This has led to growing concerns over user privacy, since human interaction data typically needs to be transmitted to the cloud in order to train and improve such systems. Federated learning (FL) has recently emerged as a method for training ML models on edge devices using sensitive user data and is seen as a way to mitigate concerns over data privacy. However, since ML models are most commonly trained with label supervision, we need a way to extract labels on edge to make FL viable. In this work, we propose a strategy for training FL models using positive and negative user feedback. We also design a novel framework to study different noise patterns in user feedback, and explore how well standard noise-robust objectives can help mitigate this noise when training models in a federated setting. We evaluate our proposed training setup through detailed experiments on two text classification datasets and analyze the effects of varying levels of user reliability and feedback noise on model performance. We show that our method improves substantially over a self-training baseline, achieving performance closer to models trained with full supervision.

Rahul Sharma, Shrikanth Narayanan

Active speaker detection in videos addresses associating a source face, visible in the video frames, with the underlying speech in the audio modality. The two primary sources of information to derive such a speech-face relationship are i) visual activity and its interaction with the speech signal and ii) co-occurrences of speakers' identities across modalities in the form of face and speech. The two approaches have their limitations: the audio-visual activity models get confused with other frequently occurring vocal activities, such as laughing and chewing, while the speakers' identity-based methods are limited to videos having enough disambiguating information to establish a speech-face association. Since the two approaches are independent, we investigate their complementary nature in this work. We propose a novel unsupervised framework to guide the speakers' cross-modal identity association with the audio-visual activity for active speaker detection. Through experiments on entertainment media videos from two benchmark datasets, the AVA active speaker (movies) and Visual Person Clustering Dataset (TV shows), we show that a simple late fusion of the two approaches enhances the active speaker detection performance.

Rahul Sharma, Shrikanth Narayanan

An essential goal of computational media intelligence is to support understanding how media stories -- be it news, commercial or entertainment media -- represent and reflect society and these portrayals are perceived. People are a central element of media stories. This paper focuses on understanding the representation and depiction of background characters in media depictions, primarily movies and TV shows. We define the background characters as those who do not participate vocally in any scene throughout the movie and address the problem of localizing background characters in videos. We use an active speaker localization system to extract high-confidence face-speech associations and generate audio-visual profiles for talking characters in a movie by automatically clustering them. Using a face verification system, we then prune all the face-tracks which match any of the generated character profiles and obtain the background character face-tracks. We curate a background character dataset which provides annotations for background character for a set of TV shows, and use it to evaluate the performance of the background character detection framework.

Wonyeol Lee, Rahul Sharma, Alex Aiken · stanford

When training deep neural networks, keeping all tensors in high precision (e.g., 32-bit or even 16-bit floats) is often wasteful. However, keeping all tensors in low precision (e.g., 8-bit floats) can lead to unacceptable accuracy loss. Hence, it is important to use a precision assignment -- a mapping from all tensors (arising in training) to precision levels (high or low) -- that keeps most of the tensors in low precision and leads to sufficiently accurate models. We provide a technique that explores this memory-accuracy tradeoff by generating precision assignments for convolutional neural networks that (i) use less memory and (ii) lead to more accurate convolutional networks at the same time, compared to the precision assignments considered by prior work in low-precision floating-point training. We evaluate our technique on image classification tasks by training convolutional networks on CIFAR-10, CIFAR-100, and ImageNet. Our method typically provides > 2x memory reduction over a baseline precision assignment while preserving training accuracy, and gives further reductions by trading off accuracy. Compared to other baselines which sometimes cause training to diverge, our method provides similar or better memory reduction while avoiding divergence.

Rahul Sharma, Shrikanth Narayanan

We present a cross-modal unsupervised framework for active speaker detection in media content such as TV shows and movies. Machine learning advances have enabled impressive performance in identifying individuals from speech and facial images. We leverage speaker identity information from speech and faces, and formulate active speaker detection as a speech-face assignment task such that the active speaker's face and the underlying speech identify the same person (character). We express the speech segments in terms of their associated speaker identity distances, from all other speech segments, to capture a relative identity structure for the video. Then we assign an active speaker's face to each speech segment from the concurrently appearing faces such that the obtained set of active speaker faces displays a similar relative identity structure. Furthermore, we propose a simple and effective approach to address speech segments where speakers are present off-screen. We evaluate the proposed system on three benchmark datasets -- Visual Person Clustering dataset, AVA-active speaker dataset, and Columbia dataset -- consisting of videos from entertainment and broadcast media, and show competitive performance to state-of-the-art fully supervised methods.

Shikhar Jaiswal, Rahul Kiran Kranti Goli, Aayan Kumar et al.

Running machine learning inference on tiny devices, known as TinyML, is an emerging research area. This task requires generating inference code that uses memory frugally, a task that standard ML frameworks are ill-suited for. A deployment framework for TinyML must be a) parametric in the number representation to take advantage of the emerging representations like posits, b) carefully assign high-precision to a few tensors so that most tensors can be kept in low-precision while still maintaining model accuracy, and c) avoid memory fragmentation. We describe MinUn, the first TinyML framework that holistically addresses these issues to generate efficient code for ARM microcontrollers (e.g., Arduino Uno, Due and STM32H747) that outperforms the prior TinyML frameworks.

Rahul Sharma, Kunal Dhawan, Balakrishna Pailla

This work presents a novel methodology for calculating the phonetic similarity between words taking motivation from the human perception of sounds. This metric is employed to learn a continuous vector embedding space that groups similar sounding words together and can be used for various downstream computational phonology tasks. The efficacy of the method is presented for two different languages (English, Hindi) and performance gains over previous reported works are discussed on established tests for predicting phonetic similarity. To address limited benchmarking mechanisms in this field, we also introduce a heterographic pun dataset based evaluation methodology to compare the effectiveness of acoustic similarity algorithms. Further, a visualization of the embedding space is presented with a discussion on the various possible use-cases of this novel algorithm. An open-source implementation is also shared to aid reproducibility and enable adoption in related tasks.

Rahul Sharma, Shrikanth Narayanan

Speaker diarization is one of the critical components of computational media intelligence as it enables a character-level analysis of story portrayals and media content understanding. Automated audio-based speaker diarization of entertainment media poses challenges due to the diverse acoustic conditions present in media content, be it background music, overlapping speakers, or sound effects. At the same time, speaking faces in the visual modality provide complementary information and not prone to the errors seen in the audio modality. In this paper, we address the problem of speaker diarization in TV shows using the active speaker faces. We perform face clustering on the active speaker faces and show superior speaker diarization performance compared to the state-of-the-art audio-based diarization methods. We additionally report a systematic analysis of the impact of active speaker face detection quality on the diarization performance. We also observe that a moderately well-performing active speaker system could outperform the audio-based diarization systems.

Naman Jain, Skanda Vaidyanath, Arun Iyer et al.

Large pre-trained language models such as GPT-3, Codex, and Google's language model are now capable of generating code from natural language specifications of programmer intent. We view these developments with a mixture of optimism and caution. On the optimistic side, such large language models have the potential to improve productivity by providing an automated AI pair programmer for every programmer in the world. On the cautionary side, since these large language models do not understand program semantics, they offer no guarantees about quality of the suggested code. In this paper, we present an approach to augment these large language models with post-processing steps based on program analysis and synthesis techniques, that understand the syntax and semantics of programs. Further, we show that such techniques can make use of user feedback and improve with usage. We present our experiences from building and evaluating such a tool jigsaw, targeted at synthesizing code for using Python Pandas API using multi-modal inputs. Our experience suggests that as these large language models evolve for synthesizing code from intent, jigsaw has an important role to play in improving the accuracy of the systems.

Rahul Sharma, Soumya Banerjee, Dootika Vats et al.

We present a variational inference (VI) framework that unifies and leverages sequential Monte-Carlo (particle filtering) with \emph{approximate} rejection sampling to construct a flexible family of variational distributions. Furthermore, we augment this approach with a resampling step via Bernoulli race, a generalization of a Bernoulli factory, to obtain a low-variance estimator of the marginal likelihood. Our framework, Variational Rejection Particle Filtering (VRPF), leads to novel variational bounds on the marginal likelihood, which can be optimized efficiently with respect to the variational parameters and generalizes several existing approaches in the VI literature. We also present theoretical properties of the variational bound and demonstrate experiments on various models of sequential data, such as the Gaussian state-space model and variational recurrent neural net (VRNN), on which VRPF outperforms various existing state-of-the-art VI methods.

Rahul Sharma, Krishna Somandepalli, Shrikanth Narayanan

An objective understanding of media depictions, such as inclusive portrayals of how much someone is heard and seen on screen such as in film and television, requires the machines to discern automatically who, when, how, and where someone is talking, and not. Speaker activity can be automatically discerned from the rich multimodal information present in the media content. This is however a challenging problem due to the vast variety and contextual variability in the media content, and the lack of labeled data. In this work, we present a cross-modal neural network for learning visual representations, which have implicit information pertaining to the spatial location of a speaker in the visual frames. Avoiding the need for manual annotations for active speakers in visual frames, acquiring of which is very expensive, we present a weakly supervised system for the task of localizing active speakers in movie content. We use the learned cross-modal visual representations, and provide weak supervision from movie subtitles acting as a proxy for voice activity, thus requiring no manual annotations. We evaluate the performance of the proposed system on the AVA active speaker dataset and demonstrate the effectiveness of the cross-modal embeddings for localizing active speakers in comparison to fully supervised systems. We also demonstrate state-of-the-art performance for the task of voice activity detection in an audio-visual framework, especially when speech is accompanied by noise and music.

Rahul Sharma, Abhishek Kumar, Piyush Rai

We present an approximate inference method, based on a synergistic combination of Rényi $α$-divergence variational inference (RDVI) and rejection sampling (RS). RDVI is based on minimization of Rényi $α$-divergence $D_α(p||q)$ between the true distribution $p(x)$ and a variational approximation $q(x)$; RS draws samples from a distribution $p(x) = \tilde{p}(x)/Z_{p}$ using a proposal $q(x)$, s.t. $Mq(x) \geq \tilde{p}(x), \forall x$. Our inference method is based on a crucial observation that $D_\infty(p||q)$ equals $\log M(θ)$ where $M(θ)$ is the optimal value of the RS constant for a given proposal $q_θ(x)$. This enables us to develop a \emph{two-stage} hybrid inference algorithm. Stage-1 performs RDVI to learn $q_θ$ by minimizing an estimator of $D_α(p||q)$, and uses the learned $q_θ$ to find an (approximately) optimal $\tilde{M}(θ)$. Stage-2 performs RS using the constant $\tilde{M}(θ)$ to improve the approximate distribution $q_θ$ and obtain a sample-based approximation. We prove that this two-stage method allows us to learn considerably more accurate approximations of the target distribution as compared to RDVI. We demonstrate our method's efficacy via several experiments on synthetic and real datasets.

Chantat Eksombatchai, Pranav Jindal, Jerry Zitao Liu et al.

User experience in modern content discovery applications critically depends on high-quality personalized recommendations. However, building systems that provide such recommendations presents a major challenge due to a massive pool of items, a large number of users, and requirements for recommendations to be responsive to user actions and generated on demand in real-time. Here we present Pixie, a scalable graph-based real-time recommender system that we developed and deployed at Pinterest. Given a set of user-specific pins as a query, Pixie selects in real-time from billions of possible pins those that are most related to the query. To generate recommendations, we develop Pixie Random Walk algorithm that utilizes the Pinterest object graph of 3 billion nodes and 17 billion edges. Experiments show that recommendations provided by Pixie lead up to 50% higher user engagement when compared to the previous Hadoop-based production system. Furthermore, we develop a graph pruning strategy at that leads to an additional 58% improvement in recommendations. Last, we discuss system aspects of Pixie, where a single server executes 1,200 recommendation requests per second with 60 millisecond latency. Today, systems backed by Pixie contribute to more than 80% of all user engagement on Pinterest.

Rahul Sharma, Tanaya Guha, Gaurav Sharma

Public speaking is an important aspect of human communication and interaction. The majority of computational work on public speaking concentrates on analyzing the spoken content, and the verbal behavior of the speakers. While the success of public speaking largely depends on the content of the talk, and the verbal behavior, non-verbal (visual) cues, such as gestures and physical appearance also play a significant role. This paper investigates the importance of visual cues by estimating their contribution towards predicting the popularity of a public lecture. For this purpose, we constructed a large database of more than $1800$ TED talk videos. As a measure of popularity of the TED talks, we leverage the corresponding (online) viewers' ratings from YouTube. Visual cues related to facial and physical appearance, facial expressions, and pose variations are extracted from the video frames using convolutional neural network (CNN) models. Thereafter, an attention-based long short-term memory (LSTM) network is proposed to predict the video popularity from the sequence of visual features. The proposed network achieves state-of-the-art prediction accuracy indicating that visual cues alone contain highly predictive information about the popularity of a talk. Furthermore, our network learns a human-like attention mechanism, which is particularly useful for interpretability, i.e. how attention varies with time, and across different visual cues by indicating their relative importance.

Rahul Anand Sharma, Bharath Bhat, Vineet Gandhi et al.

In this paper, we propose a novel method to register football broadcast video frames on the static top view model of the playing surface. The proposed method is fully automatic in contrast to the current state of the art which requires manual initialization of point correspondences between the image and the static model. Automatic registration using existing approaches has been difficult due to the lack of sufficient point correspondences. We investigate an alternate approach exploiting the edge information from the line markings on the field. We formulate the registration problem as a nearest neighbour search over a synthetically generated dictionary of edge map and homography pairs. The synthetic dictionary generation allows us to exhaustively cover a wide variety of camera angles and positions and reduce this problem to a minimal per-frame edge map matching procedure. We show that the per-frame results can be improved in videos using an optimization framework for temporal camera stabilization. We demonstrate the efficacy of our approach by presenting extensive results on a dataset collected from matches of football World Cup 2014.

Rahul Anand Sharma, Pramod Sankar K, CV Jawahar

The recognition of human activities is one of the key problems in video understanding. Action recognition is challenging even for specific categories of videos, such as sports, that contain only a small set of actions. Interestingly, sports videos are accompanied by detailed commentaries available online, which could be used to perform action annotation in a weakly-supervised setting. For the specific case of Cricket videos, we address the challenge of temporal segmentation and annotation of ctions with semantic descriptions. Our solution consists of two stages. In the first stage, the video is segmented into "scenes", by utilizing the scene category information extracted from text-commentary. The second stage consists of classifying video-shots as well as the phrases in the textual description into various categories. The relevant phrases are then suitably mapped to the video-shots. The novel aspect of this work is the fine temporal scale at which semantic information is assigned to the video. As a result of our approach, we enable retrieval of specific actions that last only a few seconds, from several hours of video. This solution yields a large number of labeled exemplars, with no manual effort, that could be used by machine learning algorithms to learn complex actions.