Rex Liu

Surya Narayanan Hari, Rex Liu, Matt Thomson

Currently, over a thousand LLMs exist that are multi-purpose and are capable of performing real world tasks, including Q&A, text summarization, content generation, etc. However, accessibility, scale and reliability of free models prevents them from being widely deployed in everyday use cases. To address the first two issues of access and scale, organisations such as HuggingFace have created model repositories where users have uploaded model weights and quantized versions of models trained using different paradigms, as well as model cards describing their training process. While some models report performance on commonly used benchmarks, not all do, and interpreting the real world impact of trading off performance on a benchmark for model deployment cost, is unclear. Here, we show that a herd of open source models can match or exceed the performance of proprietary models via an intelligent router. We show that a Herd of open source models is able to match the accuracy of ChatGPT, despite being composed of models that are effectively 2.5x smaller. We show that in cases where GPT is not able to answer the query, Herd is able to identify a model that can, at least 40% of the time.

Rex Liu, Xin Liu

With the exponential growth of multimedia data, leveraging multimodal sensors presents a promising approach for improving accuracy in human activity recognition. Nevertheless, accurately identifying these activities using both video data and wearable sensor data presents challenges due to the labor-intensive data annotation, and reliance on external pretrained models or additional data. To address these challenges, we introduce Multimodal Masked Autoencoders-Based One-Shot Learning (Mu-MAE). Mu-MAE integrates a multimodal masked autoencoder with a synchronized masking strategy tailored for wearable sensors. This masking strategy compels the networks to capture more meaningful spatiotemporal features, which enables effective self-supervised pretraining without the need for external data. Furthermore, Mu-MAE leverages the representation extracted from multimodal masked autoencoders as prior information input to a cross-attention multimodal fusion layer. This fusion layer emphasizes spatiotemporal features requiring attention across different modalities while highlighting differences from other classes, aiding in the classification of various classes in metric-based one-shot learning. Comprehensive evaluations on MMAct one-shot classification show that Mu-MAE outperforms all the evaluated approaches, achieving up to an 80.17% accuracy for five-way one-shot multimodal classification, without the use of additional data.

Michael J. Zellinger, Rex Liu, Matt Thomson

LLM cascades deploy small LLMs to answer most queries, limiting the use of large and expensive LLMs to difficult queries. This approach can significantly reduce costs without impacting performance. However, risk-sensitive domains such as finance or medicine place an additional premium on avoiding model errors. Since even the most expensive models are susceptible to making mistakes, applications in these domains benefit from allowing LLM systems to completely abstain from answering difficult queries. Introducing abstention poses a design question for LLM cascades: should abstention only be allowed at the final model or also at earlier models? Since the error patterns of small and large models are correlated, allowing earlier models to abstain may reduce inference costs and latency by anticipating abstention decisions by expensive and slow models, thus avoiding the need to run these models. We investigate the benefits of such "early abstention" in LLM cascades and find that it reduces overall test loss by 2.2% on average across six benchmarks (GSM8K, MedMCQA, MMLU, TriviaQA, TruthfulQA, and XSum). These gains result from a more effective use of abstention, trading a 4.1% average increase in the overall abstention rate for a 13.0% reduction in cost and a 5.0% reduction in error rate. Our findings demonstrate the possibility of leveraging correlations between the error patterns of different language models to drive performance improvements for LLM systems with abstention.

Rex Liu, Huanle Zhang, Hamed Pirsiavash et al.

We propose MASTAF, a Model-Agnostic Spatio-Temporal Attention Fusion network for few-shot video classification. MASTAF takes input from a general video spatial and temporal representation,e.g., using 2D CNN, 3D CNN, and Video Transformer. Then, to make the most of such representations, we use self- and cross-attention models to highlight the critical spatio-temporal region to increase the inter-class variations and decrease the intra-class variations. Last, MASTAF applies a lightweight fusion network and a nearest neighbor classifier to classify each query video. We demonstrate that MASTAF improves the state-of-the-art performance on three few-shot video classification benchmarks(UCF101, HMDB51, and Something-Something-V2), e.g., by up to 91.6%, 69.5%, and 60.7% for five-way one-shot video classification, respectively.

Rex Liu, Sarina A Fazio, Huanle Zhang et al.

With the development of the Internet of Things(IoT) and Artificial Intelligence(AI) technologies, human activity recognition has enabled various applications, such as smart homes and assisted living. In this paper, we target a new healthcare application of human activity recognition, early mobility recognition for Intensive Care Unit(ICU) patients. Early mobility is essential for ICU patients who suffer from long-time immobilization. Our system includes accelerometer-based data collection from ICU patients and an AI model to recognize patients' early mobility. To improve the model accuracy and stability, we identify features that are insensitive to sensor orientations and propose a segment voting process that leverages a majority voting strategy to recognize each segment's activity. Our results show that our system improves model accuracy from 77.78\% to 81.86\% and reduces the model instability (standard deviation) from 16.69\% to 6.92\%, compared to the same AI model without our feature engineering and segment voting process.

Albara Ah Ramli, Xin Liu, Kelly Berndt et al.

Differences in gait patterns of children with Duchenne muscular dystrophy (DMD) and typically-developing (TD) peers are visible to the eye, but quantifications of those differences outside of the gait laboratory have been elusive. In this work, we measured vertical, mediolateral, and anteroposterior acceleration using a waist-worn iPhone accelerometer during ambulation across a typical range of velocities. Fifteen TD and fifteen DMD children from 3-16 years of age underwent eight walking/running activities, including five 25 meters walk/run speed-calibration tests at a slow walk to running speeds (SC-L1 to SC-L5), a 6-minute walk test (6MWT), a 100 meters fast-walk/jog/run (100MRW), and a free walk (FW). For clinical anchoring purposes, participants completed a Northstar Ambulatory Assessment (NSAA). We extracted temporospatial gait clinical features (CFs) and applied multiple machine learning (ML) approaches to differentiate between DMD and TD children using extracted temporospatial gait CFs and raw data. Extracted temporospatial gait CFs showed reduced step length and a greater mediolateral component of total power (TP) consistent with shorter strides and Trendelenberg-like gait commonly observed in DMD. ML approaches using temporospatial gait CFs and raw data varied in effectiveness at differentiating between DMD and TD controls at different speeds, with an accuracy of up to 100%. We demonstrate that by using ML with accelerometer data from a consumer-grade smartphone, we can capture DMD-associated gait characteristics in toddlers to teens.

Rex Liu, Albara Ah Ramli, Huanle Zhang et al.

With the rapid development of the internet of things (IoT) and artificial intelligence (AI) technologies, human activity recognition (HAR) has been applied in a variety of domains such as security and surveillance, human-robot interaction, and entertainment. Even though a number of surveys and review papers have been published, there is a lack of HAR overview papers focusing on healthcare applications that use wearable sensors. Therefore, we fill in the gap by presenting this overview paper. In particular, we present our projects to illustrate the system design of HAR applications for healthcare. Our projects include early mobility identification of human activities for intensive care unit (ICU) patients and gait analysis of Duchenne muscular dystrophy (DMD) patients. We cover essential components of designing HAR systems including sensor factors (e.g., type, number, and placement location), AI model selection (e.g., classical machine learning models versus deep learning models), and feature engineering. In addition, we highlight the challenges of such healthcare-oriented HAR systems and propose several research opportunities for both the medical and the computer science community.

Albara Ah Ramli, Rex Liu, Rahul Krishnamoorthy et al.

Amyotrophic lateral sclerosis (ALS) is a progressive neurodegenerative disease of the brain and the spinal cord, which leads to paralysis of motor functions. Patients retain their ability to blink, which can be used for communication. Here, We present an Artificial Intelligence (AI) system that uses eye-blinks to communicate with the outside world, running on real-time Internet-of-Things (IoT) devices. The system uses a Convolutional Neural Network (CNN) to find the blinking pattern, which is defined as a series of Open and Closed states. Each pattern is mapped to a collection of words that manifest the patient's intent. To investigate the best trade-off between accuracy and latency, we investigated several Convolutional Network architectures, such as ResNet, SqueezeNet, DenseNet, and InceptionV3, and evaluated their performance. We found that the InceptionV3 architecture, after hyper-parameter fine-tuning on the specific task led to the best performance with an accuracy of 99.20% and 94ms latency. This work demonstrates how the latest advances in deep learning architectures can be adapted for clinical systems that ameliorate the patient's quality of life regardless of the point-of-care.

Drew Linsley, Alekh Karkada Ashok, Lakshmi Narasimhan Govindarajan et al.

Primate vision depends on recurrent processing for reliable perception. A growing body of literature also suggests that recurrent connections improve the learning efficiency and generalization of vision models on classic computer vision challenges. Why then, are current large-scale challenges dominated by feedforward networks? We posit that the effectiveness of recurrent vision models is bottlenecked by the standard algorithm used for training them, "back-propagation through time" (BPTT), which has O(N) memory-complexity for training an N step model. Thus, recurrent vision model design is bounded by memory constraints, forcing a choice between rivaling the enormous capacity of leading feedforward models or trying to compensate for this deficit through granular and complex dynamics. Here, we develop a new learning algorithm, "contractor recurrent back-propagation" (C-RBP), which alleviates these issues by achieving constant O(1) memory-complexity with steps of recurrent processing. We demonstrate that recurrent vision models trained with C-RBP can detect long-range spatial dependencies in a synthetic contour tracing task that BPTT-trained models cannot. We further show that recurrent vision models trained with C-RBP to solve the large-scale Panoptic Segmentation MS-COCO challenge outperform the leading feedforward approach, with fewer free parameters. C-RBP is a general-purpose learning algorithm for any application that can benefit from expansive recurrent dynamics. Code and data are available at https://github.com/c-rbp.