Jacob Whitehill

Pitipat Kongsomjit, Suryansh Goyal, Jacob Whitehill

Soft prompt tuning is a parameter-efficient method for adapting LLMs to specific tasks, but suffers from a lack of interpretability. Building on recent work on interpreting soft prompts (Ramati et al., 2024), we explore how training a dedicated soft prompt to natural language translation model can yield higher translation quality. In particular, in both quantitative and qualitative comparisons on multiple Datasets of Datasets (DoDs), we demonstrate that our translator produces fluent, accurate verbalizations that outperforms existing training-free methods like InSPEcT. In addition to advancing interpretability, our work suggests a promising downstream application: soft prompts optimized on small, open-source models can be translated into portable text prompts that, when deployed on larger closed-API models, exceed the performance of the original soft prompt and, in some cases, even few-shot learning.

Xinlu He, Jacob Whitehill

Monaural multi-speaker automatic speech recognition (ASR) remains challenging due to data scarcity and the intrinsic difficulty of recognizing and attributing words to individual speakers, particularly in overlapping speech. Recent advances have driven the shift from cascade systems to end-to-end (E2E) architectures, which reduce error propagation and better exploit the synergy between speech content and speaker identity. Despite rapid progress in E2E multi-speaker ASR, the field lacks a comprehensive review of recent developments. This survey provides a systematic taxonomy of E2E neural approaches for multi-speaker ASR, highlighting recent advances and comparative analysis. Specifically, we analyze: (1) architectural paradigms (SIMO vs.~SISO) for pre-segmented audio, analyzing their distinct characteristics and trade-offs; (2) recent architectural and algorithmic improvements based on these two paradigms; (3) extensions to long-form speech, including segmentation strategy and speaker-consistent hypothesis stitching. Further, we (4) evaluate and compare methods across standard benchmarks. We conclude with a discussion of open challenges and future research directions towards building robust and scalable multi-speaker ASR.

Jacob Whitehill, Jennifer LoCasale-Crouch

With the aim to provide teachers with more specific, frequent, and actionable feedback about their teaching, we explore how Large Language Models (LLMs) can be used to estimate ``Instructional Support'' domain scores of the CLassroom Assessment Scoring System (CLASS), a widely used observation protocol. We design a machine learning architecture that uses either zero-shot prompting of Meta's Llama2, and/or a classic Bag of Words (BoW) model, to classify individual utterances of teachers' speech (transcribed automatically using OpenAI's Whisper) for the presence of Instructional Support. Then, these utterance-level judgments are aggregated over a 15-min observation session to estimate a global CLASS score. Experiments on two CLASS-coded datasets of toddler and pre-kindergarten classrooms indicate that (1) automatic CLASS Instructional Support estimation accuracy using the proposed method (Pearson $R$ up to $0.48$) approaches human inter-rater reliability (up to $R=0.55$); (2) LLMs generally yield slightly greater accuracy than BoW for this task, though the best models often combined features extracted from both LLM and BoW; and (3) for classifying individual utterances, there is still room for improvement of automated methods compared to human-level judgments. Finally, (4) we illustrate how the model's outputs can be visualized at the utterance level to provide teachers with explainable feedback on which utterances were most positively or negatively correlated with specific CLASS dimensions.

Yiwen Guan, Viet Anh Trinh, Vivek Voleti et al.

Decoder-only discrete-token language models have recently achieved significant success in automatic speech recognition. However, systematic analyses of how different modalities impact performance in specific scenarios remain limited. In this paper, we investigate the effects of multiple modalities on recognition accuracy on both synthetic and real-world datasets. Our experiments suggest that: (1) Integrating more modalities can increase accuracy; in particular, our paper is, to our best knowledge, the first to show the benefit of combining audio, image context, and lip information; (2) Images as a supplementary modality for speech recognition provide the greatest benefit at moderate noise levels, moreover, they exhibit a different trend compared to inherently synchronized modalities like lip movements; (3) Performance improves on both synthetic and real-world datasets when the most relevant visual information is filtered as a preprocessing step.

Evan Gibson Smith, Jacob Whitehill, Fatemeh Ganji

Quantization is a natural complement to the sparse, event-driven computation of Spiking Neural Networks, reducing memory bandwidth and arithmetic cost for deployment on resource-constrained hardware. However, existing SNN quantization evaluation focuses almost exclusively on accuracy, overlooking whether a quantized network preserves the firing behavior of its full-precision counterpart. We demonstrate that quantization method, clipping range, and bit-width can produce substantially different firing distributions at equivalent accuracy, differences invisible to standard metrics but relevant to deployment, where firing activity governs effective sparsity, state storage, and event-processing load. To capture this gap, we propose Earth Mover's Distance as a diagnostic metric for firing distribution divergence, and apply it systematically across weight and membrane quantization on SEW-ResNet architectures trained on CIFAR-10 and CIFAR-100. We find that uniform quantization induces distributional drift even when accuracy is preserved, while LQ-Net style learned quantization maintains firing behavior close to the full-precision baseline. Our results suggest that behavior preservation should be treated as an evaluation criterion alongside accuracy, and that EMD provides a principled tool for assessing it.

Cutter Beck, Evan Smith, Khagendra Katuwal et al.

Coronal holes (CHs) are low-activity, low-density solar coronal regions with open magnetic field lines (Cranmer 2009). In the extreme ultraviolet (EUV) spectrum, CHs appear as dark patches. Using daily hand-drawn maps from the Space Weather Prediction Center (SWPC), we developed a semi-automated pipeline to digitize the SWPC maps into binary segmentation masks. The resulting masks constitute the CHASM-SWPC dataset, a high-quality dataset to train and test automated CH detection models, which is released with this paper. We developed CHASM (Coronal Hole Annotation using Semi-automatic Methods), a software tool for semi-automatic annotation that enables users to rapidly and accurately annotate SWPC maps. The CHASM tool enabled us to annotate 1,111 CH masks, comprising the CHASM-SWPC-1111 dataset. We then trained multiple CHRONNOS (Coronal Hole RecOgnition Neural Network Over multi-Spectral-data) architecture (Jarolim et al. 2021) neural networks using the CHASM-SWPC dataset and compared their performance. Training the CHRONNOS neural network on these data achieved an accuracy of 0.9805, a True Skill Statistic (TSS) of 0.6807, and an intersection-over-union (IoU) of 0.5668, which is higher than the original pretrained CHRONNOS model Jarolim et al. (2021) achieved an accuracy of 0.9708, a TSS of 0.6749, and an IoU of 0.4805, when evaluated on the CHASM-SWPC-1111 test set.

Xinlu He, Yiwen Guan, Badrivishal Paurana et al.

Most automatic speech processing systems operate in "open loop" mode without user feedback about who said what; yet, human-in-the-loop workflows can potentially enable higher accuracy. We propose an LLM-assisted speaker diarization correction system that lets users fix speaker attribution errors in real time. The pipeline performs streaming ASR and diarization, uses an LLM to deliver concise summaries to the users, and accepts brief verbal feedback that is immediately incorporated without disrupting interactions. Moreover, we develop techniques to make the workflow more effective: First, a split-when-merged (SWM) technique detects and splits multi-speaker segments that the ASR erroneously attributes to just a single speaker. Second, online speaker enrollments are collected based on users' diarization corrections, thus helping to prevent speaker diarization errors from occurring in the future. LLM-driven simulations on the AMI test set indicate that our system substantially reduces DER by 9.92% and speaker confusion error by 44.23%. We further analyze correction efficacy under different settings, including summary vs full transcript display, the number of online enrollments limitation, and correction frequency.

Yiwen Guan, Jacob Whitehill

Multilingual translation faces challenges of computational redundancy and limited accuracy for low-resource languages, especially in speech translation. To address this, we propose a novel hierarchical Transformer Encoder Tree (TET) combined with non-autoregressive encoder-only models trained with Connectionist Temporal Classification for multilingual translation. By sharing intermediate representations among linguistically similar target languages, TET can improve accuracy on low-resource languages, reduce computational redundancy, and allow generating all target languages in a single forward pass, thus eliminating sequential bottlenecks and improving parallelism. For speech translation, combining TET with a non-autoregressive speech recognition backbone (wav2vec2) shows promising results in terms of translation quality compared to autoregressive systems while being 7-14 times faster.

Yiwen Guan, Viet Anh Trinh, Vivek Voleti et al.

Recent advances in multi-modal large language models (MLLMs) have opened new possibilities for unified modeling of speech, text, images, and other modalities. Building on our prior work, this paper examines the conditions and model architectures under which multiple input modalities can improve automatic speech recognition (ASR) accuracy in noisy environments. Through experiments on synthetic and real-world data, we find that (1) harnessing more modalities usually improves ASR accuracy, as each modality provides complementary information, but the improvement depends on the amount of auditory noise. (2) Synchronized modalities (e.g., lip movements) are more useful at high noise levels whereas unsynchronized modalities (e.g., image context) are most helpful at moderate noise levels. (3) Higher-quality visual representations consistently improve ASR accuracy, highlighting the importance of developing more powerful visual encoders. (4) Mamba exhibits similar trends regarding the benefits of multimodality as do Transformers. (5) The input order of modalities as well as their weights in the loss function can significantly impact accuracy. These findings both offer practical insights and help to deepen our understanding of multi-modal speech recognition under challenging conditions.

Viet Anh Trinh, Xinlu He, Jacob Whitehill

With recent advances in modeling and the increasing amount of supervised training data, automatic speech recognition (ASR) systems have achieved remarkable performance on general speech. However, the word error rate (WER) of state-of-the-art ASR remains high for named entities. Since named entities are often the most critical keywords, misrecognizing them can affect all downstream applications, especially when the ASR system functions as the front end of a complex system. In this paper, we introduce a large language model (LLM) revision mechanism to revise incorrect named entities in ASR predictions by leveraging the LLM's reasoning ability as well as local context (e.g., lecture notes) containing a set of correct named entities. Finally, we introduce the NER-MIT-OpenCourseWare dataset, containing 45 hours of data from MIT courses for development and testing. On this dataset, our proposed technique achieves up to 30\% relative WER reduction for named entities.

Viet Anh Trinh, Rosy Southwell, Yiwen Guan et al.

Recent work on discrete speech tokenization has paved the way for models that can seamlessly perform multiple tasks across modalities, e.g., speech recognition, text to speech, speech to speech translation. Moreover, large language models (LLMs) pretrained from vast text corpora contain rich linguistic information that can improve accuracy in a variety of tasks. In this paper, we present a decoder-only Discrete Multimodal Language Model (DMLM), which can be flexibly applied to multiple tasks (ASR, T2S, S2TT, etc.) and modalities (text, speech, vision). We explore several critical aspects of discrete multi-modal models, including the loss function, weight initialization, mixed training supervision, and codebook. Our results show that DMLM benefits significantly, across multiple tasks and datasets, from a combination of supervised and unsupervised training. Moreover, for ASR, it benefits from initializing DMLM from a pretrained LLM, and from a codebook derived from Whisper activations.

Zeqian Li, Xinlu He, Jacob Whitehill

We consider a novel clustering task in which clusters can have compositional relationships, e.g., one cluster contains images of rectangles, one contains images of circles, and a third (compositional) cluster contains images with both objects. In contrast to hierarchical clustering in which a parent cluster represents the intersection of properties of the child clusters, our problem is about finding compositional clusters that represent the union of the properties of the constituent clusters. This task is motivated by recently developed few-shot learning and embedding models can distinguish the label sets, not just the individual labels, assigned to the examples. We propose three new algorithms -- Compositional Affinity Propagation (CAP), Compositional k-means (CKM), and Greedy Compositional Reassignment (GCR) -- that can partition examples into coherent groups and infer the compositional structure among them. We show promising results, compared to popular algorithms such as Gaussian mixtures, Fuzzy c-means, and Agglomerative Clustering, on the OmniGlot and LibriSpeech datasets. Our work has applications to open-world multi-label object recognition and speaker identification & diarization with simultaneous speech from multiple speakers.

Anand Ramakrishnan, Minh Pham, Jacob Whitehill

For the task of face verification, we explore the utility of harnessing auxiliary facial emotion labels to impose explicit geometric constraints on the embedding space when training deep embedding models. We introduce several novel loss functions that, in conjunction with a standard Triplet Loss [43], or ArcFace loss [10], provide geometric constraints on the embedding space; the labels for our loss functions can be provided using either manually annotated or automatically detected auxiliary emotion labels. Our method is implemented purely in terms of the loss function and does not require any changes to the neural network backbone of the embedding function.

Zeqian Li, Jacob Whitehill

We propose a new method for speaker diarization that can handle overlapping speech with 2+ people. Our method is based on compositional embeddings [1]: Like standard speaker embedding methods such as x-vector [2], compositional embedding models contain a function f that separates speech from different speakers. In addition, they include a composition function g to compute set-union operations in the embedding space so as to infer the set of speakers within the input audio. In an experiment on multi-person speaker identification using synthesized LibriSpeech data, the proposed method outperforms traditional embedding methods that are only trained to separate single speakers (not speaker sets). In a speaker diarization experiment on the AMI Headset Mix corpus, we achieve state-of-the-art accuracy (DER=22.93%), slightly higher than the previous best result (23.82% from [3]).

Anand Ramakrishnan, Brian Zylich, Erin Ottmar et al.

In this work we present a multi-modal machine learning-based system, which we call ACORN, to analyze videos of school classrooms for the Positive Climate (PC) and Negative Climate (NC) dimensions of the CLASS observation protocol that is widely used in educational research. ACORN uses convolutional neural networks to analyze spectral audio features, the faces of teachers and students, and the pixels of each image frame, and then integrates this information over time using Temporal Convolutional Networks. The audiovisual ACORN's PC and NC predictions have Pearson correlations of $0.55$ and $0.63$ with ground-truth scores provided by expert CLASS coders on the UVA Toddler dataset (cross-validation on $n=300$ 15-min video segments), and a purely auditory ACORN predicts PC and NC with correlations of $0.36$ and $0.41$ on the MET dataset (test set of $n=2000$ videos segments). These numbers are similar to inter-coder reliability of human coders. Finally, using Graph Convolutional Networks we make early strides (AUC=$0.70$) toward predicting the specific moments (45-90sec clips) when the PC is particularly weak/strong. Our findings inform the design of automatic classroom observation and also more general video activity recognition and summary recognition systems.

Zeqian Li, Michael C. Mozer, Jacob Whitehill

We present a compositional embedding framework that infers not just a single class per input image, but a set of classes, in the setting of one-shot learning. Specifically, we propose and evaluate several novel models consisting of (1) an embedding function f trained jointly with a "composition" function g that computes set union operations between the classes encoded in two embedding vectors; and (2) embedding f trained jointly with a "query" function h that computes whether the classes encoded in one embedding subsume the classes encoded in another embedding. In contrast to prior work, these models must both perceive the classes associated with the input examples and encode the relationships between different class label sets, and they are trained using only weak one-shot supervision consisting of the label-set relationships among training examples. Experiments on the OmniGlot, Open Images, and COCO datasets show that the proposed compositional embedding models outperform existing embedding methods. Our compositional embedding models have applications to multi-label object recognition for both one-shot and supervised learning.

Jacob Whitehill, Anand Ramakrishnan

Automatic machine learning-based detectors of various psychological and social phenomena (e.g., emotion, stress, engagement) have great potential to advance basic science. However, when a detector $d$ is trained to approximate an existing measurement tool (e.g., a questionnaire, observation protocol), then care must be taken when interpreting measurements collected using $d$ since they are one step further removed from the underlying construct. We examine how the accuracy of $d$, as quantified by the correlation $q$ of $d$'s outputs with the ground-truth construct $U$, impacts the estimated correlation between $U$ (e.g., stress) and some other phenomenon $V$ (e.g., academic performance). In particular: (1) We show that if the true correlation between $U$ and $V$ is $r$, then the expected sample correlation, over all vectors $\mathcal{T}^n$ whose correlation with $U$ is $q$, is $qr$. (2) We derive a formula for the probability that the sample correlation (over $n$ subjects) using $d$ is positive given that the true correlation is negative (and vice-versa); this probability can be substantial (around $20-30\%$) for values of $n$ and $q$ that have been used in recent affective computing studies. %We also show that this probability decreases monotonically in $n$ and in $q$. (3) With the goal to reduce the variance of correlations estimated by an automatic detector, we show that training multiple neural networks $d^{(1)},\ldots,d^{(m)}$ using different training architectures and hyperparameters for the same detection task provides only limited ``coverage'' of $\mathcal{T}^n$.

Jacob Whitehill

Recent work on privacy-preserving machine learning has considered how data-mining competitions such as Kaggle could potentially be "hacked", either intentionally or inadvertently, by using information from an oracle that reports a classifier's accuracy on the test set. For binary classification tasks in particular, one of the most common accuracy metrics is the Area Under the ROC Curve (AUC), and in this paper we explore the mathematical structure of how the AUC is computed from an n-vector of real-valued "guesses" with respect to the ground-truth labels. We show how knowledge of a classifier's AUC on the test set can constrain the set of possible ground-truth labelings, and we derive an algorithm both to compute the exact number of such labelings and to enumerate efficiently over them. Finally, we provide empirical evidence that, surprisingly, the number of compatible labelings can actually decrease as n grows, until a test set-dependent threshold is reached.

Jacob Whitehill

In the context of data-mining competitions (e.g., Kaggle, KDDCup, ILSVRC Challenge), we show how access to an oracle that reports a contestant's log-loss score on the test set can be exploited to deduce the ground-truth of some of the test examples. By applying this technique iteratively to batches of $m$ examples (for small $m$), all of the test labels can eventually be inferred. In this paper, (1) We demonstrate this attack on the first stage of a recent Kaggle competition (Intel & MobileODT Cancer Screening) and use it to achieve a log-loss of $0.00000$ (and thus attain a rank of #4 out of 848 contestants), without ever training a classifier to solve the actual task. (2) We prove an upper bound on the batch size $m$ as a function of the floating-point resolution of the probability estimates that the contestant submits for the labels. (3) We derive, and demonstrate in simulation, a more flexible attack that can be used even when the oracle reports the accuracy on an unknown (but fixed) subset of the test set's labels. These results underline the importance of evaluating contestants based only on test data that the oracle does not examine.

Jacob Whitehill, Kiran Mohan, Daniel Seaton et al.

In order to obtain reliable accuracy estimates for automatic MOOC dropout predictors, it is important to train and test them in a manner consistent with how they will be used in practice. Yet most prior research on MOOC dropout prediction has measured test accuracy on the same course used for training the classifier, which can lead to overly optimistic accuracy estimates. In order to understand better how accuracy is affected by the training+testing regime, we compared the accuracy of a standard dropout prediction architecture (clickstream features + logistic regression) across 4 different training paradigms. Results suggest that (1) training and testing on the same course ("post-hoc") can overestimate accuracy by several percentage points; (2) dropout classifiers trained on proxy labels based on students' persistence are surprisingly competitive with post-hoc training (87.33% versus 90.20% AUC averaged over 8 weeks of 40 HarvardX MOOCs); and (3) classifier performance does not vary significantly with the academic discipline. Finally, we also research new dropout prediction architectures based on deep, fully-connected, feed-forward neural networks and find that (4) networks with as many as 5 hidden layers can statistically significantly increase test accuracy over that of logistic regression.

Jacob Whitehill, Margo Seltzer

We investigated the feasibility of crowdsourcing full-fledged tutorial videos from ordinary people on the Web on how to solve math problems related to logarithms. This kind of approach (a form of learnersourcing) to efficiently collecting tutorial videos and other learning resources could be useful for realizing personalized learning-at-scale, whereby students receive specific learning resources -- drawn from a large and diverse set -- that are tailored to their individual and time-varying needs. Results of our study, in which we collected 399 videos from 66 unique "teachers" on Mechanical Turk, suggest that (1) approximately 100 videos -- over $80\%$ of which are mathematically fully correct -- can be crowdsourced per week for \$5/video; (2) the crowdsourced videos exhibit significant diversity in terms of language style, presentation media, and pedagogical approach; (3) the average learning gains (posttest minus pretest score) associated with watching the videos was stat.~sig.~higher than for a control video ($0.105$ versus $0.045$); and (4) the average learning gains ($0.1416$) from watching the best tested crowdsourced videos was comparable to the learning gains ($0.1506$) from watching a popular Khan Academy video on logarithms.

Jacob Whitehill

In machine learning contests such as the ImageNet Large Scale Visual Recognition Challenge and the KDD Cup, contestants can submit candidate solutions and receive from an oracle (typically the organizers of the competition) the accuracy of their guesses compared to the ground-truth labels. One of the most commonly used accuracy metrics for binary classification tasks is the Area Under the Receiver Operating Characteristics Curve (AUC). In this paper we provide proofs-of-concept of how knowledge of the AUC of a set of guesses can be used, in two different kinds of attacks, to improve the accuracy of those guesses. On the other hand, we also demonstrate the intractability of one kind of AUC exploit by proving that the number of possible binary labelings of $n$ examples for which a candidate solution obtains a AUC score of $c$ grows exponentially in $n$, for every $c\in (0,1)$.

Jacob Whitehill

The ACT-R theory of cognition developed by John Anderson and colleagues endeavors to explain how humans recall chunks of information and how they solve problems. ACT-R also serves as a theoretical basis for "cognitive tutors", i.e., automatic tutoring systems that help students learn mathematics, computer programming, and other subjects. The official ACT-R definition is distributed across a large body of literature spanning many articles and monographs, and hence it is difficult for an "outsider" to learn the most important aspects of the theory. This paper aims to provide a tutorial to the core components of the ACT-R theory.