Abstract

Machine learning methods have seen a meteoric rise in their applications in the scientific community. However, little effort has been put into understanding these “black box” models. We show how one can apply integrated gradients (IGs) to understand these models by designing different baselines, by taking an example case study in particle physics. We find that the zero-vector baseline does not provide good feature attributions and that an averaged baseline sampled from the background events provides consistently more reasonable attributions.

Abstract

For people with hearing impairments, lipreading is a crucial skill, as it bridges the gap between spoken language and understanding, improving the ability to communicate effectively. However, the creation of lipreading resources is pervasive and labor-intensive, leading to a scarcity of high-quality, structured, and widely accessible educational materials. To address this, we developed LipMOOC, a Massive Open Online Course (MOOC) tailored for lipreading education. Our approach leverages synthetically generated content, reducing the dependence on human-recorded videos and building an accessible and effective learning option. Using the state-of-the-art models, we created a comprehensive lipreading course and conducted a study to evaluate its effectiveness. We statistically proved that (1) Synthetic videos can replace human-recorded videos without affecting the learning outcomes, and (2) Learner performance on a structured course, i.e., LipMOOC, is much better than a standard course. A total of 120 participants with hearing impairment and no prior lipreading experience volunteered. These results highlight the potential of AI-driven content generation and scientific instructional design.

Abstract

Submodular functions are known to satisfy various forms of fractional subadditivity. This work investigates the conditions for equality to hold exactly or approximately in the fractional subadditivity of submodular functions. We establish that a small gap in the inequality implies that the function is close to being modular, and that the gap is zero if and only if the function is modular. We then present natural implications of these results for special cases of submodular functions, such as entropy, relative entropy, and matroid rank. As a consequence, we characterize the necessary and sufficient conditions for equality to hold in Shearer's lemma, recovering a result of Ellis et al. (2016) as a special case. We leverage our results to propose a new multivariate mutual information, which generalizes Watanabe's total correlation (1960), Han's dual total correlation (1975), and Csiszar and Narayan's shared information (2004), and analyze its properties. Among these properties, we extend Watanabe's characterization of total correlation as the maximum correlation over partitions to fractional partitions. When applied to matrix determinantal inequalities for positive definite matrices, our results recover the equality conditions of the classical determinantal inequalities of Hadamard, Szasz, and Fischer as special cases.

Abstract

Foundational Language Models perform significantly better on downstream tasks in specialised domains (such as law, computer science, and medical science) upon being further pre-trained on extensive domain-specific corpora, but this continual pre-training incurs heavy computational costs. Indeed, some of the most performant specialised language models such as BioBERT incur even higher computing costs during domain-specific training than the pre-training cost of the foundational models they are initialised from. In this paper, we argue that much of the extended pre-training is redundant, with models seemingly wasting valuable resources re-learning lexical and semantic patterns already well-represented in their foundational models such as BERT, T5 and GPT. Focusing on Masked Language Models, we introduce a novel domain-specific masking strategy that is designed to facilitate continual learning while minimizing the training cost. Using this approach, we train and present a BERT-based model trained on a biomedical corpus that matches or surpasses traditionally trained biomedical language models in performance across several downstream classification tasks while incurring up to 11 times lower training costs.

Abstract

Foundational Language Models perform significantly better on downstream tasks in the biomedical domain upon being further pre-trained on extensive biomedical corpora, but this continual pre-training incurs heavy computational costs. Indeed, some of the most performant biomedical language models incur even more computing costs during domain-specific training than the entire training cost of the foundational models they are warm-started from. In this paper, we argue that much of the extended pre-training is redundant, with models seemingly wasting valuable resources re-learning lexical and semantic patterns already well-represented in their foundational models such as BERT, T5 and GPT. Focusing on Masked Language Models, we introduce a novel domain-specific masking strategy that is designed to facilitate continual learning while minimizing the training cost. Using this approach, we train and present a BERT-based model that matches or surpasses traditionally trained biomedical language models in performance across several downstream classification tasks while incurring up to 11 times lower training costs.

Abstract

One of the central themes of Algebraic Complexity Theory is to understand the relative computation power of algebraic complexity classes VP and VNP. VP is a class of polynomial families which can be computed efficiently by algebraic circuits. VNP is a class of polynomials which can be efficiently expressed through polynomial families in VP, but we do not know if they can be computed efficiently. It is a long standing open problem in this area to show that VP is a strict subset of VNP. At this juncture, it is fair to believe that newer characterization of these complexity classes could help us understand these models better (and possibly help us in resolving this question in restricted settings like non-commutativity at the very least). It is well known that Algebraic Branching Programs (ABPs) characterize the class VBP (or equivalently VPskew ). Recently, Mengel [MFCS, 2013] showed that ABPs can be extended with memory and the computational models thus obtained can be used to characterize VP and VNP. In particular, Mengel showed that ABPs with a single stack characterize VP, and branching programs with random-access memory characterize VNP. In this work we show that algebraic branching programs with just 2 stacks efficiently simulates the polynomial families in VNP, and two stack branching programs are VNP-complete. Further, we observe that k-stack branching programs (for all polynomially bounded k) are no more powerful than 2-stack branching programs (upto a polynomial blowup in size). This strengthens the work of Mengel in the following way – VBP vs. VP vs. VNP are characterized by algebraic branching programs with 0, 1, and 2 stacks, respectively, or no-memory, a stack as memory, and a queue as memory respectively, which hold even in the non-commutative setting. We also refine Mengel’s characterization of VP through stack-height characterization.

Abstract

In information-theoretic private information retrieval (PIR), a client wants to retrieve one desired file out of $M$ files, stored across $N$ servers, while keeping the index of the desired file private from each $T$-sized subset of servers. A PIR protocol must ideally maximize the rate, which is the ratio of the file size to the total quantum of the download from the servers, while ensuring such privacy. In Weak-PIR (WPIR), the criterion of perfect information-theoretic privacy is relaxed. This enables higher rates to be achieved, while some information about the desired file index leaks to the servers. This leakage is captured by various known privacy metrics. By leveraging the well-established capacity-achieving schemes of Sun and Jafar under non-colluding ($T=1$) and colluding ($1