Imagine to Hear: Auditory Knowledge Generation can be an Effective Assistant for Language Models

Language models pretrained on text-only corpora often struggle with tasks that require auditory commonsense knowledge. Previous work addresses this problem by augmenting the language model to retrieve knowledge from external audio databases. This approach has several limitations, such as the potential lack of relevant audio in databases and the high costs associated with constructing and querying the databases. To address these issues, we propose Imagine to Hear, a novel approach that dynamically generates auditory knowledge using generative models. Our framework detects multiple audio-related textual spans from the given prompt and generates corresponding auditory knowledge. We develop several mechanisms to efficiently process multiple auditory knowledge, including a CLAP-based rejection sampler and a language-audio fusion module. Our experiments show that our method achieves state-of-the-art performance on AuditoryBench without relying on external databases, highlighting the effectiveness of our generation-based approach.

On the Internal Representations of Graph Metanetworks

Weight space learning is an emerging paradigm in the deep learning community. The primary goal of weight space learning is to extract informative features from a set of parameters using specially designed neural networks, often referred to as \emph{metanetworks}. However, it remains unclear how these metanetworks learn solely from parameters. To address this, we take the first step toward understanding \emph{representations} of metanetworks, specifically graph metanetworks (GMNs), which achieve state-of-the-art results in this field, using centered kernel alignment (CKA). Through various experiments, we reveal that GMNs and general neural networks (\textit{e.g.,} multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs)) differ in terms of their representation space.

Prompt-based Depth Pruning of Large Language Models

Depth pruning aims to reduce the inference cost of a large language model without any hardware-specific complications, by simply removing several less important transformer blocks. However, our empirical findings suggest that the importance of a transformer block may be highly task-dependent -- a block that is crucial for a task can be removed without degrading the accuracy on another task. Based on this observation, we develop a dynamic depth pruning algorithm, coined PuDDing (Prompt-routed Dynamic Depth Pruning), which determines which blocks to omit from the model based on the input prompt. PuDDing operates by training a lightweight router to predict the best omission set among a set of options, where this option set has also been constructed in a data-driven manner. Empirical results on commonsense reasoning benchmarks demonstrate that PuDDing effectively accelerates the inference language models, and achieves better on-task performance than static depth pruning baselines.

Fast Training of Sinusoidal Neural Fields via Scaling Initialization

Neural fields are an emerging paradigm that represent data as continuous functions parameterized by neural networks. Despite many advantages, neural fields often have a high training cost, which prevents a broader adoption. In this paper, we focus on a popular family of neural fields, called sinusoidal neural fields (SNFs), and study how it should be initialized to maximize the training speed. We find that the standard initialization scheme for SNFs -- designed based on the signal propagation principle -- is suboptimal. In particular, we show that by simply multiplying each weight (except for the last layer) by a constant, we can accelerate SNF training by 10$\times$. This method, coined $\textit{weight scaling}$, consistently provides a significant speedup over various data domains, allowing the SNFs to train faster than more recently proposed architectures. To understand why the weight scaling works well, we conduct extensive theoretical and empirical analyses which reveal that the weight scaling not only resolves the spectral bias quite effectively but also enjoys a well-conditioned optimization trajectory.

Constructing a Singing Style Caption Dataset

Singing voice synthesis and conversion have emerged as significant subdomains of voice generation, leading to much demands on prompt-conditioned generation. Unlike common voice data, generating a singing voice requires an understanding of various associated vocal and musical characteristics, such as the vocal tone of the singer or emotional expressions. However, existing open-source audio-text datasets for voice generation tend to capture only a very limited range of attributes, often missing musical characteristics of the audio. To fill this gap, we introduce S2Cap, an audio-text pair dataset with a diverse set of attributes. S2Cap consists of pairs of textual prompts and music audio samples with a wide range of vocal and musical attributes, including pitch, volume, tempo, mood, singer's gender and age, and musical genre and emotional expression. Utilizing S2Cap, we suggest an effective novel baseline algorithm for singing style captioning. Singing style captioning is a relative task to voice generation that generates text descriptions of vocal characteristics, which we first suggested. First, to mitigate the misalignment between the audio encoder and the text decoder, we present a novel mechanism called CRESCENDO, which utilizes positive-pair similarity learning to synchronize the embedding spaces of a pretrained audio encoder to get similar embeddings with a text encoder. We additionally supervise the model using the singer's voice, which is demixed by the accompaniment. This supervision allows the model to more accurately capture vocal characteristics, leading to improved singing style captions that better reflect the style of the singer. The dataset and the codes are available at \bulurl{https://github.com/HJ-Ok/S2cap}.

AudioBERT: Audio Knowledge Augmented Language Model

Recent studies have identified that language models, pretrained on text-only datasets, often lack elementary visual knowledge, \textit{e.g.,} colors of everyday objects. Motivated by this observation, we ask whether a similar shortcoming exists in terms of the \textit{auditory} knowledge. To answer this question, we construct a new dataset called AuditoryBench, which consists of two novel tasks for evaluating auditory knowledge. Based on our analysis using the benchmark, we find that language models also suffer from a severe lack of auditory knowledge. To address this limitation, we propose AudioBERT, a novel method to augment the auditory knowledge of BERT through a retrieval-based approach. First, we detect auditory knowledge spans in prompts to query our retrieval model efficiently. Then, we inject audio knowledge into BERT and switch on low-rank adaptation for effective adaptation when audio knowledge is required. Our experiments demonstrate that AudioBERT is quite effective, achieving superior performance on the AuditoryBench. The dataset and code are available at \bulurl{https://github.com/HJ-Ok/AudioBERT}.

Decoding with Limited Teacher Supervision Requires Understanding When to Trust the Teacher

How can sLLMs efficiently utilize the supervision of LLMs to improve their generative quality? This question has been well studied in scenarios where there is no restriction on the number of LLM supervisions one can use, giving birth to many decoding algorithms that utilize supervision without further training. However, it is still unclear what is an effective strategy under the limited supervision scenario, where we assume that no more than a few tokens can be generated by LLMs. To this end, we develop an algorithm to effectively aggregate the sLLM and LLM predictions on initial tokens so that the generated tokens can more accurately condition the subsequent token generation by sLLM only. Critically, we find that it is essential to adaptively overtrust or disregard the LLM prediction based on the confidence of the sLLM. Through our experiments on a wide range of models and datasets, we demonstrate that our method provides a consistent improvement over conventional decoding strategies.

Towards Federated Low-Rank Adaptation with Rank-Heterogeneous Communication

Low-rank adaptation (LoRA) is an attractive alternative of adapting full weights for the federated fine-tuning of large pretrained models, which can significantly reduce the memory and communication burden. In principle, federated LoRA can provide an effective mean to allocate different resources to each client by tuning ranks for each client, which can be useful in achieving a better communication-performance tradeoff. We find, however, that the empirical performance of LoRA is highly unstable with respect to such rank-heterogeneity, severely limiting the applicability to the scenarios where it is desirable or even required to allocate nonuniform communication bandwidth to each client due to constrained total bandwidth. Our investigation reveals that the root cause of this instability is the zero-padding-based aggregation strategy adopted in conventional federated LoRA frameworks, which causes the information from high rank clients to get diluted during the aggregation process. To address this issue, we propose a new replication-based padding strategy, which allows us to better leverage the information from clients with high-quality datasets. This method ensures that valuable information from high rank clients is retained during the aggregation process, accelerating the convergence speed and enhancing the overall prediction quality of the global model.

Rethinking Pruning Large Language Models: Benefits and Pitfalls of Reconstruction Error Minimization

This work suggests fundamentally rethinking the current practice of pruning large language models (LLMs). The way it is done is by divide and conquer: split the model into submodels, sequentially prune them, and reconstruct predictions of the dense counterparts on small calibration data one at a time; the final model is obtained simply by putting the resulting sparse submodels together. While this approach enables pruning under memory constraints, it generates high reconstruction errors. In this work, we first present an array of reconstruction techniques that can significantly reduce this error by more than $90\%$. Unwittingly, however, we discover that minimizing reconstruction error is not always ideal and can overfit the given calibration data, resulting in rather increased language perplexity and poor performance at downstream tasks. We find out that a strategy of self-generating calibration data can mitigate this trade-off between reconstruction and generalization, suggesting new directions in the presence of both benefits and pitfalls of reconstruction for pruning LLMs.

Prefixing Attention Sinks can Mitigate Activation Outliers for Large Language Model Quantization

Despite recent advances in LLM quantization, activation quantization remains to be challenging due to the activation outliers. Conventional remedies, e.g., mixing precisions for different channels, introduce extra overhead and reduce the speedup. In this work, we develop a simple yet effective strategy to facilitate per-tensor activation quantization by preventing the generation of problematic tokens. Precisely, we propose a method to find a set of key-value cache, coined CushionCache, which mitigates outliers in subsequent tokens when inserted as a prefix. CushionCache works in two steps: First, we greedily search for a prompt token sequence that minimizes the maximum activation values in subsequent tokens. Then, we further tune the token cache to regularize the activations of subsequent tokens to be more quantization-friendly. The proposed method successfully addresses activation outliers of LLMs, providing a substantial performance boost for per-tensor activation quantization methods. We thoroughly evaluate our method over a wide range of models and benchmarks and find that it significantly surpasses the established baseline of per-tensor W8A8 quantization and can be seamlessly integrated with the recent activation quantization method.