Abstract:Large language models learn and continually learn through the accumulation of gradient-based updates, but how individual pieces of new information affect existing knowledge, leading to both beneficial generalization and problematic hallucination, remains poorly understood. We demonstrate that when learning new information, LLMs exhibit a "priming" effect: learning a new fact can cause the model to inappropriately apply that knowledge in unrelated contexts. To systematically study this phenomenon, we introduce "Outlandish," a carefully curated dataset of 1320 diverse text samples designed to probe how new knowledge permeates through an LLM's existing knowledge base. Using this dataset, we show that the degree of priming after learning new information can be predicted by measuring the token probability of key words before learning. This relationship holds robustly across different model architectures (PALM-2, Gemma, Llama), sizes, and training stages. Finally, we develop two novel techniques to modulate how new knowledge affects existing model behavior: (1) a ``stepping-stone'' text augmentation strategy and (2) an ``ignore-k'' update pruning method. These approaches reduce undesirable priming effects by 50-95\% while preserving the model's ability to learn new information. Our findings provide both empirical insights into how LLMs learn and practical tools for improving the specificity of knowledge insertion in language models. Further materials: https://sunchipsster1.github.io/projects/outlandish/
Abstract:Long context processing is critical for the adoption of LLMs, but existing methods often introduce architectural complexity that hinders their practical adoption. Gisting, an in-context compression method with no architectural modification to the decoder transformer, is a promising approach due to its simplicity and compatibility with existing frameworks. While effective for short instructions, we demonstrate that gisting struggles with longer contexts, with significant performance drops even at minimal compression rates. Surprisingly, a simple average pooling baseline consistently outperforms gisting. We analyze the limitations of gisting, including information flow interruptions, capacity limitations and the inability to restrict its attention to subsets of the context. Motivated by theoretical insights into the performance gap between gisting and average pooling, and supported by extensive experimentation, we propose GistPool, a new in-context compression method. GistPool preserves the simplicity of gisting, while significantly boosting its performance on long context compression tasks.
Abstract:Neural Audio Synthesis (NAS) models offer interactive musical control over high-quality, expressive audio generators. While these models can operate in real-time, they often suffer from high latency, making them unsuitable for intimate musical interaction. The impact of architectural choices in deep learning models on audio latency remains largely unexplored in the NAS literature. In this work, we investigate the sources of latency and jitter typically found in interactive NAS models. We then apply this analysis to the task of timbre transfer using RAVE, a convolutional variational autoencoder for audio waveforms introduced by Caillon et al. in 2021. Finally, we present an iterative design approach for optimizing latency. This culminates with a model we call BRAVE (Bravely Realtime Audio Variational autoEncoder), which is low-latency and exhibits better pitch and loudness replication while showing timbre modification capabilities similar to RAVE. We implement it in a specialized inference framework for low-latency, real-time inference and present a proof-of-concept audio plugin compatible with audio signals from musical instruments. We expect the challenges and guidelines described in this document to support NAS researchers in designing models for low-latency inference from the ground up, enriching the landscape of possibilities for musicians.
Abstract:What happens when a new piece of knowledge is introduced into the training data and how long does it last while a large language model (LM) continues to train? We investigate this question by injecting facts into LMs from a new probing dataset, "Outlandish", which is designed to permit the testing of a spectrum of different fact types. When studying how robust these memories are, there appears to be a sweet spot in the spectrum of fact novelty between consistency with world knowledge and total randomness, where the injected memory is the most enduring. Specifically we show that facts that conflict with common knowledge are remembered for tens of thousands of training steps, while prompts not conflicting with common knowledge (mundane), as well as scrambled prompts (randomly jumbled) are both forgotten much more rapidly. Further, knowledge-conflicting facts can "prime'' how the language model hallucinates on logically unrelated prompts, showing their propensity for non-target generalization, while both mundane and randomly jumbled facts prime significantly less. Finally, we show that impacts of knowledge-conflicting facts in LMs, though they can be long lasting, can be largely erased by novel application of multi-step sparse updates, even while the training ability of the model is preserved. As such, this very simple procedure has direct implications for mitigating the effects of data poisoning in training.
Abstract:This paper presents an examination of State Space Models (SSM) and Koopman-based deep learning methods for modelling the dynamics of both linear and non-linear stiff strings. Through experiments with datasets generated under different initial conditions and sample rates, we assess the capacity of these models to accurately model the complex behaviours observed in string dynamics. Our findings indicate that our proposed Koopman-based model performs as well as or better than other existing approaches in non-linear cases for long-sequence modelling. We inform the design of these architectures with the structure of the problems at hand. Although challenges remain in extending model predictions beyond the training horizon (i.e., extrapolation), the focus of our investigation lies in the models' ability to generalise across different initial conditions within the training time interval. This research contributes insights into the physical modelling of dynamical systems (in particular those addressing musical acoustics) by offering a comparative overview of these and previous methods and introducing innovative strategies for model improvement. Our results highlight the efficacy of these models in simulating non-linear dynamics and emphasise their wide-ranging applicability in accurately modelling dynamical systems over extended sequences.
Abstract:In modern deep learning, the models are learned by applying gradient updates using an optimizer, which transforms the updates based on various statistics. Optimizers are often hand-designed and tuning their hyperparameters is a big part of the training process. Learned optimizers have shown some initial promise, but are generally unsuccessful as a general optimization mechanism applicable to every problem. In this work we explore a different direction: instead of learning general optimizers, we instead specialize them to a specific training environment. We propose a novel optimizer technique that learns a layer-specific linear combination of update directions provided by a set of base optimizers, effectively adapting its strategy to the specific model and dataset. When evaluated on image classification tasks, this specialized optimizer significantly outperforms both traditional off-the-shelf methods such as Adam, as well as existing general learned optimizers. Moreover, it demonstrates robust generalization with respect to model initialization, evaluating on unseen datasets, and training durations beyond its meta-training horizon.
Abstract:Recent research has demonstrated that transformers, particularly linear attention models, implicitly execute gradient-descent-like algorithms on data provided in-context during their forward inference step. However, their capability in handling more complex problems remains unexplored. In this paper, we prove that any linear transformer maintains an implicit linear model and can be interpreted as performing a variant of preconditioned gradient descent. We also investigate the use of linear transformers in a challenging scenario where the training data is corrupted with different levels of noise. Remarkably, we demonstrate that for this problem linear transformers discover an intricate and highly effective optimization algorithm, surpassing or matching in performance many reasonable baselines. We reverse-engineer this algorithm and show that it is a novel approach incorporating momentum and adaptive rescaling based on noise levels. Our findings show that even linear transformers possess the surprising ability to discover sophisticated optimization strategies.
Abstract:Tone Transfer is a novel deep-learning technique for interfacing a sound source with a synthesizer, transforming the timbre of audio excerpts while keeping their musical form content. Due to its good audio quality results and continuous controllability, it has been recently applied in several audio processing tools. Nevertheless, it still presents several shortcomings related to poor sound diversity, and limited transient and dynamic rendering, which we believe hinder its possibilities of articulation and phrasing in a real-time performance context. In this work, we present a discussion on current Tone Transfer architectures for the task of controlling synthetic audio with musical instruments and discuss their challenges in allowing expressive performances. Next, we introduce Envelope Learning, a novel method for designing Tone Transfer architectures that map musical events using a training objective at the synthesis parameter level. Our technique can render note beginnings and endings accurately and for a variety of sounds; these are essential steps for improving musical articulation, phrasing, and sound diversity with Tone Transfer. Finally, we implement a VST plugin for real-time live use and discuss possibilities for improvement.
Abstract:Differentiable digital signal processing (DDSP) techniques, including methods for audio synthesis, have gained attention in recent years and lend themselves to interpretability in the parameter space. However, current differentiable synthesis methods have not explicitly sought to model the transient portion of signals, which is important for percussive sounds. In this work, we present a unified synthesis framework aiming to address transient generation and percussive synthesis within a DDSP framework. To this end, we propose a model for percussive synthesis that builds on sinusoidal modeling synthesis and incorporates a modulated temporal convolutional network for transient generation. We use a modified sinusoidal peak picking algorithm to generate time-varying non-harmonic sinusoids and pair it with differentiable noise and transient encoders that are jointly trained to reconstruct drumset sounds. We compute a set of reconstruction metrics using a large dataset of acoustic and electronic percussion samples that show that our method leads to improved onset signal reconstruction for membranophone percussion instruments.
Abstract:Transformers have become the dominant model in deep learning, but the reason for their superior performance is poorly understood. Here, we hypothesize that the strong performance of Transformers stems from an architectural bias towards mesa-optimization, a learned process running within the forward pass of a model consisting of the following two steps: (i) the construction of an internal learning objective, and (ii) its corresponding solution found through optimization. To test this hypothesis, we reverse-engineer a series of autoregressive Transformers trained on simple sequence modeling tasks, uncovering underlying gradient-based mesa-optimization algorithms driving the generation of predictions. Moreover, we show that the learned forward-pass optimization algorithm can be immediately repurposed to solve supervised few-shot tasks, suggesting that mesa-optimization might underlie the in-context learning capabilities of large language models. Finally, we propose a novel self-attention layer, the mesa-layer, that explicitly and efficiently solves optimization problems specified in context. We find that this layer can lead to improved performance in synthetic and preliminary language modeling experiments, adding weight to our hypothesis that mesa-optimization is an important operation hidden within the weights of trained Transformers.