In recent years, foundation models (FMs) such as large language models (LLMs) and latent diffusion models (LDMs) have profoundly impacted diverse sectors, including music. This comprehensive review examines state-of-the-art (SOTA) pre-trained models and foundation models in music, spanning from representation learning, generative learning and multimodal learning. We first contextualise the significance of music in various industries and trace the evolution of AI in music. By delineating the modalities targeted by foundation models, we discover many of the music representations are underexplored in FM development. Then, emphasis is placed on the lack of versatility of previous methods on diverse music applications, along with the potential of FMs in music understanding, generation and medical application. By comprehensively exploring the details of the model pre-training paradigm, architectural choices, tokenisation, finetuning methodologies and controllability, we emphasise the important topics that should have been well explored, like instruction tuning and in-context learning, scaling law and emergent ability, as well as long-sequence modelling etc. A dedicated section presents insights into music agents, accompanied by a thorough analysis of datasets and evaluations essential for pre-training and downstream tasks. Finally, by underscoring the vital importance of ethical considerations, we advocate that following research on FM for music should focus more on such issues as interpretability, transparency, human responsibility, and copyright issues. The paper offers insights into future challenges and trends on FMs for music, aiming to shape the trajectory of human-AI collaboration in the music realm.

Backdoor attacks change a small portion of training data by introducing hand-crafted triggers and rewiring the corresponding labels towards a desired target class. Training on such data injects a backdoor which causes malicious inference in selected test samples. Most defenses mitigate such attacks through various modifications of the discriminative learning procedure. In contrast, this paper explores an approach based on generative modelling of per-class distributions in a self-supervised representation space. Interestingly, these representations get either preserved or heavily disturbed under recent backdoor attacks. In both cases, we find that per-class generative models allow to detect poisoned data and cleanse the dataset. Experiments show that training on cleansed dataset greatly reduces the attack success rate and retains the accuracy on benign inputs.

In visual Reinforcement Learning (RL), learning from pixel-based observations poses significant challenges on sample efficiency, primarily due to the complexity of extracting informative state representations from high-dimensional data. Previous methods such as contrastive-based approaches have made strides in improving sample efficiency but fall short in modeling the nuanced evolution of states. To address this, we introduce MOOSS, a novel framework that leverages a temporal contrastive objective with the help of graph-based spatial-temporal masking to explicitly model state evolution in visual RL. Specifically, we propose a self-supervised dual-component strategy that integrates (1) a graph construction of pixel-based observations for spatial-temporal masking, coupled with (2) a multi-level contrastive learning mechanism that enriches state representations by emphasizing temporal continuity and change of states. MOOSS advances the understanding of state dynamics by disrupting and learning from spatial-temporal correlations, which facilitates policy learning. Our comprehensive evaluation on multiple continuous and discrete control benchmarks shows that MOOSS outperforms previous state-of-the-art visual RL methods in terms of sample efficiency, demonstrating the effectiveness of our method. Our code is released at https://github.com/jsun57/MOOSS.

One of the grand challenges of artificial general intelligence is developing agents capable of conducting scientific research and discovering new knowledge. While frontier models have already been used as aides to human scientists, e.g. for brainstorming ideas, writing code, or prediction tasks, they still conduct only a small part of the scientific process. This paper presents the first comprehensive framework for fully automatic scientific discovery, enabling frontier large language models to perform research independently and communicate their findings. We introduce The AI Scientist, which generates novel research ideas, writes code, executes experiments, visualizes results, describes its findings by writing a full scientific paper, and then runs a simulated review process for evaluation. In principle, this process can be repeated to iteratively develop ideas in an open-ended fashion, acting like the human scientific community. We demonstrate its versatility by applying it to three distinct subfields of machine learning: diffusion modeling, transformer-based language modeling, and learning dynamics. Each idea is implemented and developed into a full paper at a cost of less than $15 per paper. To evaluate the generated papers, we design and validate an automated reviewer, which we show achieves near-human performance in evaluating paper scores. The AI Scientist can produce papers that exceed the acceptance threshold at a top machine learning conference as judged by our automated reviewer. This approach signifies the beginning of a new era in scientific discovery in machine learning: bringing the transformative benefits of AI agents to the entire research process of AI itself, and taking us closer to a world where endless affordable creativity and innovation can be unleashed on the world's most challenging problems. Our code is open-sourced at https://github.com/SakanaAI/AI-Scientist

Multimodal VAEs often struggle with generating high-quality outputs, a challenge that extends beyond the inherent limitations of the VAE framework. The core issue lies in the restricted joint representation of the latent space, particularly when complex modalities like images are involved. Feedforward decoders, commonly used for these intricate modalities, inadvertently constrain the joint latent space, leading to a degradation in the quality of the other modalities as well. Although recent studies have shown improvement by introducing modality-specific representations, the issue remains significant. In this work, we demonstrate that incorporating a flexible diffusion decoder specifically for the image modality not only enhances the generation quality of the images but also positively impacts the performance of the other modalities that rely on feedforward decoders. This approach addresses the limitations imposed by conventional joint representations and opens up new possibilities for improving multimodal generation tasks using the multimodal VAE framework.

This thesis investigates three areas targeted at improving the reliability of machine learning; fairness in machine learning, strategic classification, and algorithmic robustness. Each of these domains has special properties or structure that can complicate learning. A theme throughout this thesis is thinking about ways in which a'plain' empirical risk minimization algorithm will be misleading or ineffective because of a mis-match between classical learning theory assumptions and specific properties of some data distribution in the wild. The overarching research goal for these related topics is to provide a crisp mathematical model for each learning scenario that exposes different failure modes and makes trade-offs between important metrics explicit in order to provide algorithmic advice or recommendations to practitioners and expose gaps for future research. By tuning our learning algorithms to be more distribution specific in these scenarios, the resulting learned system will exhibit higher utility and avoid catastrophic failure modes. This research is grounded in the theory of machine learning and is fundamentally mathematical in nature, with empirical support when appropriate. Theory is particularly important in these sensitive domains as it is unclear which poor behavior in deployed systems is a natural or benign consequence of a learning system with the underlying distribution,contrasting with problematic but correctable behavior caused by an error in algorithm design or implementation, how to mitigate these issues, or what a successful outcome even looks like in each problem. Theoretical understanding in each domain can help guide best practices and allow for the design of effective, reliable, and robust systems.

Text classification, a classic task in natural language processing (NLP), involves assigning predefined categories to textual data and is crucial for applications ranging from sentiment analysis to spam detection. This thesis advances text classification by harnessing the intrinsic knowledge of Pretrained Language Models (PLMs) to address three challenging scenarios: distractor selection for multiple-choice cloze questions, improving robustness for prompt-based zero-shot text classification, and demonstration selection for retrieval-based in-context learning. Firstly, we focus on selecting distractors for multiple-choice cloze questions, ensuring that they are misleading yet incorrect. We assess the relationship between human experts' annotations (accept/reject) and various features, including context-free features (e.g., word frequency) and context-sensitive features (e.g., conditional probabilities of fillin-the-blank words). We utilize pretrained embeddings and follow annotation instructions for context-free feature design, and we find that using contextualized word representations from PLMs as features drastically improves performance over traditional feature-based models, even rivaling human performance (Chapter 3).

Recent advancements for large-scale pre-training with neural signals such as electroencephalogram (EEG) have shown promising results, significantly boosting the development of brain-computer interfaces (BCIs) and healthcare. However, these pre-trained models often require full fine-tuning on each downstream task to achieve substantial improvements, limiting their versatility and usability, and leading to considerable resource wastage. To tackle these challenges, we propose NeuroLM, the first multi-task foundation model that leverages the capabilities of Large Language Models (LLMs) by regarding EEG signals as a foreign language, endowing the model with multi-task learning and inference capabilities. Our approach begins with learning a text-aligned neural tokenizer through vector-quantized temporal-frequency prediction, which encodes EEG signals into discrete neural tokens. These EEG tokens, generated by the frozen vector-quantized (VQ) encoder, are then fed into an LLM that learns causal EEG information via multi-channel autoregression. Consequently, NeuroLM can understand both EEG and language modalities. Finally, multi-task instruction tuning adapts NeuroLM to various downstream tasks. We are the first to demonstrate that, by specific incorporation with LLMs, NeuroLM unifies diverse EEG tasks within a single model through instruction tuning. The largest variant NeuroLM-XL has record-breaking 1.7B parameters for EEG signal processing, and is pre-trained on a large-scale corpus comprising approximately 25,000-hour EEG data. When evaluated on six diverse downstream datasets, NeuroLM showcases the huge potential of this multi-task learning paradigm.

We present GameNGen, the first game engine powered entirely by a neural model that enables real-time interaction with a complex environment over long trajectories at high quality. GameNGen can interactively simulate the classic game DOOM at over 20 frames per second on a single TPU. Next frame prediction achieves a PSNR of 29.4, comparable to lossy JPEG compression. Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation. GameNGen is trained in two phases: (1) an RL-agent learns to play the game and the training sessions are recorded, and (2) a diffusion model is trained to produce the next frame, conditioned on the sequence of past frames and actions. Conditioning augmentations enable stable auto-regressive generation over long trajectories.

The conventional approach to the fronthaul design for cell-free massive MIMO system follows the compress-and-precode (CP) paradigm. Accordingly, encoded bits and precoding coefficients are shared by the distributed unit (DU) on the fronthaul links, and precoding takes place at the radio units (RUs). Previous theoretical work has shown that CP can be potentially improved by a significant margin by precode-and-compress (PC) methods, in which all baseband processing is carried out at the DU, which compresses the precoded signals for transmission on the fronthaul links. The theoretical performance gain of PC methods are particularly pronounced when the DU implements multivariate quantization (MQ), applying joint quantization across the signals for all the RUs. However, existing solutions for MQ are characterized by a computational complexity that grows exponentially with the sum-fronthaul capacity from the DU to all RUs. This work sets out to design scalable MQ strategies for PC-based cell-free massive MIMO systems. For the low-fronthaul capacity regime, we present alpha-parallel MQ (alpha-PMQ), whose complexity is exponential only in the fronthaul capacity towards an individual RU, while performing close to full MQ. alpha-PMQ tailors MQ to the topology of the network by allowing for parallel local quantization steps for RUs that do not interfere too much with each other. For the high-fronthaul capacity regime, we then introduce neural MQ, which replaces the exhaustive search in MQ with gradient-based updates for a neural-network-based decoder, attaining a complexity that grows linearly with the sum-fronthaul capacity. Numerical results demonstrate that the proposed scalable MQ strategies outperform CP for both the low and high-fronthaul capacity regimes at the cost of increased computational complexity at the DU (but not at the RUs).