Oscar Chang

Oscar Chang, Hank Liao, Dmitriy Serdyuk et al. · cmu

Visual speech recognition models extract visual features in a hierarchical manner. At the lower level, there is a visual front-end with a limited temporal receptive field that processes the raw pixels depicting the lips or faces. At the higher level, there is an encoder that attends to the embeddings produced by the front-end over a large temporal receptive field. Previous work has focused on improving the visual front-end of the model to extract more useful features for speech recognition. Surprisingly, our work shows that complex visual front-ends are not necessary. Instead of allocating resources to a sophisticated visual front-end, we find that a linear visual front-end paired with a larger Conformer encoder results in lower latency, more efficient memory usage, and improved WER performance. We achieve a new state-of-the-art of 12.8% WER for visual speech recognition on the TED LRS3 dataset, which rivals the performance of audio-only models from just four years ago.

Tiansheng Huang, Virat Shejwalkar, Oscar Chang et al.

Instilling reasoning capabilities in large models (LMs) using reasoning training (RT) significantly improves LMs' performances. Thus Audio Reasoning Models (ARMs), i.e., audio LMs that can reason, are becoming increasingly popular. However, no work has studied the safety of ARMs against jailbreak attacks that aim to elicit harmful responses from target models. To this end, first, we show that standard RT with appropriate safety reasoning data can protect ARMs from vanilla audio jailbreaks, but cannot protect them against our proposed simple yet effective jailbreaks. We show that this is because of the significant representation drift between vanilla and advanced jailbreaks which forces the target ARMs to emit harmful responses. Based on this observation, we propose Rebellion, a robust RT that trains ARMs to be robust to the worst-case representation drift. All our results are on Qwen2-Audio; they demonstrate that Rebellion: 1) can protect against advanced audio jailbreaks without compromising performance on benign tasks, and 2) significantly improves accuracy-safety trade-off over standard RT method.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Gemma Team, Thomas Mesnard, Cassidy Hardin et al. · deepmind

This work introduces Gemma, a family of lightweight, state-of-the art open models built from the research and technology used to create Gemini models. Gemma models demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. We release two sizes of models (2 billion and 7 billion parameters), and provide both pretrained and fine-tuned checkpoints. Gemma outperforms similarly sized open models on 11 out of 18 text-based tasks, and we present comprehensive evaluations of safety and responsibility aspects of the models, alongside a detailed description of model development. We believe the responsible release of LLMs is critical for improving the safety of frontier models, and for enabling the next wave of LLM innovations.

Oscar Chang, Dongseong Hwang, Olivier Siohan

In streaming settings, speech recognition models have to map sub-sequences of speech to text before the full audio stream becomes available. However, since alignment information between speech and text is rarely available during training, models need to learn it in a completely self-supervised way. In practice, the exponential number of possible alignments makes this extremely challenging, with models often learning peaky or sub-optimal alignments. Prima facie, the exponential nature of the alignment space makes it difficult to even quantify the uncertainty of a model's alignment distribution. Fortunately, it has been known for decades that the entropy of a probabilistic finite state transducer can be computed in time linear to the size of the transducer via a dynamic programming reduction based on semirings. In this work, we revisit the entropy semiring for neural speech recognition models, and show how alignment entropy can be used to supervise models through regularization or distillation. We also contribute an open-source implementation of CTC and RNN-T in the semiring framework that includes numerically stable and highly parallel variants of the entropy semiring. Empirically, we observe that the addition of alignment distillation improves the accuracy and latency of an already well-optimized teacher-student distillation model, achieving state-of-the-art performance on the Librispeech dataset in the streaming scenario.

Oscar Chang, Lampros Flokas, Hod Lipson

Hypernetworks are meta neural networks that generate weights for a main neural network in an end-to-end differentiable manner. Despite extensive applications ranging from multi-task learning to Bayesian deep learning, the problem of optimizing hypernetworks has not been studied to date. We observe that classical weight initialization methods like Glorot & Bengio (2010) and He et al. (2015), when applied directly on a hypernet, fail to produce weights for the mainnet in the correct scale. We develop principled techniques for weight initialization in hypernets, and show that they lead to more stable mainnet weights, lower training loss, and faster convergence.

Oscar Chang, Lampros Flokas, Hod Lipson et al.

SATNet is an award-winning MAXSAT solver that can be used to infer logical rules and integrated as a differentiable layer in a deep neural network. It had been shown to solve Sudoku puzzles visually from examples of puzzle digit images, and was heralded as an impressive achievement towards the longstanding AI goal of combining pattern recognition with logical reasoning. In this paper, we clarify SATNet's capabilities by showing that in the absence of intermediate labels that identify individual Sudoku digit images with their logical representations, SATNet completely fails at visual Sudoku (0% test accuracy). More generally, the failure can be pinpointed to its inability to learn to assign symbols to perceptual phenomena, also known as the symbol grounding problem, which has long been thought to be a prerequisite for intelligent agents to perform real-world logical reasoning. We propose an MNIST based test as an easy instance of the symbol grounding problem that can serve as a sanity check for differentiable symbolic solvers in general. Naive applications of SATNet on this test lead to performance worse than that of models without logical reasoning capabilities. We report on the causes of SATNet's failure and how to prevent them.

Laura Weidinger, Joslyn Barnhart, Jenny Brennan et al.

Safety and responsibility evaluations of advanced AI models are a critical but developing field of research and practice. In the development of Google DeepMind's advanced AI models, we innovated on and applied a broad set of approaches to safety evaluation. In this report, we summarise and share elements of our evolving approach as well as lessons learned for a broad audience. Key lessons learned include: First, theoretical underpinnings and frameworks are invaluable to organise the breadth of risk domains, modalities, forms, metrics, and goals. Second, theory and practice of safety evaluation development each benefit from collaboration to clarify goals, methods and challenges, and facilitate the transfer of insights between different stakeholders and disciplines. Third, similar key methods, lessons, and institutions apply across the range of concerns in responsibility and safety - including established and emerging harms. For this reason it is important that a wide range of actors working on safety evaluation and safety research communities work together to develop, refine and implement novel evaluation approaches and best practices, rather than operating in silos. The report concludes with outlining the clear need to rapidly advance the science of evaluations, to integrate new evaluations into the development and governance of AI, to establish scientifically-grounded norms and standards, and to promote a robust evaluation ecosystem.

Oscar Chang, Otavio Braga, Hank Liao et al.

It has been shown that learning audiovisual features can lead to improved speech recognition performance over audio-only features, especially for noisy speech. However, in many common applications, the visual features are partially or entirely missing, e.g.~the speaker might move off screen. Multi-modal models need to be robust: missing video frames should not degrade the performance of an audiovisual model to be worse than that of a single-modality audio-only model. While there have been many attempts at building robust models, there is little consensus on how robustness should be evaluated. To address this, we introduce a framework that allows claims about robustness to be evaluated in a precise and testable way. We also conduct a systematic empirical study of the robustness of common audiovisual speech recognition architectures on a range of acoustic noise conditions and test suites. Finally, we show that an architecture-agnostic solution based on cascades can consistently achieve robustness to missing video, even in settings where existing techniques for robustness like dropout fall short.

Oscar Chang, Dung N. Tran, Kazuhito Koishida

Generalization remains a major problem in supervised learning of single-channel speech enhancement. In this work, we propose learnable loss mixup (LLM), a simple and effortless training diagram, to improve the generalization of deep learning-based speech enhancement models. Loss mixup, of which learnable loss mixup is a special variant, optimizes a mixture of the loss functions of random sample pairs to train a model on virtual training data constructed from these pairs of samples. In learnable loss mixup, by conditioning on the mixed data, the loss functions are mixed using a non-linear mixing function automatically learned via neural parameterization. Our experimental results on the VCTK benchmark show that learnable loss mixup achieves 3.26 PESQ, outperforming the state-of-the-art.

Oscar Chang, Hod Lipson

The success of gradient-based meta-learning is primarily attributed to its ability to leverage related tasks to learn task-invariant information. However, the absence of interactions between different tasks in the inner loop leads to task-specific over-fitting in the initial phase of meta-training. While this is eventually corrected by the presence of these interactions in the outer loop, it comes at a significant cost of slower meta-learning. To address this limitation, we explicitly encode task relatedness via an inner loop regularization mechanism inspired by multi-task learning. Our algorithm shares gradient information from previously encountered tasks as well as concurrent tasks in the same task batch, and scales their contribution with meta-learned parameters. We show using two popular few-shot classification datasets that gradient sharing enables meta-learning under bigger inner loop learning rates and can accelerate the meta-training process by up to 134%.

Abhishek Jindal, Dmitry Kalashnikov, Oscar Chang et al.

When AI interacts with the physical world -- as a robot or an assistive agent -- new safety challenges emerge beyond those of purely ``digital AI". In such interactions, the potential for physical harm is direct and immediate. How well do state-of-the-art foundation models understand common-sense facts about physical safety, e.g. that a box may be too heavy to lift, or that a hot cup of coffee should not be handed to a child? In this paper, our contributions are three-fold: first, we develop a highly scalable approach to continuous physical safety benchmarking of Embodied AI systems, grounded in real-world injury narratives and operational safety constraints. To probe multi-modal safety understanding, we turn these narratives and constraints into photorealistic images and videos capturing transitions from safe to unsafe states, using advanced generative models. Secondly, we comprehensively analyze the ability of major foundation models to perceive risks, reason about safety, and trigger interventions; this yields multi-faceted insights into their deployment readiness for safety-critical agentic applications. Finally, we develop a post-training paradigm to teach models to explicitly reason about embodiment-specific safety constraints provided through system instructions. The resulting models generate thinking traces that make safety reasoning interpretable and transparent, achieving state of the art performance in constraint satisfaction evaluations. The benchmark will be released at https://asimov-benchmark.github.io/v2

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Oscar Chang, Hod Lipson

Weight-sharing plays a significant role in the success of many deep neural networks, by increasing memory efficiency and incorporating useful inductive priors about the problem into the network. But understanding how weight-sharing can be used effectively in general is a topic that has not been studied extensively. Chen et al. [2015] proposed HashedNets, which augments a multi-layer perceptron with a hash table, as a method for neural network compression. We generalize this method into a framework (ArbNets) that allows for efficient arbitrary weight-sharing, and use it to study the role of weight-sharing in neural networks. We show that common neural networks can be expressed as ArbNets with different hash functions. We also present two novel hash functions, the Dirichlet hash and the Neighborhood hash, and use them to demonstrate experimentally that balanced and deterministic weight-sharing helps with the performance of a neural network.

Oscar Chang, Yuling Yao, David Williams-King et al.

Two main obstacles preventing the widespread adoption of variational Bayesian neural networks are the high parameter overhead that makes them infeasible on large networks, and the difficulty of implementation, which can be thought of as "programming overhead." MC dropout [Gal and Ghahramani, 2016] is popular because it sidesteps these obstacles. Nevertheless, dropout is often harmful to model performance when used in networks with batch normalization layers [Li et al., 2018], which are an indispensable part of modern neural networks. We construct a general variational family for ensemble-based Bayesian neural networks that encompasses dropout as a special case. We further present two specific members of this family that work well with batch normalization layers, while retaining the benefits of low parameter and programming overhead, comparable to non-Bayesian training. Our proposed methods improve predictive accuracy and achieve almost perfect calibration on a ResNet-18 trained with ImageNet.

Oscar Chang, Hod Lipson

We present seven myths commonly believed to be true in machine learning research, circa Feb 2019. This is an archival copy of the blog post at https://crazyoscarchang.github.io/2019/02/16/seven-myths-in-machine-learning-research/ Myth 1: TensorFlow is a Tensor manipulation library Myth 2: Image datasets are representative of real images found in the wild Myth 3: Machine Learning researchers do not use the test set for validation Myth 4: Every datapoint is used in training a neural network Myth 5: We need (batch) normalization to train very deep residual networks Myth 6: Attention $>$ Convolution Myth 7: Saliency maps are robust ways to interpret neural networks

Oscar Chang, Robert Kwiatkowski, Siyuan Chen et al.

We show that it is possible to reduce a high-dimensional object like a neural network agent into a low-dimensional vector representation with semantic meaning that we call agent embeddings, akin to word or face embeddings. This can be done by collecting examples of existing networks, vectorizing their weights, and then learning a generative model over the weight space in a supervised fashion. We investigate a pole-balancing task, Cart-Pole, as a case study and show that multiple new pole-balancing networks can be generated from their agent embeddings without direct access to training data from the Cart-Pole simulator. In general, the learned embedding space is helpful for mapping out the space of solutions for a given task. We observe in the case of Cart-Pole the surprising finding that good agents make different decisions despite learning similar representations, whereas bad agents make similar (bad) decisions while learning dissimilar representations. Linearly interpolating between the latent embeddings for a good agent and a bad agent yields an agent embedding that generates a network with intermediate performance, where the performance can be tuned according to the coefficient of interpolation. Linear extrapolation in the latent space also results in performance boosts, up to a point.

Payel Das, Kahini Wadhawan, Oscar Chang et al.

Given the emerging global threat of antimicrobial resistance, new methods for next-generation antimicrobial design are urgently needed. We report a peptide generation framework PepCVAE, based on a semi-supervised variational autoencoder (VAE) model, for designing novel antimicrobial peptide (AMP) sequences. Our model learns a rich latent space of the biological peptide context by taking advantage of abundant, unlabeled peptide sequences. The model further learns a disentangled antimicrobial attribute space by using the feedback from a jointly trained AMP classifier that uses limited labeled instances. The disentangled representation allows for controllable generation of AMPs. Extensive analysis of the PepCVAE-generated sequences reveals superior performance of our model in comparison to a plain VAE, as PepCVAE generates novel AMP sequences with higher long-range diversity, while being closer to the training distribution of biological peptides. These features are highly desired in next-generation antimicrobial design.

Oscar Chang, Hod Lipson

Self-replication is a key aspect of biological life that has been largely overlooked in Artificial Intelligence systems. Here we describe how to build and train self-replicating neural networks. The network replicates itself by learning to output its own weights. The network is designed using a loss function that can be optimized with either gradient-based or non-gradient-based methods. We also describe a method we call regeneration to train the network without explicit optimization, by injecting the network with predictions of its own parameters. The best solution for a self-replicating network was found by alternating between regeneration and optimization steps. Finally, we describe a design for a self-replicating neural network that can solve an auxiliary task such as MNIST image classification. We observe that there is a trade-off between the network's ability to classify images and its ability to replicate, but training is biased towards increasing its specialization at image classification at the expense of replication. This is analogous to the trade-off between reproduction and other tasks observed in nature. We suggest that a self-replication mechanism for artificial intelligence is useful because it introduces the possibility of continual improvement through natural selection.

Robert Kwiatkowski, Oscar Chang

In this paper we introduce a novel method of gradient normalization and decay with respect to depth. Our method leverages the simple concept of normalizing all gradients in a deep neural network, and then decaying said gradients with respect to their depth in the network. Our proposed normalization and decay techniques can be used in conjunction with most current state of the art optimizers and are a very simple addition to any network. This method, although simple, showed improvements in convergence time on state of the art networks such as DenseNet and ResNet on image classification tasks, as well as on an LSTM for natural language processing tasks.