To improve performance in contemporary deep learning, one is interested in scaling up the neural network in terms of both the number and the size of the layers. When ramping up the width of a single layer, graceful scaling of training has been linked to the need to normalize the weights and their updates in the "natural norm" particular to that layer. In this paper, we significantly generalize this idea by defining the modular norm, which is the natural norm on the full weight space of any neural network architecture. The modular norm is defined recursively in tandem with the network architecture itself. We show that the modular norm has several promising applications. On the practical side, the modular norm can be used to normalize the updates of any base optimizer so that the learning rate becomes transferable across width and depth. This means that the user does not need to compute optimizer-specific scale factors in order to scale training. On the theoretical side, we show that for any neural network built from "well-behaved" atomic modules, the gradient of the network is Lipschitz-continuous in the modular norm, with the Lipschitz constant admitting a simple recursive formula. This characterization opens the door to porting standard ideas in optimization theory over to deep learning. We have created a Python package called Modula that automatically normalizes weight updates in the modular norm of the architecture. The package is available via "pip install modula" with source code at https://github.com/jxbz/modula.

The scalability of deep learning models is fundamentally limited by computing resources, memory, and communication. Although methods like low-rank adaptation (LoRA) have reduced the cost of model finetuning, its application in model pre-training remains largely unexplored. This paper explores extending LoRA to model pre-training, identifying the inherent constraints and limitations of standard LoRA in this context. We introduce LoRA-the-Explorer (LTE), a novel bi-level optimization algorithm designed to enable parallel training of multiple low-rank heads across computing nodes, thereby reducing the need for frequent synchronization. Our approach includes extensive experimentation on vision transformers using various vision datasets, demonstrating that LTE is competitive with standard pre-training.

We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the ways by which different neural networks represent data are becoming more aligned. Next, we demonstrate convergence across data modalities: as vision models and language models get larger, they measure distance between datapoints in a more and more alike way. We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality. We term such a representation the platonic representation and discuss several possible selective pressures toward it. Finally, we discuss the implications of these trends, their limitations, and counterexamples to our analysis.

This work examines the challenges of training neural networks using vector quantization using straight-through estimation. We find that a primary cause of training instability is the discrepancy between the model embedding and the code-vector distribution. We identify the factors that contribute to this issue, including the codebook gradient sparsity and the asymmetric nature of the commitment loss, which leads to misaligned code-vector assignments. We propose to address this issue via affine re-parameterization of the code vectors. Additionally, we introduce an alternating optimization to reduce the gradient error introduced by the straight-through estimation. Moreover, we propose an improvement to the commitment loss to ensure better alignment between the codebook representation and the model embedding. These optimization methods improve the mathematical approximation of the straight-through estimation and, ultimately, the model performance. We demonstrate the effectiveness of our methods on several common model architectures, such as AlexNet, ResNet, and ViT, across various tasks, including image classification and generative modeling.

We introduce a new approach to image forensics: placing physical refractive objects, which we call totems, into a scene so as to protect any photograph taken of that scene. Totems bend and redirect light rays, thus providing multiple, albeit distorted, views of the scene within a single image. A defender can use these distorted totem pixels to detect if an image has been manipulated. Our approach unscrambles the light rays passing through the totems by estimating their positions in the scene and using their known geometric and material properties. To verify a totem-protected image, we detect inconsistencies between the scene reconstructed from totem viewpoints and the scene's appearance from the camera viewpoint. Such an approach makes the adversarial manipulation task more difficult, as the adversary must modify both the totem and image pixels in a geometrically consistent manner without knowing the physical properties of the totem. Unlike prior learning-based approaches, our method does not require training on datasets of specific manipulations, and instead uses physical properties of the scene and camera to solve the forensics problem.

Modern deep neural networks are highly over-parameterized compared to the data on which they are trained, yet they often generalize remarkably well. A flurry of recent work has asked: why do deep networks not overfit to their training data? In this work, we make a series of empirical observations that investigate and extend the hypothesis that deeper networks are inductively biased to find solutions with lower effective rank embeddings. We conjecture that this bias exists because the volume of functions that maps to low effective rank embedding increases with depth. We show empirically that our claim holds true on finite width linear and non-linear models on practical learning paradigms and show that on natural data, these are often the solutions that generalize well. We then show that the simplicity bias exists at both initialization and after training and is resilient to hyper-parameters and learning methods. We further demonstrate how linear over-parameterization of deep non-linear models can be used to induce low-rank bias, improving generalization performance on CIFAR and ImageNet without changing the modeling capacity.

Communication requires having a common language, a lingua franca, between agents. This language could emerge via a consensus process, but it may require many generations of trial and error. Alternatively, the lingua franca can be given by the environment, where agents ground their language in representations of the observed world. We demonstrate a simple way to ground language in learned representations, which facilitates decentralized multi-agent communication and coordination. We find that a standard representation learning algorithm - autoencoding - is sufficient for arriving at a grounded common language. When agents broadcast these representations, they learn to understand and respond to each other's utterances and achieve surprisingly strong task performance across a variety of multi-agent communication environments.

We present a method for projecting an input image into the space of a class-conditional generative neural network. We propose a method that optimizes for transformation to counteract the model biases in generative neural networks. Specifically, we demonstrate that one can solve for image translation, scale, and global color transformation, during the projection optimization to address the object-center bias and color bias of a Generative Adversarial Network. This projection process poses a difficult optimization problem, and purely gradient-based optimizations fail to find good solutions. We describe a hybrid optimization strategy that finds good projections by estimating transformations and class parameters. We show the effectiveness of our method on real images and further demonstrate how the corresponding projections lead to better editability of these images. The project page and the code is available at https://minyoungg.github.io/GAN-Transform-and-Project/ .

Optical imaging and stimulation are widely used to study biological events. However, scattering processes limit the depth to which externally focused light can penetrate tissue. Optical fibers and waveguides are commonly inserted into tissue when delivering light deeper than a few millimeters. This approach, however, introduces complications arising from tissue damage. In addition, it makes it difficult to steer light. Here, we demonstrate that ultrasound can be used to define and steer the trajectory of light within scattering media by exploiting local pressure differences created by acoustic waves that result in refractive index contrasts. We show that virtual light pipes can be created deep into the tissue (>18 scattering mean free paths). We demonstrate the application of this technology in confining light through mouse brain tissue. This technology is likely extendable to form arbitrary light patterns within tissue, extending both the reach and the flexibility of light-based methods.

We propose feedback adversarial learning (FAL) framework that can improve existing generative adversarial networks by leveraging spatial feedback from the discriminator. We formulate the generation task as a recurrent framework, in which the discriminator’s feedback is integrated into the feedforward path of the generation process. Specifically, the generator conditions on the discriminator’s spatial output response, and its previous generation to improve generation quality over time – allowing the generator to attend and fix its previous mistakes. To effectively utilize the feedback, we propose an adaptive spatial transform layer, which learns to spatially modulate feature maps from its previous generation and the error signal from the discriminator. We demonstrate that one can easily adapt FAL to existing adversarial learning frameworks on a wide range of tasks, including image generation, image-to-image translation, and voxel generation.

In this paper, we address the task of multi-view novel view synthesis, where we are interested in synthesizing a target image with an arbitrary camera pose from given source images. We propose an end-to-end trainable framework that learns to exploit multiple viewpoints to synthesize a novel view without any 3D supervision. Specifically, our model consists of a flow prediction module and a pixel generation module to directly leverage information presented in source views as well as hallucinate missing pixels from statistical priors. To merge the predictions produced by the two modules given multi-view source images, we introduce a self-learned confidence aggregation mechanism. We evaluate our model on images rendered from 3D object models as well as real and synthesized scenes. We demonstrate that our model is able to achieve state-of-the-art results as well as progressively improve its predictions when more source images are available.

Advances in photo editing and manipulation tools have made it significantly easier to create fake imagery. Learning to detect such manipulations, however, remains a challenging problem due to the lack of sufficient amounts of manipulated training data. In this paper, we propose a learning algorithm for detecting visual image manipulations that is trained only using a large dataset of real photographs. The algorithm uses the automatically recorded photo EXIF metadata as supervisory signal for training a model to determine whether an image is self-consistent -- that is, whether its content could have been produced by a single imaging pipeline. We apply this self-consistency model to the task of detecting and localizing image splices. The proposed method obtains state-of-the-art performance on several image forensics benchmarks, despite never seeing any manipulated images at training. That said, it is merely a step in the long quest for a truly general purpose visual forensics tool.

Adversarial learning [6] has significantly influenced the landscape of data-driven representation learning. Classical representation learning requires the task to be explicitly defined, so is the objective function used to learn representations. Adversarial learning, on the other hand, relies on modeling the data distribution via an adversarial game. The generator network learns to generate samples of data that are close to the underlying true data distribution, while the discriminator network learns to determine if the generated data samples are close to this underlying distribution. Richard Feynman’s famous quote (mentioned above) rings true several decades later the ultimate expression to understanding something is the ability to generate it from the ground up. If a Generative Adversarial Network(GAN) is trained successfully, which by itself is not a trivial task, then the network will learn how to apply its learned knowledge of the data distribution to myriad tasks. In the past, GANs have been utilized for learning style transfer [7], anime character generation [8], 3D object generation and reconstruction [9], and image super-resolution [10]. Recently, Isola et al. [11] showed brilliant results about transferring knowledge at the pixel-level conditionally onto output data to generate structurally and semantically coherent output images. A few other works have also followed this line of inquiry and are discussed in the Related Work section.

The tremendous success of ImageNet-trained deep features on a wide range of transfer tasks begs the question: what are the properties of the ImageNet dataset that are critical for learning good, general-purpose features? This work provides an empirical investigation of various facets of this question: Is more pre-training data always better? How does feature quality depend on the number of training examples per class? Does adding more object classes improve performance? For the same data budget, how should the data be split into classes? Is fine-grained recognition necessary for learning good features? Given the same number of training classes, is it better to have coarse classes or fine-grained classes? Which is better: more classes or more examples per class? To answer these and related questions, we pre-trained CNN features on various subsets of the ImageNet dataset and evaluated transfer performance on PASCAL detection, PASCAL action classification, and SUN scene classification tasks. Our overall findings suggest that most changes in the choice of pre-training data long thought to be critical do not significantly affect transfer performance.? Given the same number of training classes, is it better to have coarse classes or fine-grained classes? Which is better: more classes or more examples per class?

Here we present the first non-invasive methodology for optical delivery and steering deep inside the brain through creating reconfigurable light paths by ultrasonic waves via modulating the refractive and diffractive properties of the medium.