We introduce a simple post-hoc method that improves generative models by fine-tuning on synthetic data and then reversing the degradation direction. Neon achieves state-of-the-art FID of 1.02 on ImageNet-256 with only 0.36% additional compute by exploiting the predictable anti-alignment between synthetic and population (infinite) real data gradients caused by mode-seeking samplers.

This work introduces SIMS, a novel training framework that uses a model’s own generated synthetic data as negative guidance to improve diffusion model performance while preventing model collapse—or “model autophagy disorder” (MAD). The method outperforms prior approaches by setting new Fréchet Inception Distance (FID) records on CIFAR‑10 and ImageNet‑64, and enables controlled bias adjustments in the synthetic data distribution.

We study the phenomenon of training new generative models with synthetic data from previous generative models. Our primary conclusion is that without enough fresh real data in each generation of a self-consuming or autophagous loop, future generative models are doomed to have their quality (precision) or diversity (recall) progressively decrease.

We introduce the Recurrent Neural Tangent Kernel (RNTK), a kernel derived from the infinite-width limit of recurrent neural networks, which captures variable-length sequence inputs and provides improved performance on synthetic and real-world datasets.

We introduce a novel approach that integrates wavelet transforms with state-space models to effectively capture long‑range dependencies in sequences.

We present SaFARi, a frame‑agnostic extension of state‑space models that generalizes HiPPO to support any functional basis, enabling flexible and efficient long-range sequence representation.

We introduce TITAN, which enhances implicit neural representations by integrating deep image priors via a residual deep decoder, significantly improving interpolation quality in super-resolution and CT applications.

We propose a framework that allows tangent features to adapt during training—equivalent to structured regularization—and demonstrate this adaptivity yields significantly lower sample complexity and better kernel alignment than fixed-feature models on MNIST and CIFAR‑10.

We show that commonly used covariate balancing techniques in randomized trials can be vulnerable to adversarial manipulation, undermining their reliability in worst-case scenarios.

We introduce a framework that enhances RNN interpretability by quantifying how each hidden state temporally contributes to model decisions, offering a clear, time-resolved understanding of sequence processing.

We propose NFT‑K, a novel neural architecture that models each layer of a deep network with its own individual tangent kernel—enhancing interpretability and performance over single-kernel approaches.

We introduce a method for compressing variable-length sequences by combining recurrent neural tangent kernels with learned masks, enabling efficient and structured representations of long, complex data streams.

We extend infinite-width recurrent neural tangent kernels to bidirectional and pooled RNN architectures, implement a fast GPU version, and demonstrate superior performance on 90 non‑time‑series UCI datasets.