Kaiming He
Kaiming He is a prominent researcher in the field of computer vision and deep learning, known for his contributions to neural network architectures.
Neural networks
Model merging (neural networks)
Model merging in neural networks is a technique that combines multiple trained neural network models into a single model. This approach aims to integrate knowledge from different models to improve performance, efficiency, or adaptability.
Logit bias
Logit bias refers to the intentional adjustment of the logit values in machine learning models to influence prediction probabilities. It is commonly used in natural language processing and classification tasks to control or steer model outputs.
Concept activation vectors (CAV)
Concept activation vectors (CAV) are a method used in machine learning to interpret neural networks by associating specific directions in the latent space with human-understandable concepts.
VoiceBox (non-autoregressive TTS)
VoiceBox is a non-autoregressive text-to-speech (TTS) system designed to generate natural-sounding speech efficiently by predicting audio features in parallel rather than sequentially. It leverages advanced neural network architectures to improve synthesis speed while maintaining high audio quality.
Hyper-deep ensembles
Hyper-deep ensembles are advanced machine learning models that combine multiple deep neural networks to improve predictive performance, robustness, and uncertainty estimation. They extend traditional ensemble methods by leveraging very large or highly complex models in a coordinated manner.
Neural volume rendering
Neural volume rendering is a computational technique that leverages neural networks to synthesize images by modeling volumetric scenes. It combines principles from volume rendering and deep learning to generate photorealistic or novel views from sparse input data.
Differentiable predictive coding (DPC)
Differentiable predictive coding (DPC) is a computational framework that integrates predictive coding principles with differentiable neural network architectures. It enables end-to-end learning of hierarchical prediction models by minimizing prediction errors through gradient-based optimization.
DiffWave (diffusion waveform model)
DiffWave is a generative model based on diffusion processes for high-quality waveform synthesis, primarily used in speech generation. It leverages a denoising diffusion probabilistic model to produce natural audio waveforms from noise, offering an alternative to traditional autoregressive and adversarial approaches.
Griffin (gated linear recurrent unit)
Griffin is a variant of the gated linear recurrent unit (GLRU), a type of recurrent neural network architecture designed to improve sequence modeling by combining gating mechanisms and linear recurrent dynamics. It aims to enhance efficiency and gradient flow in tasks involving sequential data.