HCNAF: Hyper-Conditioned Neural Autoregressive Flow and its Application for Probabilistic Occupancy Map Forecasting

Abstract

We introduce Hyper-Conditioned Neural Autoregressive Flow (HCNAF); a powerful universal distribution approximator designed to model arbitrarily complex conditional probability density functions. HCNAF consists of a neural-net based conditional autoregressive flow (AF) and a hyper-network that can take large conditions in nonautoregressive fashion and outputs the network parameters of the AF. Like other flow models, HCNAF performs exact likelihood inference. We demonstrate the effectiveness and attributes of HCNAF, including its generalization capability over unseen conditions and show that HCNAF outperforms recent flow models in a conditional density estimation task for MNIST. We also show that HCNAF scales up to complex high-dimensional prediction problems of the magnitude of self-driving and that HCNAF yields a state-of-the-art performance in a public self-driving dataset.

CVAE-H: Conditionalizing Variational Autoencoders via Hypernetworks and Trajectory Forecasting for Autonomous Driving

“…Furthermore, a group of generative models that are used to model a conditional distribution p(x o |x i ) are called conditional generative models. Examples include conditional VAE [18], conditional GAN [19], and conditional flow [5].…”

“…Normalizing flow, or Flow, is also a type of explicit generative model. While flow-based models [4,5,22] have access to the exact probability density p(x), flow-based models are typically better suited for density estimations than generative tasks (e.g., trajectory forecasting).…”

“…It is not trivial for a prediction model to satisfy all four attributes. First of all, not all neural-network models are probabilistic unless they build on explicit density models such as Mixture Density Network (e.g., Gaussian Mixture Models) [2,3], Normalizing-flow models [4,5], Variational Autoencoder (VAE) [6]. Secondly, not all probabilistic models are multi-modal, for instance, a uni-modal Gaussian.…”

Lightweight and Interpretable Neural Modeling of an Audio Distortion Effect Using Hyperconditioned Differentiable Biquads

“…Conventionally, conditioning information is injected into networks via addition [20], concatenation [21], or adaptive instance normalization [22]. Conversely, HC learns transformations of conditioning signals which ultimately control the underlying model weights (for instance, the coefficients of a convolutional kernel) [12]. This type of adaptation is more direct and expressive, validating its use in recent audio style transfer applications [13].…”

“…We continue the trend towards lightweight, interpretable audio deep learning models, using several IIR filters as building blocks. Unlike previous attempts to learn trainable IIR filters using deep learning [9,11], we use hyperconditioning (HC) [12,13] in order to capture changes in filtering stages based on the user-facing parameters controlling the effect. We consider many cascaded biquad (CB) representations and introduce activation functions necessary for their stable training.…”