[Paper Review] DISCO Nets: DISsimilarity COefficient Networks

Diane Bouchacourt, Manish Kumar+1 more|arXiv (Cornell University)|Jun 8, 2016

Generative Adversarial Networks and Image Synthesis21 references3 citations

TL;DR

DISCO Nets introduce a novel probabilistic deep learning framework that models uncertainty in structured prediction by minimizing a dissimilarity coefficient between true and estimated posterior distributions. By tailoring the training objective to the task-specific loss, DISCO Nets outperform non-probabilistic and existing probabilistic models in hand pose estimation, accurately capturing output uncertainty and achieving state-of-the-art performance with minimal architectural constraints.

ABSTRACT

We present a new type of probabilistic model which we call DISsimilarity COefficient Networks (DISCO Nets). DISCO Nets allow us to efficiently sample from a posterior distribution parametrised by a neural network. During training, DISCO Nets are learned by minimising the dissimilarity coefficient between the true distribution and the estimated distribution. This allows us to tailor the training to the loss related to the task at hand. We empirically show that (i) by modeling uncertainty on the output value, DISCO Nets outperform equivalent non-probabilistic predictive networks and (ii) DISCO Nets accurately model the uncertainty of the output, outperforming existing probabilistic models based on deep neural networks.

Motivation & Objective

• To address the challenge of modeling uncertainty in structured prediction tasks where ground-truth outputs are ambiguous due to data noise or occlusions.
• To develop a probabilistic deep learning framework that allows efficient posterior sampling and uncertainty quantification.
• To enable training objectives that align with task-specific evaluation losses, improving generalization and predictive performance.
• To overcome limitations of existing models like GANs and VAEs, which require complex architectures and training procedures.
• To provide a flexible, architecture-agnostic framework that can be easily adapted to diverse prediction tasks with uncertainty.

Proposed method

• DISCO Nets represent the true posterior distribution P using a neural network-parametrized distribution Q.
• The model is trained by minimizing a dissimilarity coefficient between P and Q, derived from a symmetric, non-negative loss function.
• This dissimilarity coefficient, based on Rao's work, allows the training objective to be explicitly tailored to the task-specific loss function Δ_task.
• The method does not require adversarial training or specific network architectures, enabling broad applicability.
• Sampling from the posterior is efficient, as Q is directly parameterized by the neural network.
• The framework supports end-to-end training using backpropagation, with no need for inference networks or approximate posterior approximations.

Experimental results

Research questions

• RQ1 Can a deep learning model that explicitly minimizes a task-specific dissimilarity coefficient outperform non-probabilistic models in structured prediction tasks?
• RQ2 Does tailoring the training objective to the evaluation loss lead to improved performance on uncertain or ambiguous prediction tasks?
• RQ3 Can DISCO Nets accurately model output uncertainty in real-world scenarios such as hand pose estimation from depth images?
• RQ4 How does DISCO Nets compare to state-of-the-art probabilistic models like cGANs and VAEs in terms of predictive accuracy and training stability?
• RQ5 To what extent can DISCO Nets be generalized across different architectures and tasks without requiring architectural modifications?

Key findings

• DISCO Nets outperform equivalent non-probabilistic predictive networks on hand pose estimation, achieving a MeJEE of 20.7 ± 0.121 mm and MaJEE of 45.1 ± 0.246 mm.
• The model achieves state-of-the-art performance comparable to NYU-Prior-Refined (MeJEE: 19.7 ± 0.157 mm) and NYU-Feedback (MeJEE: 16.0 ± 0.096 mm), despite using a simpler single-network architecture.
• When trained with β=1, γ=0.5, DISCO Nets achieve a 93.250% fraction of frames within 80 mm error, outperforming NYU-Init (86.537%) and NYU-Feedback (97.334%) in the 80 mm threshold.
• cGAN-based models only achieve competitive results when the generator and discriminator convolutional layers are initialized with a pre-trained DISCO Net and kept fixed, indicating training instability and architectural dependency.
• DISCO Nets significantly outperform standard GANs and cGANs in both accuracy and training reliability, demonstrating the advantage of a principled, loss-tailored objective function.
• The framework enables accurate uncertainty modeling without requiring complex inference networks or adversarial training, making it more efficient and easier to deploy than existing probabilistic models.