Michele Marzollo, Jiawei Zhuang, Niklas Roemer, Niklas Zwingenberger, Lorenz K Muller, Lukas Cavigelli. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Neural networks whose weights are the output of a predictor (HyperNetworks) achieve excellent performance on many tasks. In ConvNets, kernel prediction layers are a popular type of HyperNetwork. Previous theoretical work has argued that a hierarchy of multiplicative interactions exists in which gating is at the bottom and full weight prediction, as in HyperNetworks, is at the top. In this paper, we constructively demonstrate an equivalence between gating combined with fixed weight layers and weight prediction, relativizing the notion of a hierarchy of multiplicative interactions. We further derive an equivalence between a restricted type of HyperNetwork and factorization machines. Finally, we find empirically that gating layers can learn to imitate weight prediction layers with an SGD variant and show a novel practical application in image denoising using kernel prediction networks. Our reformulation of predicted kernels, combining fixed layers and gating, reduces memory requirements.

Monte-Carlo path tracing is a powerful technique for realistic image synthesis but suffers from high levels of noise at low sample counts, limiting its use in real-time applications. To address this, we propose a framework with end-to-end training of a sampling importance network, a latent space encoder network, and a denoiser network. Our approach uses reinforcement learning to optimize the sampling importance network, thus avoiding explicit numerically approximated gradients. Our method does not aggregate the sampled values per pixel by averaging but keeps all sampled values which are then fed into the latent space encoder. The encoder replaces handcrafted spatiotemporal heuristics by learned representations in a latent space. Finally, a neural denoiser is trained to refine the output image. Our approach increases visual quality on several challenging datasets and reduces rendering times for equal quality by a factor of 1.6x compared to the previous state-of-the-art, making it a promising solution for real-time applications.

Camera sensors rely on global or rolling shutter functions to expose an image. This fixed function approach severely limits the sensors' ability to capture high-dynamic-range (HDR) scenes and resolve high-speed dynamics. Spatially varying pixel exposures have been introduced as a powerful computational photography approach to optically encode irradiance on a sensor and computationally recover additional information of a scene, but existing approaches rely on heuristic coding schemes and bulky spatial light modulators to optically implement these exposure functions. Here, we introduce neural sensors as a methodology to optimize per-pixel shutter functions jointly with a differentiable image processing method, such as a neural network, in an end-to-end fashion. Moreover, we demonstrate how to leverage emerging programmable and re-configurable sensor-processors to implement the optimized exposure functions directly on the sensor. Our system takes specific limitations of the sensor into account to optimize physically feasible optical codes and we evaluate its performance for snapshot HDR and high-speed compressive imaging both in simulation and experimentally with real scenes.

In this paper we introduce a novel neural network architecture, in which weight matrices are re-parametrized in terms of low-dimensional vectors, interacting through kernel functions. A layer of our network can be interpreted as introducing a (potentially infinitely wide) linear layer between input and output. We describe the theory underpinning this model and validate it with concrete examples, exploring how it can be used to impose structure on neural networks in diverse applications ranging from data visualization to recommender systems. We achieve state-of-the-art performance in a collaborative filtering task (MovieLens).

Memristive devices have emerged as compact nonvolatile memory elements which can be used as synapses in neuromorphic architectures. However, the intrinsic stochasticity in their switching behavior, non-linear characteristics, and variability limit their operation in real systems. In this paper we propose spike-based learning circuits designed to exploit the stochastic properties of memristors. This implements a probabilistic version of a local gradient descent rule, namely the delta rule, for online learning in neuromorphic chips. The circuits proposed translate the delta error to the slope of a ramp voltage which modulates the probability of resistive switching in very low resolution (i.e. binary) memristive devices. We demonstrate the feasibility and computational power of such approach, using a spiking neural network simulator to carry out system level behavioral simulations of the neuromorphic architecture applied to a classification task of digits 0 to 4 in the MNIST data-set.

In this paper, we present a 3D imaging system providing a semi-dense depth map, using a passive, low-power, compact, static, monocular camera. The demonstrated depth estimation system reconstructs 32 depth-levels in real-time at 25FPS drawing less than 1.9W of power. This is achieved by performing computation on an analog focal-plane processor that analyses frames captured through a vibrating liquid focus-tunable lens. The optical system provides shallow depth of focus images and fast sweeps of optical power, while the use of pixel-level processing removes the sensor-processor bandwidth limitations of depth-from-focus systems built using conventional imaging and processor technologies. All-in-focus images are also obtained.

We demonstrate a 3D imaging system that produces sparse depth maps. It consists in a liquid focus-tunable lens whose focal power can be changed at high speed, placed in front of a SCAMP5 vision-chip embedding processing capabilities in each pixel. The focus-tunable lens performs focal sweeps with shallow depth of fields. These are sampled by the vision chip taking multiple images at different focus and analyzed on-chip to produce a single depth frame. The combination of the focus tunable-lens with the vision-chip, enabling near-focal plane processing, allows us to present a compact passive system that is static, monocular, real-time (> 25FPS) and low-power (<; 1.6W).

Training neural networks with low-resolution synaptic weights raised much interest recently and inference in neural networks with binary activation and binary weights has been shown to be able to achieve near state-of-the-art performance in a wide range of tasks. However, the current methods for training such networks rely on high-resolution gradients or update probabilities. Low resolution training methods would be useful for neuromorphic architectures that support lower power hardware implementations as well as emerging memory technologies based on memristive devices that do not always support fine-grained state changes. In this paper, we propose a training method, Randomized Unregulated Step Descent (RUSD), as an alternative to gradient descent that uses only a single bit of information about the gradient; we show how it is compatible with low-resolution integer arithmetic platforms and is resilient to some of the prominent non-idealities of memristive memories. We verify the performance of RUSD several standard machine-learning benchmarks.

To improve computational efficiency, it may be advantageous to transfer part of the intelligence lying in the core of a system to its sensors. Vision sensors equipped with small programmable processors at each pixel allow us to follow this principle in so-called near-focal plane processing, which is performed on-chip directly where light is being collected. Such devices need then only to communicate relevant pre-processed visual information to other parts of the system. In this work, we demonstrate how two classical problems, namely high dynamic range imaging and auto-focus, can be solved efficiently using two simple parallel algorithms implemented on such a chip. We illustrate with these two examples that embedding uncomplicated algorithms on-chip, directly where information acquisition takes place can replace more complex dedicated post-processing. Adapting data acquisition by bringing processing at the sensor level allows us to explore solutions that would not be feasible in a conventional sensor-ADC-processor pipeline.

Gamma-band rhythmic inhibition is a ubiquitous phenomenon in neural circuits, yet its computational role remains elusive. We show that a model of gamma-band rhythmic inhibition allows networks of coupled cortical circuit motifs to search for network configurations that best reconcile external inputs with an internal consistency model encoded in the network connectivity. We show that Hebbian plasticity allows the networks to learn the consistency model by example. The search dynamics driven by rhythmic inhibition enable the described networks to solve difficult constraint satisfaction problems without making assumptions about the form of stochastic fluctuations in the network. We show that the search dynamics are well approximated by a stochastic sampling process. We use the described networks to reproduce perceptual multistability phenomena with switching times that are a good match to experimental data and show that they provide a general neural framework that can be used to model other perceptual inference phenomena.

We present a recurrent neuronal network, modeled as a continuous-time dynamical system, that can solve constraint satisfaction problems. Discrete variables are represented by coupled Winner-Take-All (WTA) networks, and their values are encoded in localized patterns of oscillations that are learned by the recurrent weights in these networks. Constraints over the variables are encoded in the network connectivity. Although there are no sources of noise, the network can escape from local optima in its search for solutions that satisfy all constraints by modifying the effective network connectivity through oscillations. If there is no solution that satisfies all constraints, the network state changes in a pseudo-random manner and its trajectory approximates a sampling procedure that selects a variable assignment with a probability that increases with the fraction of constraints satisfied by this assignment. External evidence, or input to the network, can force variables to specific values. When new inputs are applied, the network re-evaluates the entire set of variables in its search for the states that satisfy the maximum number of constraints, while being consistent with the external input. Our results demonstrate that the proposed network architecture can perform a deterministic search for the optimal solution to problems with non-convex cost functions. The network is inspired by canonical microcircuit models of the cortex and suggests possible dynamical mechanisms to solve constraint satisfaction problems that can be present in biological networks, or implemented in neuromorphic electronic circuits.