Neural information processing

The Fortieth Annual Conference on Neural Information Processing Systems brings together interdisciplinary machine-learning research through peer-reviewed sessions, invited talks, demonstrations, tutorials, workshops, and an exposition.

The brain is one of the most energy intense organs. Some of this energyis used for neural information processing, however, fruitfly experiments have shown that also learning is metabolically costly. We will present estimates of this cost and introduce a general model of this cost, and compare it to costs in computers. Next, we turn to a supervised artificial network setting and explore a number of strategies that cansave energy need for plasticity. Either by modifying the objective function, by restricting plasticity, or by using less costly transient forms of plasticity. Finally, we will disc

Mounting evidence suggests that traveling waves are a key dynamical motif in biological neural information processing systems [Muller et al., 2018], yet their precise computational role remains under investigation. While the Transformer architecture [Vaswani et al., 2023] has achieved state-of-the-art performance across various data modalities, it relies heavily on an explicit `context’ – processing the entire input sequence simultaneously. This approach is both biologically implausible and computationally inefficient, resulting in quadratic complexity relative to sequence length. Recent inter

Temporal sequences are an important feature of neural information processing in biology. Neurons can fire a spike with millisecond precision, and, at the network level, repetitions of spatiotemporal spike patterns are observed in neurobiological data. However, methods for detecting precise temporal patterns in neural activity suffer from high computational complexity and poor robustness to noise, and quantitative detection of these repetitive patterns remains an open problem. Here, we propose a new method to extract spike patterns embedded in raster plots using a 1D convolutional autoencoder w

A well-established hypothesis in neuroscience proposes that the fundamental unit of brain computation consists of synchronized groups of neurons, known as cell assemblies. Our team demonstrated the existence of reader neurons, responsive to the activity of cell assemblies from upstream brain structures, and showed that functional links between assemblies and their readers are modified through learning processes. On the other hand, there is accumulating evidence that brain-wide dynamics are characterized by neuronal avalanches, namely aperiodic bursts of activity spreading across the population

How does the organization of neural information processing enable humans’ sophisticated cognition? Here we decompose functional interactions between brain regions into synergistic and redundant components, revealing their distinct information-processing roles. Combining functional, structural, and diffusion-weighted MRI with meta-analytic results from NeuroSynth, we demonstrate that redundant interactions are predominantly associated with structurally coupled, modular sensorimotor processing. Synergistic interactions instead support integrative processes and complex cognition across higher-or

Understanding neural information processing is extremely challenging. This is because often, it is not even clear what information neurons are really processing. However, we can access the dynamic fingerprint of the processing with information-theoretic and statistical methods. Recent work has thus focused on the autocorrelation time of neural activity, also termed intrinsic timescale, which serves as a proxy for how long information is stored in the activity of the network. This work revealed a hierarchy of intrinsic timescales, which suggests that along the processing hierarchy, the brain