Deep Learning

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 Fortieth Annual Conference on Neural Information Processing Systems, organized by the Neural Information Processing Systems Foundation, with the main meeting in Sydney, Australia (Dec 6-12) and satellite venues in Atlanta, Georgia and Paris, France (Dec 9-13).

Useful internal representations should explain the patterns of regularities and dependencies among observations. Probabilistic graphical models promise a principled way to uncover latent factors as such, but they are hard to scale to handle high-dimensional sensory observations and complicated dependencies structures. Neural-networks, on the other hand, excel at approximating complicated high-dimensional functions, but their internal representations do not easily lend themselves to a probabilistic interpretation. Despite some successes, a general unified approach is still missing for integrating the two approaches. I will describe a novel approach towards merging adaptive neural-network components into a probabilistic framework, based on three core ideas. The first is to train a set of networks to collectively perform inference, leveraging the ability of pattern-recognition methods to amortise complicated transformations. The second is to constrain the way in which the outputs of these networks are interpreted, transformed, and combined together. These constraints, together with the learning objective itself, are derived directly from probabilistic considerations encoded in a graphical model. Finally, the third core idea is that of recognition-parametrisation, allowing the inference ("recognition") procedure to directly define the model itself, without requiring an explicit "generative" decoder. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2026-03-04. Recording duration: 00:48:26.

Surround suppression is ubiquitous in visual processing and perception. Neurophysiology and modeling studies reveal a rich set of spatial context effects in the Primary Visual Cortex, whereby surrounding stimuli nonlinearly modulate (e.g., suppress) the responses of a center target stimulus. Studies suggest that the strength of surround suppression depends on the image statistics of the center and surround pattern. Namely, if the center and surround share similar statistics, the neural responses are more suppressed by the surround. This behavior has been connected to computational benefits such as improving coding efficiency, and relates to highlighting salient image structure. Though Convolutional Neural Networks (CNNs) have been popular models for biological visual systems, it is unknown if they also have such surround suppression effects. We took advantage of the feature visualization method to visualize the surround pattern that is most suppressive to the center pattern. In detail, we came up with a two-step optimization method: first we found the optimal center pattern, and next the most suppressive surround pattern. Our results match the general understanding of surround suppression effects in the visual cortex. We found that the most suppressive surround is visually similar to the center, while the most facilitative surround is visually distinct from the center. We also found that the most suppressive surround can follow the change in the center, i.e. when presenting a non-optimal center. When the center color was altered, the most suppressive surround was also changed to match the center color. More surprisingly, when the center texture was changed entirely, some neurons in deeper layers could still follow the center texture. Our study not only suggests a new view towards the efficiency of CNNs, but also a new experimental paradigm that could be used in future neurophysiology studies.

1. Can we reconstruct images that a person saw, directly from their fMRI brain recordings? 2. Can we reconstruct the training data that a deep-network trained on, directly from the parameters of the network? The answer to both of these intriguing questions is “Yes!” In this talk I will present some of our work in both domains. I will then show how combining the power of Brains and Machines can lead to significant breakthroughs in both areas, and potentially bridge the gap between Minds and Machines. Finally, I will show how combining the power of Multiple Brains (with NO shared data) may lead to new breakthrough discoveries in Brain-Science, and allow mapping of information between different brains. Presented in the van Vreeswijk Theoretical Neuroscience Seminar series (formerly WWTNS) on 2026-02-18. Recording duration: 00:50:24.

The Brain Image Analysis Unit at the RIKEN Center for Brain Science seeks a postdoctoral researcher to develop deep-learning methods for large-scale microscopy and MRI of the marmoset brain within Brain/MINDS 2.0. The programme combines two-photon and bright-field microscopy, spatial transcriptomics, anterograde tracing and diffusion MRI into a whole-brain multimodal atlas. Research topics include neuron and axon tracing in dense 3D data, generative models for synthetic tracing ground truth, images exceeding GPU memory, cross-modal image registration, biomedical foundation models, and diffusion-MRI tractography validated against same-animal tracer data. Biomedical image-analysis experience is beneficial but is not mandatory. The role is based at Wako.

Supports learned and generative processing of limited sensor observations in compact, low-power devices. Research topics include generating visual experiences beyond capture constraints, efficient visual encoding for on-device vision-language models, and learned multimodal sensor fusion for object tracking. Proposals should address realistic power, bandwidth, noise and capture constraints and explain how research advances perception. Use the ARA proposal template; four pages excluding appendices are encouraged. Decisions are expected in February 2027. Applications close on 4 November 2026 at 11:59pm Pacific Time.

Supports multilingual and culturally contextual evaluation of AI systems, including benchmarks and datasets for underrepresented languages and contexts, with attention to the Global South. Proposals may develop responsible evaluation infrastructure, human judgment in evaluation, and methods for assessing AI deployment in emerging markets. Use the ARA proposal template; four pages excluding appendices are encouraged. Decisions are expected in February 2027. Applications close on 4 November 2026 at 11:59pm Pacific Time.