Arno Solin

🔥 #LeAVJEPA: A simple recipe for self-supervised learning from sound and vision 🔥 One shared encoder, trained without class labels. Dropping a modality makes audio & video learn a common representation. 🧵

5/ In the video, audio-to-video attention follows the sound-producing objects, although the model received no localisation supervision. With attentive probes on frozen ViT-L features, we obtain 36.0 mAP on AudioSet-20K and 91.3% accuracy on ESC-50.

🔥 #LeAVJEPA: A simple recipe for self-supervised learning from sound and vision 🔥 One shared encoder, trained without class labels. Dropping a modality makes audio & video learn a common representation. 🧵

3/ The important detail is how we construct the views. When every view contains both modalities, matching them can be solved by relying on just one. Modality dropout makes that shortcut harder, and our controlled ablation shows a clear improvement in the learned features.

🔥 #LeAVJEPA: A simple recipe for self-supervised learning from sound and vision 🔥 One shared encoder, trained without class labels. Dropping a modality makes audio & video learn a common representation. 🧵

3D Gaussian splats are amazing--until you realise how much storage space they need. Smol-GS makes #3DGS actually small: explicit splats, tiny learned features w/ positional awareness, octree-coded geometry, and fast rendering. 🥇Now leading the 3DGS.zip benchmark.

7/ For details, check the paper pre-print (rendering speeds, training details, etc.): arxiv.org/abs/2512.00850

3D Gaussian splats are amazing--until you realise how much storage space they need. Smol-GS makes #3DGS actually small: explicit splats, tiny learned features w/ positional awareness, octree-coded geometry, and fast rendering. 🥇Now leading the 3DGS.zip benchmark.

5/ Tiny MLPs decode each splat into what the renderer needs: color, opacity, shape, rotation, and view-dependent effects. The features and scale controllers are quantized and entropy-coded. The result is a compact neural splat representation without anchor-offset bookkeeping.

3D Gaussian splats are amazing--until you realise how much storage space they need. Smol-GS makes #3DGS actually small: explicit splats, tiny learned features w/ positional awareness, octree-coded geometry, and fast rendering. 🥇Now leading the 3DGS.zip benchmark.

3/ Most strong 3DGS compression methods are anchor-based: Store shared anchors, use local offsets, and decode splats around them. That works well and has dominated recent benchmarks. But Smol-GS takes a different approach.

3D Gaussian splats are amazing--until you realise how much storage space they need. Smol-GS makes #3DGS actually small: explicit splats, tiny learned features w/ positional awareness, octree-coded geometry, and fast rendering. 🥇Now leading the 3DGS.zip benchmark.

1/ 🔥 New paper: Differentiable Vector Quantization (DiVeQ) 🔥 Vector quantization (VQ) is a core tool in modern AI. It connects continuous data like images and audio to discrete tokens used by transformers. It underpins compression, generation, and multimodal modelling.

7/ DiVeQ is also included in the popular vector-quantize-pytorch package. To use it there, enable: directional_reparam=True

1/ 🔥 New paper: Differentiable Vector Quantization (DiVeQ) 🔥 Vector quantization (VQ) is a core tool in modern AI. It connects continuous data like images and audio to discrete tokens used by transformers. It underpins compression, generation, and multimodal modelling.

5/ The result is a direct and general way to do end-to-end trainable quantization, without many of the complications of earlier approaches. We also see improved performance in image compression, image generation, and speech coding.

1/ 🔥 New paper: Differentiable Vector Quantization (DiVeQ) 🔥 Vector quantization (VQ) is a core tool in modern AI. It connects continuous data like images and audio to discrete tokens used by transformers. It underpins compression, generation, and multimodal modelling.

3/ In our #ICLR2026 paper, we introduce DiVeQ. The idea is simple: keep the hard quantization behavior we want, but make training behave as if learning can still pass through it.

1/ 🔥 New paper: Differentiable Vector Quantization (DiVeQ) 🔥 Vector quantization (VQ) is a core tool in modern AI. It connects continuous data like images and audio to discrete tokens used by transformers. It underpins compression, generation, and multimodal modelling.

Statement from #AISTATS2026 organizers regarding the @openreview.bsky.social API Security Incident

I recently gave my installation talk after being tenured. The video of the talk is now available on the university's YouTube channel: youtu.be/R1UQoflPTDg 1/n

My own research, together with my group, focuses less on building the giant models and more on designing the building blocks behind them: model components, inductive biases, training principles, and inference methods that make AI systems more robust, data-efficient, and uncertainty-aware. 3/n

I recently gave my installation talk after being tenured. The video of the talk is now available on the university's YouTube channel: youtu.be/R1UQoflPTDg 1/n

📣 Please share: We invite submissions to the 29th International Conference on Artificial Intelligence and Statistics (#AISTATS 2026) and welcome paper submissions at the intersection of AI, machine learning, statistics, and related areas. [1/3]

Accepted papers will be presented in person in Morocco, May 2–5, 2026. The full Call for Papers is available here: virtual.aistats.org/Conferences/... [2/3]

Real-world #3DGS scenes are messy—occluders, moving objects, and clutter often ruin reconstruction. This #CVPR2025 paper presents DeSplat, which separates static scene content from distractors, all without requiring external semantic models. [1/n]

Qualitative visualization of static distractor elements achieved by our model, DeSplat. [3/n]

Real-world #3DGS scenes are messy—occluders, moving objects, and clutter often ruin reconstruction. This #CVPR2025 paper presents DeSplat, which separates static scene content from distractors, all without requiring external semantic models. [1/n]

Have you thought that in computer memory model weights are given in terms of discrete values in any case. Thus, why not do probabilistic inference on the discrete (quantized) parameters. @trappmartin.bsky.social is presenting our work at #AABI2025 today. [1/3]

We introduce BitVI, a novel approach for variational inference with discrete bitstring representations of continuous parameters. We use a deterministic probabilistic circuit structure to model the distribution over bitstrings, allowing for exact and efficient probabilistic inference. [2/3]

Excited to share "Plan*RAG: Efficient Test-Time Planning for Retrieval Augmented Generation", presented at the #ICLR2025 "Workshop on Reasoning and Planning for LLMs" on Monday! 🚀 1/3

This work was born out of Prakhar's internship with Microsoft Research (\w Sukruta Prakash Midigeshi, Gaurav Sinha, Arno Solin, Nagarajan Natarajan, and Amit Sharma). 2/3