Accelerate Token Production in AI Factories Using Unified Services and Real-Time AI | NVIDIA Technical Blog - NVIDIA Developer
<a href="https://news.google.com/rss/articles/CBMiugFBVV95cUxPM0dKX2tGMjFwTHBER0JaRnFSd0dfMXQzc051ZTZ5eWhiQnBMQUlfVlFDX3U2R3FjWk05MGZJLU1NT1pHUjlPaW1ETGdCTGNjaWFKTTZid3JjeXBvMmJWQ2lYako4Z0R2V1ZqdjVZV1d2ZmZXQ2hCWURaUUFuYkRCWEV6Y1VkN25iQXZTdHRvT05HazdlTFB1LXIwNWlWMUVSQ1BmdXJHV3dpWTNHMnduOF96NGpOVnFHa0E?oc=5" target="_blank">Accelerate Token Production in AI Factories Using Unified Services and Real-Time AI | NVIDIA Technical Blog</a> <font color="#6f6f6f">NVIDIA Developer</font>
Could not retrieve the full article text.
Read on Google News: Machine Learning →Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
productservice![[P] I replaced Dot-Product Attention with distance-based RBF-Attention (so you don't have to...)](https://images.unsplash.com/photo-1551288049-bebda4e38f71?w=600&q=80)
[P] I replaced Dot-Product Attention with distance-based RBF-Attention (so you don't have to...)
<!-- SC_OFF --><div class="md"><p>I recently asked myself what would happen if we replaced the standard dot-product in self-attention with a different distance metric, e.g. an rbf-kernel?</p> <p>Standard dot-product attention has this quirk where a key vector can "bully" the softmax simply by having a massive magnitude. A random key that points in roughly the right direction but is huge will easily outscore a perfectly aligned but shorter key. Distance-based (RBF) attention could fix this. To get a high attention score, Q and K <em>actually</em> have to be close to each other in high-dimensional space. You can't cheat by just being large.</p> <p>I thought this would be a quick 10-minute PyTorch experiment, but it was a reminder on how deeply the dot-product is hardcoded into the entire ML
[D] Self-Promotion Thread
<!-- SC_OFF --><div class="md"><p>Please post your personal projects, startups, product placements, collaboration needs, blogs etc.</p> <p>Please mention the payment and pricing requirements for products and services.</p> <p>Please do not post link shorteners, link aggregator websites , or auto-subscribe links.</p> <p>--</p> <p>Any abuse of trust will lead to bans.</p> <p>Encourage others who create new posts for questions to post here instead!</p> <p>Thread will stay alive until next one so keep posting after the date in the title.</p> <p>--</p> <p>Meta: This is an experiment. If the community doesnt like this, we will cancel it. This is to encourage those in the community to promote their work by not spamming the main threads.</p> </div><!-- SC_ON -->   submitted by   <a href="ht
Covalo raises €3.5M to become the shared data infrastructure for an industry where 80% of products will need reformulating by 2030
The Zurich platform, which connects 1,500+ ingredient suppliers and 6,000 brands including Givaudan, Symrise, PUIG, and La Prairie, is evolving from a discovery marketplace into a data backbone that plugs directly into suppliers’ PIM systems and brands’ R&D workflows. Hi inov led the round. Covalo, the Zurich-based platform connecting personal care ingredient suppliers with brands, […] This story continues at The Next Web
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Products
This Defense Company Made AI Agents That Blow Things Up - WIRED
<a href="https://news.google.com/rss/articles/CBMiiAFBVV95cUxNY2V3RGJkNDduUmVPV3JpYnNkRHgxNGt2dEoyNEdZUTAyOVJMbFl5REZsT08zb1Z4TVlMNWc3OFJDVFNqcWRxa0FGSEFFVTlQajg4dVRIVWRKSzhkZV9yXzN4Z3lWbXltbFk4UDIxcDZQSnJ4alhvZTlINnl6YmRoaTFnT2ZNMUtT?oc=5" target="_blank">This Defense Company Made AI Agents That Blow Things Up</a> <font color="#6f6f6f">WIRED</font>
[D] Why I abandoned YOLO for safety critical plant/fungi identification. Closed-set classification is a silent failure mode
<!-- SC_OFF --><div class="md"><p>I’ve been building an open-sourced handheld device for field identification of edible and toxic plants wild plants, and fungi, running entirely on device. Early on I trained specialist YOLO models on iNaturalist research grade data and hit 94-96% accuracy across my target species. Felt great, until I discovered a problem I don’t see discussed enough on this sub.</p> <p>YOLO’s closed set architecture has no concept of “I don’t know.” Feed it an out of distribution image and it will confidently classify it as one of its classes at near 100% confidence. In most CV cases this can be annoyance. In foraging, it’s potentially lethal.</p> <p>I tried confidence threshold fine-tuning at first, doesn’t work. The confidence scores on OOD inputs are indistinguishable f
[P] I replaced Dot-Product Attention with distance-based RBF-Attention (so you don't have to...)
<!-- SC_OFF --><div class="md"><p>I recently asked myself what would happen if we replaced the standard dot-product in self-attention with a different distance metric, e.g. an rbf-kernel?</p> <p>Standard dot-product attention has this quirk where a key vector can "bully" the softmax simply by having a massive magnitude. A random key that points in roughly the right direction but is huge will easily outscore a perfectly aligned but shorter key. Distance-based (RBF) attention could fix this. To get a high attention score, Q and K <em>actually</em> have to be close to each other in high-dimensional space. You can't cheat by just being large.</p> <p>I thought this would be a quick 10-minute PyTorch experiment, but it was a reminder on how deeply the dot-product is hardcoded into the entire ML
[D] Self-Promotion Thread
<!-- SC_OFF --><div class="md"><p>Please post your personal projects, startups, product placements, collaboration needs, blogs etc.</p> <p>Please mention the payment and pricing requirements for products and services.</p> <p>Please do not post link shorteners, link aggregator websites , or auto-subscribe links.</p> <p>--</p> <p>Any abuse of trust will lead to bans.</p> <p>Encourage others who create new posts for questions to post here instead!</p> <p>Thread will stay alive until next one so keep posting after the date in the title.</p> <p>--</p> <p>Meta: This is an experiment. If the community doesnt like this, we will cancel it. This is to encourage those in the community to promote their work by not spamming the main threads.</p> </div><!-- SC_ON -->   submitted by   <a href="ht
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!