Proud to announce our new paper, "Relevant Entity Selection: Knowledge Graph Bootstrapping via Zero-Shot Analogical Pruning" with Lucas Jarnac and Miguel Couceiro, accepted in #cikm2023

#knowledgeGraph #machineLearning #graphEmbedding #transferLearning #zeroShot #openScience

#ESC is a #ZeroShot agent that uses #ChatGPT and #SoftLogic to navigate in 3D environments with #CommonSense and #WorldKnowledge. #AI #NLP #RL
https://www.marktechpost.com/2023/07/22/researchers-from-uc-santa-cruz-and-samsung-introduce-esc-a-zero-shot-object-navigation-agent-that-leverages-commonsense-in-llms-like-chatgpt-for-navigation-decisions/

@Riedl

I have a #LLM that has 100% accuracy and #ZeroShot transfer on all blocks world problems. It is 'better than humans' who only have 80% accuracy. I call it the PDDLModel. With some prompt engineering (also called model writing), it can #zeroshot transfer to other problems like tower-of-hanoi, 8-puzzle, eaters, and what not. It even fits on a raspberry pi.

We can only imagine what this new #AGI would enable.

https://github.com/gerryai/PDDL4J/blob/master/pddl/blockworld/blocksworld.pddl

Let's all play this game!!

#Stanford Introduces #DetectGPT to Help #Educators Fight Back Against #ChatGPT Generated Papers

"This method, called ' #ZeroShot,' allows DetectGPT to detect machine-written text without any knowledge of the #AI that was used to generate it. It operates in stark contrast to other methods which require training #classifiers and datasets of real and generated passages."

#ArtificialIntelligence #Education #Teaching #OpenAI #TechNews #TechnologyNews #BreakingNews

https://www.neowin.net/news/stanford-introduces-detectgpt-to-help-educators-fight-back-against-chatgpt-generated-papers/

L'uso di modelli linguistici di grandi dimensioni (LLM) è alle stelle e con buone ragioni; è molto buono. Nelle ultime due settimane, ChatGPT è diventato il mio strumento preferito. Al lavoro, gli ho chiesto come costruire un oscuro pezzo di software Linux su un kernel moderno, e mi ha detto come. H...
https://parliamodi.news/detail/9339.html
#28Gennaio #zeroshot #training

@dx One of the papers in the #zeroshot space is "Language Models are Zero-Shot Reasoners" (https://arxiv.org/abs/2205.11916)

Followup work ("Impact of Pretraining Term Frequencies on Few-Shot Reasoning", https://arxiv.org/abs/2202.07206) found that the model only does well on arithmetic test items that appear at least 10^7 times in the training set (for the test item "what is 2 times 24?", the problem and answer appears 10^7 times in training set).

Hard to think of this as “zero” or “few” shot learning.

@cmaclell I think that's an important point. I like framing it as #transfer learning, because you have a network that has settled pretty solidly into a particular basin of the parameter space, based on the data it was trained with. This is VERY different than a system that has abstracted generalizable, relevant concepts, which is what I think people want to be true in #zeroshot learning.