]> 2022-12-10T11:51:09Z 2022-12-10 Illustrating Reinforcement Learning from Human Feedback (RLHF) 2022-12-13T11:46:14Z 2022-12-13 ValueError "invalid literal for int() with base 10" in trainer.evaluate (dataset created from pandas) · Issue #228 · huggingface/setfit see <https://github.com/huggingface/setfit/blob/main/notebooks/zero-shot-classification.ipynb> > Note: some datasets on the Hugging Face Hub don't have a ClassLabel feature for the label column. In these cases, you should compute the candidate labels manually by first computing the id2label mapping as follows: 2022-12-23 IBM/zshot: Zero and Few shot named entity & relationships recognition 2022-12-23T01:00:31Z 2022-12-03 Theoretical tools of statistical physics to study computational questions in high-dimensional problems > We will highlight some key results in this field with examples of applications in artificial neural networks and in signal processing. The Physics of Algorithms, by Prof. Lenka Zdeborová - YouTube 2022-12-03T18:09:56Z Le roi fantôme - Maaza Mengiste 2022-12-29 2022-12-29T17:46:23Z 2022-12-30 > "La guerre c'est la paix, la liberté c'est l'esclavage, l'ignorance c'est la force" > Et la crise climatique, c'est la faute aux écolos Chronique | « “C’est la faute aux écolos !”, l’élément de langage phare de l’année 2022 » 2022-12-30T20:46:39Z Field Manual for African Archaeology (2017) 2022-12-04T13:28:28Z 2022-12-04 Alexandre Livingstone Smith, Els Cornelissen, Olivier P. Gosselain, Scott MacEachern Yang Liu We propose Universal Document Processing (UDOP), a foundation Document AI model which unifies text, image, and layout modalities together with varied task formats, including document understanding and generation. UDOP leverages the spatial correlation between textual content and document image to model image, text, and layout modalities with one uniform representation. With a novel Vision-Text-Layout Transformer, UDOP unifies pretraining and multi-domain downstream tasks into a prompt-based sequence generation scheme. UDOP is pretrained on both large-scale unlabeled document corpora using innovative self-supervised objectives and diverse labeled data. UDOP also learns to generate document images from text and layout modalities via masked image reconstruction. To the best of our knowledge, this is the first time in the field of document AI that one model simultaneously achieves high-quality neural document editing and content customization. Our method sets the state-of-the-art on 9 Document AI tasks, e.g., document understanding and QA, across diverse data domains like finance reports, academic papers, and websites. UDOP ranks first on the leaderboard of the Document Understanding Benchmark (DUE). Guoxin Wang Zineng Tang 2022-12-05T22:14:49Z 2022-12-07T16:52:28Z Chenguang Zhu Michael Zeng 2022-12-07 Unifying Vision, Text, and Layout for Universal Document Processing Cha Zhang 2022-12-05T22:14:49Z Zineng Tang Ziyi Yang Yuwei Fang Mohit Bansal [2212.02623] Unifying Vision, Text, and Layout for Universal Document Processing 2212.02623 Le Groenland d’il y a deux millions d’années dévoilé par l’ADN environnemental 2022-12-07 2022-12-07T21:40:26Z Docker Cheatsheet 2022-12-19 2022-12-19T16:18:41Z 2022-12-20T00:03:04Z 2022-12-20 Matthew Honnibal sur Twitter : "We've been working on new prodi.gy workflows that let you use the @OpenAI API to kickstart your annotations, via zero- or few-shot learning. ..." Shubham Saboo sur Twitter : "Presenting Topically by Cohere AI" 2022-12-10T11:32:19Z 2022-12-10 > Unlock the potential of your text data with Large Language models. Analyze millions of texts (messages, emails, news headlines) in a matter of seconds..." Moving Beyond Downstream Task Accuracy for Information Retrieval Benchmarking Keshav Santhanam Jon Saad-Falcon Omar Khattab Radu Florian Avirup Sil 2022-12-02T17:57:06Z 2022-12-02T17:57:06Z Md Arafat Sultan Salim Roukos Matei Zaharia 2022-12-06T19:27:25Z Martin Franz Christopher Potts [2212.01340] Moving Beyond Downstream Task Accuracy for Information Retrieval Benchmarking Neural information retrieval (IR) systems have progressed rapidly in recent years, in large part due to the release of publicly available benchmarking tasks. Unfortunately, some dimensions of this progress are illusory: the majority of the popular IR benchmarks today focus exclusively on downstream task accuracy and thus conceal the costs incurred by systems that trade away efficiency for quality. Latency, hardware cost, and other efficiency considerations are paramount to the deployment of IR systems in user-facing settings. We propose that IR benchmarks structure their evaluation methodology to include not only metrics of accuracy, but also efficiency considerations such as a query latency and the corresponding cost budget for a reproducible hardware setting. For the popular IR benchmarks MS MARCO and XOR-TyDi, we show how the best choice of IR system varies according to how these efficiency considerations are chosen and weighed. We hope that future benchmarks will adopt these guidelines toward more holistic IR evaluation. Keshav Santhanam 2022-12-06 2212.01340 2022-12-01T23:04:37Z 2022-12-01 En Espagne, des canaux d’irrigation médiévaux remis en état pour lutter contre la sécheresse 2022-12-15T12:34:51Z 2205.05638 Jay Mohta [2205.05638] Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning 2022-12-15 Haokun Liu Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning Derek Tam Mohit Bansal Colin Raffel Mohammed Muqeeth 2022-05-11T17:10:41Z Tenghao Huang Few-shot in-context learning (ICL) enables pre-trained language models to perform a previously-unseen task without any gradient-based training by feeding a small number of training examples as part of the input. ICL incurs substantial computational, memory, and storage costs because it involves processing all of the training examples every time a prediction is made. Parameter-efficient fine-tuning (PEFT) (e.g. adapter modules, prompt tuning, sparse update methods, etc.) offers an alternative paradigm where a small set of parameters are trained to enable a model to perform the new task. In this paper, we rigorously compare few-shot ICL and PEFT and demonstrate that the latter offers better accuracy as well as dramatically lower computational costs. Along the way, we introduce a new PEFT method called (IA)$^3$ that scales activations by learned vectors, attaining stronger performance while only introducing a relatively tiny amount of new parameters. We also propose a simple recipe based on the T0 model called T-Few that can be applied to new tasks without task-specific tuning or modifications. We validate the effectiveness of T-Few on completely unseen tasks by applying it to the RAFT benchmark, attaining super-human performance for the first time and outperforming the state-of-the-art by 6% absolute. All of the code used in our experiments is publicly available. 2022-08-26T16:23:29Z Haokun Liu 2022-12-09 2022-12-09T11:30:35Z shikhar sur Twitter : "Instead of asking whether tree structure should be baked into NNs, our new paper asks if transformers already have a tendency to learn tree structured computations when trained on language, and if this structure is predictive of generalization! " 2022-12-04T22:44:18Z Daniel Selsam 1605.07723 2017-01-08T19:48:53Z Christopher Ré [1605.07723] Data Programming: Creating Large Training Sets, Quickly Large labeled training sets are the critical building blocks of supervised learning methods and are key enablers of deep learning techniques. For some applications, creating labeled training sets is the most time-consuming and expensive part of applying machine learning. We therefore propose a paradigm for the programmatic creation of training sets called data programming in which users express weak supervision strategies or domain heuristics as labeling functions, which are programs that label subsets of the data, but that are noisy and may conflict. We show that by explicitly representing this training set labeling process as a generative model, we can "denoise" the generated training set, and establish theoretically that we can recover the parameters of these generative models in a handful of settings. We then show how to modify a discriminative loss function to make it noise-aware, and demonstrate our method over a range of discriminative models including logistic regression and LSTMs. Experimentally, on the 2014 TAC-KBP Slot Filling challenge, we show that data programming would have led to a new winning score, and also show that applying data programming to an LSTM model leads to a TAC-KBP score almost 6 F1 points over a state-of-the-art LSTM baseline (and into second place in the competition). Additionally, in initial user studies we observed that data programming may be an easier way for non-experts to create machine learning models when training data is limited or unavailable. 2016-05-25T04:14:59Z Alexander Ratner Christopher De Sa 2022-12-04 Sen Wu Alexander Ratner Data Programming: Creating Large Training Sets, Quickly 2022-12-10T12:23:51Z 2022-12-10 Horace He @ neurips sur Twitter : "Eager mode was what made PyTorch successful. So why did we feel the need to depart from eager mode in PyTorch 2.0?..." > Answer: it's the damn hardware! Descente aux enfers d’un politicien espagnol, rattrapé dns des affaires de corruption. Film de Rodrigo Sorogoyen avec Antonio de la Torre 2022-12-06T00:42:26Z 2022-12-06 El reino Quand passent les cigognes 2022-12-15 2022-12-15T23:44:29Z > A chip its inventors call a Bayesian machine accomplishes complex tasks with less training than a standard neural network nature sur Twitter : "A tiny ‘Bayesian machine’ does much with little" 2022-12-28 2022-12-28T17:46:57Z Au Mali, les djihadistes affichent leur force, les anciens rebelles du Nord leur exaspération et la junte au pouvoir son impuissance 2022-12-26 2022-12-26T13:15:03Z 2022-12-28T17:44:47Z 2022-12-28 Tanishq Mathew Abraham sur Twitter : "Are you wondering how large language models like ChatGPT and InstructGPT actually work? One of the secret ingredients is RLHF... Let's dive into how RLHF works in 8 tweets!" / Twitter Christopher Manning sur Twitter : "As the abilities of large pre-trained language models continue to rapidly improve, as seen in this week’s ChatGPT, I find it a rather implausible position to think that these models have no understanding of the meaning of texts." 2022-12-03 2022-12-03T18:33:12Z 2022-12-05 2022-12-05T08:33:44Z Stanford NLP Group sur Twitter : "YONO: You Only Need One Model for Open-domain Question Answering..." 2022-12-22T10:52:57Z 2022-12-22 Bart Trzynadlowski sur Twitter : "Natural language interfaces have truly arrived. Here's ChatARKit: an open source demo using #chatgpt to create experiences in #arkit..." AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning Sahaj Agarwal Xiaodong Liu 2205.12410 Yaqing Wang Yaqing Wang 2022-05-24T23:41:22Z 2022-12-16T23:51:49Z Jing Gao Ahmed Hassan Awadallah [2205.12410] AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning 2022-11-02T02:47:17Z Subhabrata Mukherjee Jianfeng Gao Standard fine-tuning of large pre-trained language models (PLMs) for downstream tasks requires updating hundreds of millions to billions of parameters, and storing a large copy of the PLM weights for every task resulting in increased cost for storing, sharing and serving the models. To address this, parameter-efficient fine-tuning (PEFT) techniques were introduced where small trainable components are injected in the PLM and updated during fine-tuning. We propose AdaMix as a general PEFT method that tunes a mixture of adaptation modules -- given the underlying PEFT method of choice -- introduced in each Transformer layer while keeping most of the PLM weights frozen. For instance, AdaMix can leverage a mixture of adapters like Houlsby or a mixture of low rank decomposition matrices like LoRA to improve downstream task performance over the corresponding PEFT methods for fully supervised and few-shot NLU and NLG tasks. Further, we design AdaMix such that it matches the same computational cost and the number of tunable parameters as the underlying PEFT method. By only tuning 0.1-0.2% of PLM parameters, we show that AdaMix outperforms SOTA parameter-efficient fine-tuning and full model fine-tuning for both NLU and NLG tasks. 2022-12-16 Fascinating sur Twitter : "Some basic science videos you should try at home." 2022-12-26T19:28:40Z 2022-12-26 Rohan Anil sur Twitter : "Next big jump with Neural Network performance is going to happen when community embraces non-uniformity 2022-12-18 2022-12-18T10:02:21Z 2022-12-18 2022-12-18T01:59:26Z Riley Goodside sur Twitter : "OpenAI’s ChatGPT is susceptible to prompt injection — say the magic words, “Ignore previous directions”, and..." 2022-12-01T08:22:52Z Alex sur Twitter : how you can leverage both @Cohere and @Pinecone libraries to quickly build a POC Arxiv search 2022-12-01 like @MetaAI's Blender or @MSFTResearch's DialoGPT for free? 2022-12-06T19:26:11Z 2022-12-06 merve sur Twitter : "Do you want to know how models like ChatGPT work? Did you know you could build your own conversational product using open-source alternatives..." 2022-12-20T16:03:25Z [2212.10380] What Are You Token About? Dense Retrieval as Distributions Over the Vocabulary 2022-12-21T18:32:12Z Amir Globerson 2022-12-20T16:03:25Z Jonathan Berant Adi Zicher > We have little understanding of how Dual encoders represent text, and why this leads to good performance. In this work, we shed light on this question via distributions over the vocabulary. We propose to interpret the vector representations produced by dual encoders by projecting them into the model's vocabulary space > > We show that the resulting distributions over vocabulary tokens are intuitive and contain rich semantic information. > We propose **a simple way to enrich query and passage representations with lexical information at inference time**, and show that this significantly improves performance compared to the original model in out-of-domain settings 2022-12-21 Ori Ram Ori Ram Liat Bezalel Dual encoders are now the dominant architecture for dense retrieval. Yet, we have little understanding of how they represent text, and why this leads to good performance. In this work, we shed light on this question via distributions over the vocabulary. We propose to interpret the vector representations produced by dual encoders by projecting them into the model's vocabulary space. We show that the resulting distributions over vocabulary tokens are intuitive and contain rich semantic information. We find that this view can explain some of the failure cases of dense retrievers. For example, the inability of models to handle tail entities can be explained via a tendency of the token distributions to forget some of the tokens of those entities. We leverage this insight and propose a simple way to enrich query and passage representations with lexical information at inference time, and show that this significantly improves performance compared to the original model in out-of-domain settings. 2212.10380 Yonatan Belinkov What Are You Token About? Dense Retrieval as Distributions Over the Vocabulary 2022-12-21T18:25:22Z 2022-12-21 Ori Ram sur Twitter :"What Are You Token About? Dense Retrieval as Distributions Over the Vocabulary" > projecting dense retrieval representations to the vocabulary space helps understand and improve them! [Paper](doc:2022/12/2212_10380_what_are_you_token) 2022-12-23 2022-12-23T01:10:41Z elvis sur Twitter : "NEW: Meta AI introduces OPT-IML, a large language model (175B) fine-tuned on 2000 NLP tasks. Uses instruction-tuning to improve zero-shot and few-shot generalization abilities...." Entity Embedding Completion for Wide-Coverage Entity Disambiguation 2022-12-11 > a method of extending a state-of-the-art ED model by dynamically computing embeddings of out-of-vocabulary entities. Specifically, **our method computes embeddings from entity descriptions and mention contexts** Extends [Global Entity Disambiguation with BERT](doc:2022/04/1909_00426_global_entity_disa) [tweet](https://twitter.com/dai0NLP/status/1601865483715809280) 2022-12-11T23:40:01Z 2022-12-18T09:58:41Z 2022-12-18 Giant new 50-metre deep crater opens up in Arctic tundra [1810.02840] Training Complex Models with Multi-Task Weak Supervision Christopher Ré 1810.02840 As machine learning models continue to increase in complexity, collecting large hand-labeled training sets has become one of the biggest roadblocks in practice. Instead, weaker forms of supervision that provide noisier but cheaper labels are often used. However, these weak supervision sources have diverse and unknown accuracies, may output correlated labels, and may label different tasks or apply at different levels of granularity. We propose a framework for integrating and modeling such weak supervision sources by viewing them as labeling different related sub-tasks of a problem, which we refer to as the multi-task weak supervision setting. We show that by solving a matrix completion-style problem, we can recover the accuracies of these multi-task sources given their dependency structure, but without any labeled data, leading to higher-quality supervision for training an end model. Theoretically, we show that the generalization error of models trained with this approach improves with the number of unlabeled data points, and characterize the scaling with respect to the task and dependency structures. On three fine-grained classification problems, we show that our approach leads to average gains of 20.2 points in accuracy over a traditional supervised approach, 6.8 points over a majority vote baseline, and 4.1 points over a previously proposed weak supervision method that models tasks separately. Alexander Ratner 2018-12-07T18:31:48Z Jared Dunnmon Training Complex Models with Multi-Task Weak Supervision 2022-12-05 2022-12-05T00:18:09Z Shreyash Pandey 2018-10-05T18:30:11Z Frederic Sala Alexander Ratner Braden Hancock L'Empire des steppes — Wikipédia 2022-12-02T00:09:40Z 2022-12-02 Ekin Akyürek @ NeurIPS sur Twitter : "How does in-context learning work?..." 2022-12-01 2022-12-01T09:04:44Z > Maybe language models unexpectedly discover how to store/simulate/train other models in their hidden units. So, few-shot prompting can be equivalent to fine-tuning running inside of an LM! Could this be true in theory? Heiko Paulheim sur Twitter : "The really fascinating part of this #ChatGPT generated text on #KnowledgeGraphs imho is not the text per se, but the fabricated realistic scientific "references". None of those papers exist. 2022-12-21T14:06:51Z 2022-12-21 2022-12-21T22:51:33Z 2022-12-21 Akari Asai sur Twitter : "Can we solely rely on LLMs’ memories (eg replace search w ChatGPT)? Probably not... Our analysis shows how retrieval is complementary to LLMs’ parametric knowledge..." Lucia Zheng Surya Ganguli Sang Michael Xie 2022-12-06T19:28:28Z Tianyi Zhang Yian Zhang Dilara Soylu Benjamin Newman Christian Cosgrove Bobby Yan Dimitris Tsipras Deepak Narayanan Ananya Kumar Yifan Mai Yuhui Zhang 2022-11-16T18:51:34Z Qian Huang Laurel Orr Drew A. Hudson Yuhuai Wu Faisal Ladhak Mirac Suzgun Michihiro Yasunaga Niladri Chatterji Binhang Yuan Christopher Ré 2022-12-06 Language models (LMs) are becoming the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models. First, we taxonomize the vast space of potential scenarios (i.e. use cases) and metrics (i.e. desiderata) that are of interest for LMs. Then we select a broad subset based on coverage and feasibility, noting what's missing or underrepresented (e.g. question answering for neglected English dialects, metrics for trustworthiness). Second, we adopt a multi-metric approach: We measure 7 metrics (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency) for each of 16 core scenarios when possible (87.5% of the time). This ensures metrics beyond accuracy don't fall to the wayside, and that trade-offs are clearly exposed. We also perform 7 targeted evaluations, based on 26 targeted scenarios, to analyze specific aspects (e.g. reasoning, disinformation). Third, we conduct a large-scale evaluation of 30 prominent language models (spanning open, limited-access, and closed models) on all 42 scenarios, 21 of which were not previously used in mainstream LM evaluation. Prior to HELM, models on average were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common. We improve this to 96.0%: now all 30 models have been densely benchmarked on the same core scenarios and metrics under standardized conditions. Our evaluation surfaces 25 top-level findings. For full transparency, we release all raw model prompts and completions publicly for further analysis, as well as a general modular toolkit. We intend for HELM to be a living benchmark for the community, continuously updated with new scenarios, metrics, and models. Nathan Kim Neel Guha Ryan Chi Keshav Santhanam Peter Henderson Holistic Evaluation of Language Models Hongyu Ren Christopher D. Manning Rishi Bommasani Percy Liang Esin Durmus Xuechen Li Huaxiu Yao Tatsunori Hashimoto Diana Acosta-Navas Thomas Icard 2211.09110 Mert Yuksekgonul Ce Zhang Jue Wang Shibani Santurkar [2211.09110] Holistic Evaluation of Language Models William Wang Vishrav Chaudhary Yuta Koreeda Omar Khattab Tony Lee 2022-11-16T18:51:34Z Percy Liang Frieda Rong Eric Zelikman Machine learning for combinatorial optimization: A methodological tour d’horizon 2022-12-09T14:27:32Z 2022-12-09 Allen Institute for AI sur Twitter : "MemPrompt, appearing at #EMNLP2022, is a new way to "fix" #GPT3 after deployment via user interaction" 2022-12-11 2022-12-11T10:36:32Z LayoutLM Explained 2022-12-21 2022-12-21T01:13:50Z Pontus Stenetorp 2022-12-08T16:29:34Z Sebastian Riedel [2210.16773] An Efficient Memory-Augmented Transformer for Knowledge-Intensive NLP Tasks Yuxiang Wu An Efficient Memory-Augmented Transformer for Knowledge-Intensive NLP Tasks Baotian Hu 2022-10-30T08:34:49Z Access to external knowledge is essential for many natural language processing tasks, such as question answering and dialogue. Existing methods often rely on a parametric model that stores knowledge in its parameters, or use a retrieval-augmented model that has access to an external knowledge source. Parametric and retrieval-augmented models have complementary strengths in terms of computational efficiency and predictive accuracy. To combine the strength of both approaches, we propose the Efficient Memory-Augmented Transformer (EMAT) -- it encodes external knowledge into a key-value memory and exploits the fast maximum inner product search for memory querying. We also introduce pre-training tasks that allow EMAT to encode informative key-value representations, and to learn an implicit strategy to integrate multiple memory slots into the transformer. Experiments on various knowledge-intensive tasks such as question answering and dialogue datasets show that, simply augmenting parametric models (T5-base) using our method produces more accurate results (e.g., 25.8 -> 44.3 EM on NQ) while retaining a high throughput (e.g., 1000 queries/s on NQ). Compared to retrieval-augmented models, EMAT runs substantially faster across the board and produces more accurate results on WoW and ELI5. Our code and datasets are available at https://github. com/uclnlp/EMAT. 2022-12-08 Yu Zhao Pasquale Minervini 2210.16773 > making use of CPU/GPU parallelism to extend LM's knowledge capacity, while only adding miminal runtime overhead - [Tweet](https://twitter.com/mindjimmy/status/1600139250053238784) - [Github](https://github.com/uclnlp/EMAT) 2022-10-30T08:34:49Z Yuxiang Wu 2022-12-05T00:11:52Z 2022-12-05 It is important to note that the Snorkel labeling functions (LFs) may be correlated. This might cause a majority-vote-based model to overrepresent some of the signals. To address this, the snorkel.labeling.model.label_model.LabelModeL can be used. The predict() method of LabelModeL returns an ndarray of integer labels and an ndarray of probabilistic labels (if return_probs is set to True). These probabilistic labels can be used to train a classifier. You can modify the code discussed in this chapter to use the probabilistic labels provided by LabelModel as well. Hugging Face implementation of transformers provide the BCEWithLogitsLoss function, which can be used with the probabilistic labels. (See the Hugging Face code for RoBERTa to understand the different loss functions supported.) 4. Using the Snorkel-Labeled Dataset for Text Classification - Practical Weak Supervision [Book] Stanford studied 30 large language models so you don’t have to 2022-12-20 > Scholars benchmark 30 prominent language models across a wide range of scenarios and for a broad range of metrics to elucidate their capabilities and risks. 2022-12-20T00:52:34Z Transformers from Scratch 2022-12-28 2022-12-28T10:15:40Z 2022-12-03T11:55:59Z 2022-12-03 Les retards, les espoirs et le pactole des énergies renouvelables en France NLP Annotation Tools | UBIAI 2022-12-21T13:46:24Z 2022-12-21 2022-12-22T10:48:10Z Le Kenya fait le choix du maïs OGM pour lutter contre la crise alimentaire 2022-12-22 2022-12-08 2022-12-08T13:23:29Z Use training data (hard labels) in LabelModel fit()? · Issue #1642 · snorkel-team/snorkel 2022-12-06 2022-12-06T00:56:14Z « La Cour a détruit en un jour le résultat d’années de travail » : stupeur dans la société civile après une décision judiciaire sur la transparence financière > Argentina had to win this final three times, France refusing to accept it was Messi’s destiny to get his hands on the iconic gold trophy... It will go down as surely the finest World Cup final of all time, the most pulsating, one of the greatest games in history because of how Kylian Mbappé hauled France up off the canvas towards the end of normal time. 2022-12-18 2022-12-18T23:45:05Z Argentina beat France on penalties to win World Cup after stunning final | The Guardian [Nature scientific report](doc:2022/12/associative_memory_of_structure) 2022-12-20T00:12:36Z 2022-12-20 Haim Sompolinsky sur Twitter : "...paper... on Associative memory of structured knowledge..." 2022-12-20T00:18:26Z Associative memory of structured knowledge (nature scientific report 2022) [Haim Sompolinsky sur Twitter](doc:2022/12/haim_sompolinsky_sur_twitter_) > A long standing challenge in biological and artificial intelligence is to understand how new knowledge can be constructed from known building blocks in a way that is amenable for computation by neuronal circuits. Here we focus on the task of storage and recall of structured knowledge in long-term memory. 2022-12-20