Podcasts about Backpropagation

Artificial intelligence is evolving at an unprecedented pace—what does that mean for the future of technology, venture capital, business, and even our understanding of ourselves? Award-winning journalist and writer Anil Ananthaswamy joins us for our latest episode to discuss his latest book Why Machines Learn: The Elegant Math Behind Modern AI.Anil helps us explore the journey and many breakthroughs that have propelled machine learning from simple perceptrons to the sophisticated algorithms shaping today's AI revolution, powering GPT and other models. The discussion aims to demystify some of the underlying mathematical concepts that power modern machine learning, to help everyone grasp this technology impacting our lives–even if your last math class was in high school. Anil walks us through the power of scaling laws, the shift from training to inference optimization, and the debate among AI's pioneers about the road to AGI—should we be concerned, or are we still missing key pieces of the puzzle? The conversation also delves into AI's philosophical implications—could understanding how machines learn help us better understand ourselves? And what challenges remain before AI systems can truly operate with agency?If you enjoy this episode, please subscribe and leave us a review on your favorite podcast platform. Sign up for our newsletter at techsurgepodcast.com for exclusive insights and updates on upcoming TechSurge Live Summits.Links:Read Why Machines Learn, Anil's latest book on the math behind AIhttps://www.amazon.com/Why-Machines-Learn-Elegant-Behind/dp/0593185749Learn more about Anil Ananthaswamy's work and writinghttps://anilananthaswamy.com/Watch Anil Ananthaswamy's TED Talk on AI and intelligencehttps://www.ted.com/speakers/anil_ananthaswamyDiscover the MIT Knight Science Journalism Fellowship that shaped Anil's AI researchhttps://ksj.mit.edu/Understand the Perceptron, the foundation of neural networkshttps://en.wikipedia.org/wiki/PerceptronRead about the Perceptron Convergence Theorem and its significancehttps://www.nature.com/articles/323533a0

Anil Ananthaswamy is a renowned science writer and journalist who has written extensively on various scientific topics. In his latest book "Why Machines Learn", Anil explores the fascinating world of artificial intelligence and machine learning. He reveals the intricate mechanisms and complex algorithms that underlie these cutting-edge technologies. Join us for a fascinating conversation with science writer Anil Ananthaswamy as he shares insights from his book and sheds light on the rapidly evolving field of AI. Tune in to gain a deeper understanding of how these machines work at a basic mathematics level. Resource List - Why Machines Learn, book by Anil Ananthaswamy - https://amzn.in/d/bmirU45 Dartmouth Summer Research Project on Artificial Intelligence - https://home.dartmouth.edu/about/artificial-intelligence-ai-coined-dartmouth What is the Perceptron artificial neural network? - https://www.geeksforgeeks.org/what-is-perceptron-the-simplest-artificial-neural-network/ Read about the McCulloch-Pitts Artificial Neuron - https://towardsdatascience.com/mcculloch-pitts-model-5fdf65ac5dd1 Nobel Prize in Physics 2024 - https://www.nobelprize.org/prizes/physics/2024/press-release/ What is the Hopfield Neural Network? - https://www.geeksforgeeks.org/hopfield-neural-network/ Read about Backpropagation - https://en.wikipedia.org/wiki/Backpropagation “Learning representations by back-propagating errors”, paper by Geoffrey Hinton, David Rumelhart and Ronald Williams - https://www.nature.com/articles/323533a0 AlexNet by Geoffery Hinton and team - https://en.wikipedia.org/wiki/AlexNet What is ImageNet? - https://www.image-net.org/about.php ‘Attention Is All You Need', transformer architecture paper - https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf What are Neural Scaling Laws? - https://en.wikipedia.org/wiki/Neural_scaling_law NeuroAI - https://neuro-ai.com/ About SparX by Mukesh Bansal SparX is a podcast where we delve into cutting-edge scientific research, stories from impact-makers and tools for unlocking the secrets to human potential and growth. We believe that entrepreneurship, fitness and the science of productivity is at the forefront of the India Story; the country is at the cusp of greatness and at SparX, we wish to make these tools accessible for every generation of Indians to be able to make the most of the opportunities around us. In a new episode every Sunday, our host Mukesh Bansal (Founder Myntra and Cult.fit) will talk to guests from all walks of life and also break down everything he's learnt about the science of impact over the course of his 20-year long career. This is the India Century, and we're enthusiastic to start this journey with you. Follow us on Instagram: / sparxbymukeshbansal Website: https://www.sparxbymukeshbansal.com You can also listen to SparX on all audio platforms Fasion | Outbreak | Courtesy EpidemicSound.com

University of Washington の浜崎伸彦さん(@Nobu_Hamazaki)がゲスト。Human RA-Gastruloid論文、条件探索のコツ、ステージング問題、卵子形成論文、PI生活、今後目指す合成的発生生物学の方向性など (12/12収録) Show Notes (番組HP): 浜崎ラボHP 博士課程(の後半)での所属ラボ：中島欽一ラボ 日本での所属ラボ：林克彦ラボ (現在は阪大) 日本にいたときの卵子形成転写因子論文 かぐや論文 Parthenogenesisのレビュー 留学先の所属ラボ：Jay Shendureラボ 留学先でのHuman RA-Gastruloid論文 最初にあったマウスGastruloid論文 Matrigel おすすめの発生生物学の教科書ギルバート発生生物学、ウォルパート発生生物学とラングマン人体発生学とそのムービー JayのGestalt イベントのレコーディング シュペーマンのオーガナイザー (pdf) Turner Syndrome Wntアゴニストのカイロン 下流のBrachyury (T)という転写因子 最近出ていた脊索が現れる3D trunk organoid論文 bioRxivに最初に出た論文 とその数日後にアップロードされた競合論文 その１ ２ ３ ４ の出版論文としての結果 1 2 3 4 最近溜まっているEmbryonic single cell data 1 2 3 4 5 6 7 10x sci-RNA-seq3 Smart-seq3 Wei Yangさん（Jayラボメンバーページ） マーカージーンをGPTに入れると予測してくる Mappingする方法  UMAPのco-embedding Foundation modelを作っていく例 今村さん とやっていたgene regulation Zygotic genome activation 精子のヒストンがプロタミンに置き換わる機構 残ったヒストンがinter-generationalなepigenetic memoryに関与する？（pdf） 父が肥満だと子が肥満になりやすい、母だとそうでもない ダーウィン vs ルイセンコ iPSで精子作成している例 減数分裂？に関するプレプリント パキテンで止まっている 山中論文は24から4へ 卵子が大きくなるためにはアクアポリンが必要。

In dieser Folge graben wir uns tief in die Gehirne zweier Genies: John J. Hopfield und Geoffrey Hinton – Pioniere der Künstlichen Intelligenz, ohne die dein Smartphone, dein Laptop und die nette Dame von der Hotline einfach nur... ja, nicht existieren würden. Hopfield-Netzwerke, Boltzmann-Maschinen, Backpropagation-Algorithmen. Klingt kompliziert? Keine Sorge, wir brechen es runter. Erfahre, wie Hopfield in den 80ern ein Modell des menschlichen Gehirns entwickelte, das heute noch die Grundlage für moderne KI bildet. Oder wie Hinton mit seinen bahnbrechenden Ideen neuronale Netze auf das nächste Level brachte – und uns alle auf eine Achterbahnfahrt in die KI-Zukunft schickte. Doch wie gefährlich kann das Ganze werden? Hinton selbst warnt: KI könnte uns irgendwann überflügeln – und dann? Diese Episode liefert dir spannende Geschichten über Nobelpreise, künstliche Intelligenz und den schmalen Grat zwischen technologischem Fortschritt und Science-Fiction-Dystopie. CHAPTERS (00:00) Intro (00:01) Hast du schonmal einen Preis bekommen? (00:02) Nobelpreisgeschichte (00:07) John Hopfield (00:12) Geoffrey Hinton (00:13) Backpropagation (00:23) Wie sehen wir die Zukunft? (00:31) Outtakes (00:32) Outro LINKS https://www.tagesspiegel.de/wissen/grundlage-fur-chatgpt-und-co-physik-nobelpreis-geht-an-ki-forscher-fur-maschinen-die-lernen-12498434.html https://www.zeit.de/wissen/2024-10/nobelpreis-physik-kuenstliche-intelligenz-geoffry-hinton-john-hopfield https://www.youtube.com/watch?v=Hc-RpjTpHFg https://www.youtube.com/watch?v=BcG9KEdiZfw

Jürgen Schmidhuber, the father of generative AI shares his groundbreaking work in deep learning and artificial intelligence. In this exclusive interview, he discusses the history of AI, some of his contributions to the field, and his vision for the future of intelligent machines. Schmidhuber offers unique insights into the exponential growth of technology and the potential impact of AI on humanity and the universe. YT version: https://youtu.be/DP454c1K_vQ MLST is sponsored by Brave: The Brave Search API covers over 20 billion webpages, built from scratch without Big Tech biases or the recent extortionate price hikes on search API access. Perfect for AI model training and retrieval augmentated generation. Try it now - get 2,000 free queries monthly at http://brave.com/api. TOC 00:00:00 Intro 00:03:38 Reasoning 00:13:09 Potential AI Breakthroughs Reducing Computation Needs 00:20:39 Memorization vs. Generalization in AI 00:25:19 Approach to the ARC Challenge 00:29:10 Perceptions of Chat GPT and AGI 00:58:45 Abstract Principles of Jurgen's Approach 01:04:17 Analogical Reasoning and Compression 01:05:48 Breakthroughs in 1991: the P, the G, and the T in ChatGPT and Generative AI 01:15:50 Use of LSTM in Language Models by Tech Giants 01:21:08 Neural Network Aspect Ratio Theory 01:26:53 Reinforcement Learning Without Explicit Teachers Refs: ★ "Annotated History of Modern AI and Deep Learning" (2022 survey by Schmidhuber): ★ Chain Rule For Backward Credit Assignment (Leibniz, 1676) ★ First Neural Net / Linear Regression / Shallow Learning (Gauss & Legendre, circa 1800) ★ First 20th Century Pioneer of Practical AI (Quevedo, 1914) ★ First Recurrent NN (RNN) Architecture (Lenz, Ising, 1920-1925) ★ AI Theory: Fundamental Limitations of Computation and Computation-Based AI (Gödel, 1931-34) ★ Unpublished ideas about evolving RNNs (Turing, 1948) ★ Multilayer Feedforward NN Without Deep Learning (Rosenblatt, 1958) ★ First Published Learning RNNs (Amari and others, ~1972) ★ First Deep Learning (Ivakhnenko & Lapa, 1965) ★ Deep Learning by Stochastic Gradient Descent (Amari, 1967-68) ★ ReLUs (Fukushima, 1969) ★ Backpropagation (Linnainmaa, 1970); precursor (Kelley, 1960) ★ Backpropagation for NNs (Werbos, 1982) ★ First Deep Convolutional NN (Fukushima, 1979); later combined with Backprop (Waibel 1987, Zhang 1988). ★ Metalearning or Learning to Learn (Schmidhuber, 1987) ★ Generative Adversarial Networks / Artificial Curiosity / NN Online Planners (Schmidhuber, Feb 1990; see the G in Generative AI and ChatGPT) ★ NNs Learn to Generate Subgoals and Work on Command (Schmidhuber, April 1990) ★ NNs Learn to Program NNs: Unnormalized Linear Transformer (Schmidhuber, March 1991; see the T in ChatGPT) ★ Deep Learning by Self-Supervised Pre-Training. Distilling NNs (Schmidhuber, April 1991; see the P in ChatGPT) ★ Experiments with Pre-Training; Analysis of Vanishing/Exploding Gradients, Roots of Long Short-Term Memory / Highway Nets / ResNets (Hochreiter, June 1991, further developed 1999-2015 with other students of Schmidhuber) ★ LSTM journal paper (1997, most cited AI paper of the 20th century) ★ xLSTM (Hochreiter, 2024) ★ Reinforcement Learning Prompt Engineer for Abstract Reasoning and Planning (Schmidhuber 2015) ★ Mindstorms in Natural Language-Based Societies of Mind (2023 paper by Schmidhuber's team) https://arxiv.org/abs/2305.17066 ★ Bremermann's physical limit of computation (1982) EXTERNAL LINKS CogX 2018 - Professor Juergen Schmidhuber https://www.youtube.com/watch?v=17shdT9-wuA Discovering Neural Nets with Low Kolmogorov Complexity and High Generalization Capability (Neural Networks, 1997) https://sferics.idsia.ch/pub/juergen/loconet.pdf The paradox at the heart of mathematics: Gödel's Incompleteness Theorem - Marcus du Sautoy https://www.youtube.com/watch?v=I4pQbo5MQOs (Refs truncated, full version on YT VD)

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: ML Safety Research Advice - GabeM, published by Gabe M on July 23, 2024 on The AI Alignment Forum. This is my advice for careers in empirical ML research that might help AI safety (ML Safety). Other ways to improve AI safety, such as through AI governance and strategy, might be more impactful than ML safety research (I generally think they are). Skills can be complementary, so this advice might also help AI governance professionals build technical ML skills. 1. Career Advice 1.1 General Career Guides Preventing an AI-related catastrophe - 80,000 Hours A Survival Guide to a PhD (Andrej Karpathy) How to pursue a career in technical AI alignment - EA Forum AI safety technical research - Career review - 80,000 Hours Beneficial AI Research Career Advice 2. Upskilling 2.1 Fundamental AI Safety Knowledge AI Safety Fundamentals - BlueDot Impact AI Safety, Ethics, and Society Textbook Forming solid AI safety threat models helps you select impactful research ideas. 2.2 Speedrunning Technical Knowledge in 12 Hours Requires some basic coding, calculus, and linear algebra knowledge Build Intuition for ML (5h) Essence of linear algebra - 3Blue1Brown (3h) Neural networks - 3Blue1Brown (2h) Backpropagation, the foundation of deep learning (3h) Neural Networks: Backpropagation - CS 231N (0.5h) The spelled-out intro to neural networks and backpropagation: building micrograd (2.5h) Transformers and LLMs (4h) [1hr Talk] Intro to Large Language Models (1h) The Illustrated Transformer - Jay Alammar (1h) Let's build GPT: from scratch, in code, spelled out. (2h) 2.3 How to Build Technical Skills Traditionally, people take a couple of deep learning classes. Stanford CS 224N | Natural Language Processing with Deep Learning (lecture videos) Practical Deep Learning for Coders - Practical Deep Learning (fast.ai) Other curricula that seem good: Syllabus | Intro to ML Safety Levelling Up in AI Safety Research Engineering [Public] ARENA Maybe also check out recent topical classes like this with public lecture recordings: CS 194/294-267 Understanding Large Language Models: Foundations and Safety Beware of studying too much. You should aim to understand the fundamentals of ML through 1 or 2 classes and then practice doing many manageable research projects with talented collaborators or a good mentor who can give you time to meet. It's easy to keep taking classes, but you tend to learn many more practical ML skills through practice doing real research projects. You can also replicate papers to build experience. Be sure to focus on key results rather than wasting time replicating many experiments. "One learns from books and reels only that certain things can be done. Actual learning requires that you do those things." -Frank Herbert Note that ML engineering skills will be less relevant over time as AI systems become better at writing code. A friend didn't study computer science but got into MATS 2023 with good AI risk takes. Then, they had GPT-4 write most of their code for experiments and did very well in their stream. Personally, GitHub Copilot and language model apps with code interpreters/artifacts write a significant fraction of my code. However, fundamental deep learning knowledge is still useful for making sound decisions about what experiments to run. 2.4 Math You don't need much of it to do empirical ML research. Someone once told me, "You need the first chapter of a calculus textbook and the first 5 pages of a linear algebra textbook" to understand deep learning. You need more math for ML theory research, but theoretical research is not as popular right now. Beware mathification: authors often add unnecessary math to appease (or sometimes confuse) conference reviewers. If you don't understand some mathematical notation in an empirical paper, you can often send a screenshot to an LLM chatbot f...

Yapay zeka serimizin beşinci bölümünde sinir ağlarının eğitimine odaklanıyoruz. İşin detayına gireceğiz ve üç önemli kavram göreceğiz:Kayıp fonksiyonu veya hata fonksiyonuhatayı azaltmanın bir yöntemi olan gradient descentve onun etkili biçimde uygulanmasını sağlayan backpropagation tekniği..Konular:(01:53) Cost function(05:01) Gradient descent(11:04) Backpropagation(15:28) Vanishing Gradient(16:51) Test vs Eğitim(17:47) Patreon TeşekkürleriKaynaklar:Video: The Most Important Algorithm in Machine LearningVideo: Backpropagation explained | Part 1 - The intuitionVideo: Watching Neural Networks Learn.------- Podbee Sunar -------Bu podcast, AgeSA hakkında reklam içerir.AgeSA BES ile Yatırımda Rahat Edin. Yüksek kazançlı geniş fon seçenekleri, %30 Devlet Katkısı ve finansal danışmanlık AgeSA'da. Yatırımlarınla ilgili daha iyi hissetmek için tıkla.SummarySee Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Dr. Blake Richards discusses the world of AI and human cognition this week. Learn about the essence of intelligence, the ways AI research informs our understanding of the human brain, and discover the potential future scenarios where AI and humanity might intersect. This episode is brought to you by Gurobi (https://gurobi.com/sds), the Decision Intelligence Leader, and by CloudWolf (https://www.cloudwolf.com/sds), the Cloud Skills platform. Interested in sponsoring a SuperDataScience Podcast episode? Visit JonKrohn.com/podcast for sponsorship information. In this episode you will learn: • Blake's research and his take on intelligence [09:56] • How we can evaluate progress in artificial general intelligence [15:54] • Blake's thoughts on biomimicry [20:57] • Why Blake thinks the fears regarding AI are overdone [25:38] • The most effective strategies to mitigate AI fears without hindering innovation [35:31] • What steps can we take to ensure that AI supports human flourishing [45:23] • The importance of interpreting neuroscience data through the lens of ML [55:08] • Backpropagation, gradient descent and the brain [1:17:32] Additional materials: www.superdatascience.com/729

While an elephant may never forget, the same cannot be said for artificial neural networks. What is catastrophic forgetting, how does it affect artificial intelligence and how are engineers trying to solve the problem? See omnystudio.com/listener for privacy information.

Backpropagation was one of the innovations by Geoff Hinton that made deep learning networks a practical reality. But have you ever heard of that term before and know what it is at a high level? In this episode of the AI Today podcast hosts Kathleen Walch and Ron Schmelzer define the terms Backpropagation, Learning Rate, and Optimizer, explain how these terms relates to AI and why it's important to know about them. Continue reading AI Today Podcast: AI Glossary Series – Backpropagation, Learning Rate, and Optimizer at AI & Data Today.

In this episode of The AI Frontier, join us as we embark on a journey through the history of deep learning and artificial intelligence. From the earliest days of linear regression to the latest advancements in generative adversarial networks, we will explore the key moments and milestones that have shaped the development of this groundbreaking field. Learn about the pioneers and trailblazers who pushed the boundaries of what was possible, and discover how deep learning has revolutionized the way we think about and interact with technology. Get ready to delve deep into the history of AI!Support the Show.Keep AI insights flowing – become a supporter of the show!Click the link for details

Link to bioRxiv paper: http://biorxiv.org/cgi/content/short/2023.02.23.529770v1?rss=1 Authors: Cheng, H., Brown, J. Abstract: How episodic memories are formed in the brain is an outstanding puzzle for the neuroscience community. The brain areas that are critical for episodic learning (e.g., the hippocampus) are characterized by recurrent connectivity and generate frequent offline replay events. The function of the replay events is a subject of active debate. Recurrent connectivity, computational simulations show, enables sequence learning when combined with a suitable learning algorithm such as Backpropagation through time (BPTT). BPTT, however, is not biologically plausible. We describe here, for the first time, a biologically plausible variant of BPTT in a reversible recurrent neural network, R2N2, that critically leverages offline-replay to support episodic learning. The model uses forwards and backwards offline replay to transfer information between two recurrent neural networks, a cache and a consolidator,that perform rapid one-shot learning and statistical learning, respectively. Unlike replay in standard BPTT, this architecture requires no artificial external memory store. This architecture and approach outperform existing solutions and account for the functional significance to hippocampal replay events. We demonstrate the R2N2 network properties using benchmark tests from computer science and simulate the rodent delayed alternation T-maze task. Copy rights belong to original authors. Visit the link for more info Podcast created by Paper Player, LLC

Andrej Karpathy, AI Transformers and Backpropagation

Array Cast - August 19, 2022 Show NotesMany thanks to Marshall Lochbaum, Rodrigo Girão Serrão, Bob Therriault, Conor Hoekstra, Adám Brudzewsky and Romilly Cocking for gathering these links:[01] 00:00:03 BYTE magazine https://en.wikipedia.org/wiki/Byte_(magazine)[02] 00:01:02 Org Mode https://orgmode.org/[03] 00:02:58 Toronto Meet-up https://www.meetup.com/en-AU/programming-languages-toronto-meetup/events/287695788/ New York Meet-up https://www.meetup.com/programming-languages-toronto-meetup/events/287729348/[04] 00:04:19 Morten Kromberg episode https://www.arraycast.com/episodes/episode21-morten-kromberg[05] 00:05:01 Romilly's video 'An Excellent Return' https://dyalog.tv/Dyalog08/?v=thr-7QfQWJw[06] 00:06:12 Ferranti Pegasus computer https://en.wikipedia.org/wiki/Ferranti_Pegasus[07] 00:09:09 System 360 in APL http://keiapl.org/archive/APL360_UsersMan_Aug1968.pdf[08] 00:16:50 Mind Map https://en.wikipedia.org/wiki/Mind_map[09] 00:17:00 Dyalog https://www.dyalog.com/[10] 00:18:20 Digitalk https://winworldpc.com/product/digital-smalltalk/5x[11] 00:18:30 Smalltalk https://en.wikipedia.org/wiki/Smalltalk[12] 00:21:17 Raspberry Pi https://www.raspberrypi.org/[13] 00:22:10 Robotics on Dyalog website https://www.dyalog.com/blog/2014/08/dancing-with-the-bots/[14] 00:22:45 Neural Network https://en.wikipedia.org/wiki/Neural_network David Marr https://en.wikipedia.org/wiki/David_Marr_(neuroscientist)[15] 00:23:21 Jetson Nano https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-nano/[16] 00:23:38 Spiking neural networks https://en.wikipedia.org/wiki/Spiking_neural_network[17] 00:24:02 JAX https://jax.readthedocs.io/en/latest/notebooks/quickstart.html[18] 00:27:00 Numpy https://numpy.org/[19] 00:28:21 Nested arrays https://aplwiki.com/wiki/Nested_array[20] 00:29:07 flip Numpy https://numpy.org/doc/stable/reference/generated/numpy.flip.html flipud https://numpy.org/doc/stable/reference/generated/numpy.flipud.html#numpy.flipud[21] 00:31:07 Pegasus Autocode http://blog.rareschool.com/2014/09/pegasus-autocode-revisited.html[22] 00:32:05 Atlas computer 1966 https://en.wikipedia.org/wiki/Atlas_(computer)[23] 00:34:30 Raspberry Pi pico https://www.raspberrypi.com/products/raspberry-pi-pico/[24] 00:36:33 Booker and Morris https://dl.acm.org/doi/pdf/10.1145/364520.364521[25] 00:38:12 Romilly's Blog Markdown http://blog.rareschool.com/2022/05/apl-and-python-go-head-to-head.html[26] 00:41:30 Languages that are built from concatenation https://en.wikipedia.org/wiki/Agglutination[27] 00:44:30 Alan Kay https://en.wikipedia.org/wiki/Alan_Kay[28] 00:47:12 Clojure https://en.wikipedia.org/wiki/Alan_Kay Forth https://en.wikipedia.org/wiki/Forth_(programming_language) Haskell https://www.haskell.org/[29] 00:50:00 Cosy http://www.cosy.com/language/[30] 00:51:38 Py'n'APL https://dyalog.tv/Dyalog21/?v=gOUFXBUMv_A[31] 01:00:12 Logic Analyzer https://en.wikipedia.org/wiki/Logic_analyzer[32] 01:02:15 Back propagation in neural networks https://en.wikipedia.org/wiki/Backpropagation[33] 01:07:38 Stefan Kruger 'Learn APL' https://xpqz.github.io/learnapl/intro.html[34] 01:08:10 Rodrigo Girão Serrão videos https://www.youtube.com/channel/UCd_24S_cYacw6zrvws43AWg[35] 01:08:27 João Araújo episode https://www.arraycast.com/episodes/episode33-joao-araujo[36] 01:08:59 Rodrigo Girão Serrão Neural networks https://www.youtube.com/playlist?list=PLgTqamKi1MS3p-O0QAgjv5vt4NY5OgpiM[37] 01:10:55 Functional Geekery podcast https://www.functionalgeekery.com/[38] 01:11:36 Conor's Security talk https://www.youtube.com/watch?v=ajGX7odA87k[39] 01:12:38 SICP https://en.wikipedia.org/wiki/Structure_and_Interpretation_of_Computer_Programs[40] 01:12:55 Alan McKean Rebecca Wirfs-Brock "Object Design" https://books.google.ca/books?id=vUF72vN5MY8C&printsec=copyright&redir_esc=y#v=onepage&q&f=false[41] 01:13:35 Growing Object Oriented Guided by Tests http://www.growing-object-oriented-software.com/[42] 01:15:01 Design Patterns vs Anti pattern in APL https://www.youtube.com/watch?v=v7Mt0GYHU9A[43] 01:18:25 Pop2 https://hopl.info/showlanguage.prx?exp=298&language=POP-2 Pop2 on pdf-11 https://www.cs.bham.ac.uk/research/projects/poplog/retrieved/adrian-howard-pop11.html[44] 01:18:52 Donald Michie https://en.wikipedia.org/wiki/Donald_Michie[45] 01:21:30 Menace robot http://chalkdustmagazine.com/features/menace-machine-educable-noughts-crosses-engine/[46] 01:22:05 Menace in APL https://romilly.github.io/o-x-o/an-introduction.html

Beren is a postdoc in Oxford with a background in machine learning and computational neuroscience. He is interested in Active Inference (related to the Free Energy Principle) and how the cortex can perform long-term credit assignment as deep artificial neural networks do. We start off with some shorter questions on the Free Energy Principle and its background concepts. Next, we get onto the exploration vs exploitation dilemma in reinforcement learning and Beren's strategy on how to maximize expected reward from restaurant visits - it's a long episode :=). We also discuss multimodal representations, shallow minima, autism and enactivism. Then, we explore predictive coding going all the way from the phenomenon of visual fading, to 20-eyed reinforcement learning agents and the 'Anti-Grandmother Cell'. Finally, we discuss some open questions about backpropagation and the role of time in the brain, and finish the episode with some career advice about writing, publishing and Beren's future projects! Timestamps: (00:00) - Intro (02:11) - The Free Energy Principle, Active Inference, and Reinforcement Learning (13:40) - Exploration vs Exploitation (26:47) - Multimodal representation, shallow minima, autism (36:11) - Biased generative models, enactivism, and representation in the brain? (45:21) - Fixational eye movements, predictive coding, and 20-eyed RL (52:57) - Precision, attention, and dopamine (01:01:51) - Sparsity, negative prediction errors, and the 'Anti-Grandmother Cell' (01:11:23) - Backpropagation in the brain? (01:19:25) - Time in machine learning and the brain? (01:25:32) - Career Questions Beren's Twitter: https://twitter.com/BerenMillidge Paper: Deep active inference as variational policy gradients: https://www.sciencedirect.com/science/article/abs/pii/S0022249620300298 Paper: Predictive Coding Approximates Backprop Along Arbitrary Computation Graphs: https://direct.mit.edu/neco/article-abstract/34/6/1329/110646/Predictive-Coding-Approximates-Backprop-Along Paper: Predictive Coding: a Theoretical and Experimental Review: https://arxiv.org/abs/2107.12979 Beren's other work: https://scholar.google.gr/citations?user=3GGkFTkAAAAJ&hl=en My Twitter https://twitter.com/Embodied_AI My LinkedIn https://www.linkedin.com/in/akseli-ilmanen-842098181/

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Introducing the ML Safety Scholars Program, published by ThomasWoodside on May 4, 2022 on The Effective Altruism Forum. Program Overview The Machine Learning Safety Scholars program is a paid, 9-week summer program designed to help undergraduate students gain skills in machine learning with the aim of using those skills for empirical AI safety research in the future. Apply for the program here by May 31st. The course will have three main parts: Machine learning, with lectures and assignments from MIT Deep learning, with lectures and assignments from the University of Michigan, NYU, and Hugging Face ML safety, with lectures and assignments produced by Dan Hendrycks at UC Berkeley The first two sections are based on public materials, and we plan to make the ML safety course publicly available soon as well. The purpose of this program is not to provide proprietary lessons but to better facilitate learning: The program will have a Slack, regular office hours, and active support available for all Scholars. We hope that this will provide useful feedback over and above what's possible with self-studying. The program will have designated “work hours” where students will cowork and meet each other. We hope this will provide motivation and accountability, which can be hard to get while self-studying. We will pay Scholars a $4,500 stipend upon completion of the program. This is comparable to undergraduate research roles and will hopefully provide more people with the opportunity to study ML. MLSS will be fully remote, so participants will be able to do it from wherever they're located. Why have this program? Much of AI safety research currently focuses on existing machine learning systems, so it's necessary to understand the fundamentals of machine learning to be able to make contributions. While many students learn these fundamentals in their university courses, some might be interested in learning them on their own, perhaps because they have time over the summer or their university courses are badly timed. In addition, we don't think that any university currently devotes multiple weeks to AI Safety. There are already sources of funding for upskilling within EA, such as the Long Term Future Fund. Our program focuses specifically on ML and therefore we are able to provide a curriculum and support to Scholars in addition to funding, so they can focus on learning the content. Our hope is that this program can contribute to producing knowledgeable and motivated undergraduates who can then use their skills to contribute to the most pressing research problems within AI safety. Time Commitment The program will last 9 weeks, beginning on Monday, June 20th, and ending on August 19th. We expect each week of the program to cover the equivalent of about 3 weeks of the university lectures we are drawing our curriculum from. As a result, the program will likely take roughly 30-40 hours per week, depending on speed and prior knowledge. Preliminary Content & Schedule Machine Learning (content from the MIT open course) Week 1 - Basics, Perceptrons, Features Week 2 - Features continued, Margin Maximization (logistic regression and gradient descent), Regression Deep Learning (content from a University of Michigan course as well as an NYU course) Week 3 - Introduction, Image Classification, Linear Classifiers, Optimization, Neural Networks. ML Assignments due. Week 4 - Backpropagation, CNNs, CNN Architectures, Hardware and Software, Training Neural Nets I & II. DL Assignment 1 due. Week 5 - RNNs, Attention, NLP (from NYU), Hugging Face tutorial (parts 1-3), RL overview. DL Assignment 2 due. ML Safety Week 6 - Risk Management Background (e.g., accident models), Robustness (e.g., optimization pressure). DL Assignment 3 due. Week 7 - Monitoring (e.g., emergent capabilities), Alignment (e.g., honesty...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Introducing the ML Safety Scholars Program, published by Dan Hendrycks on May 4, 2022 on The AI Alignment Forum. Program Overview The Machine Learning Safety Scholars program is a paid, 9-week summer program designed to help undergraduate students gain skills in machine learning with the aim of using those skills for empirical AI safety research in the future. Apply for the program here by May 31st. The course will have three main parts: Machine learning, with lectures and assignments from MIT Deep learning, with lectures and assignments from the University of Michigan, NYU, and Hugging Face ML safety, with lectures and assignments produced by Dan Hendrycks at UC Berkeley The first two sections are based on public materials, and we plan to make the ML safety course publicly available soon as well. The purpose of this program is not to provide proprietary lessons but to better facilitate learning: The program will have a Slack, regular office hours, and active support available for all Scholars. We hope that this will provide useful feedback over and above what's possible with self-studying. The program will have designated “work hours” where students will cowork and meet each other. We hope this will provide motivation and accountability, which can be hard to get while self-studying. We will pay Scholars a $4,500 stipend upon completion of the program. This is comparable to undergraduate research roles and will hopefully provide more people with the opportunity to study ML. MLSS will be fully remote, so participants will be able to do it from wherever they're located. Why have this program? Much of AI safety research currently focuses on existing machine learning systems, so it's necessary to understand the fundamentals of machine learning to be able to make contributions. While many students learn these fundamentals in their university courses, some might be interested in learning them on their own, perhaps because they have time over the summer or their university courses are badly timed. In addition, we don't think that any university currently devotes multiple weeks to AI Safety. There are already sources of funding for upskilling within EA, such as the Long Term Future Fund. Our program focuses specifically on ML and therefore we are able to provide a curriculum and support to Scholars in addition to funding, so they can focus on learning the content. Our hope is that this program can contribute to producing knowledgeable and motivated undergraduates who can then use their skills to contribute to the most pressing research problems within AI safety. Time Commitment The program will last 9 weeks, beginning on Monday, June 20th, and ending on August 19th. We expect each week of the program to cover the equivalent of about 3 weeks of the university lectures we are drawing our curriculum from. As a result, the program will likely take roughly 30-40 hours per week, depending on speed and prior knowledge. Preliminary Content & Schedule Machine Learning (content from the MIT open course) Week 1 - Basics, Perceptrons, Features Week 2 - Features continued, Margin Maximization (logistic regression and gradient descent), Regression Deep Learning (content from a University of Michigan course as well as an NYU course) Week 3 - Introduction, Image Classification, Linear Classifiers, Optimization, Neural Networks. ML Assignments due. Week 4 - Backpropagation, CNNs, CNN Architectures, Hardware and Software, Training Neural Nets I & II. DL Assignment 1 due. Week 5 - RNNs, Attention, NLP (from NYU), Hugging Face tutorial (parts 1-3), RL overview. DL Assignment 2 due. ML Safety Week 6 - Risk Management Background (e.g., accident models), Robustness (e.g., optimization pressure). DL Assignment 3 due. Week 7 - Monitoring (e.g., emergent capabilities), Alignment (e.g., honesty). Proj...

Early in his career, IEEE fellow and retired National Science Foundation program director Paul Werbos developed the neural network training algorithm known as error backpropagation, which has been foundational to the vast majority of today’s advances in artificial intelligence. Listen in as he discusses his work in this area and other topics, including his tenure with the National Science Foundation,… Source

Today we are speaking with Dr. Thomas Lux, a research scientist at Meta in Silicon Valley. In some sense, all of supervised machine learning can be framed through the lens of geometry. All training data exists as points in euclidean space, and we want to predict the value of a function at all those points. Neural networks appear to be the modus operandi these days for many domains of prediction. In that light; we might ask ourselves — what makes neural networks better than classical techniques like K nearest neighbour from a geometric perspective. Our guest today has done research on exactly that problem, trying to define error bounds for approximations in terms of directions, distances, and derivatives. The insights from Thomas's work point at why neural networks are so good at problems which everything else fails at, like image recognition. The key is in their ability to ignore parts of the input space, do nonlinear dimension reduction, and concentrate their approximation power on important parts of the function. [00:00:00] Intro to Show [00:04:11] Intro to Thomas (Main show kick off) [00:04:56] Interpolation of Sparse High-Dimensional Data [00:12:19] Where does one place the basis functions to partition the space, the perennial question [00:16:20] The sampling phenomenon -- where did all those dimensions come from? [00:17:40] The placement of the MLP basis functions, they are not where you think they are [00:23:15] NNs only extrapolate when given explicit priors to do so, CNNs in the translation domain [00:25:31] Transformers extrapolate in the permutation domain [00:28:26] NN priors work by creating space junk everywhere [00:36:44] Are vector spaces the way to go? On discrete problems [00:40:23] Activation functioms [00:45:57] What can we prove about NNs? Gradients without backprop Interpolation of Sparse High-Dimensional Data [Lux] https://tchlux.github.io/papers/tchlux-2020-NUMA.pdf A Spline Theory of Deep Learning [_Balestriero_] https://proceedings.mlr.press/v80/balestriero18b.html Gradients without Backpropagation ‘22 https://arxiv.org/pdf/2202.08587.pdf

In this episode , I tried to explain the concept of Artificial Neural network with out any Maths ANN:An artificial neural network is an attempt to simulate the network of neurons that make up a human brain so that the computer will be able to learn things and make decisions in a humanlike manner. Neurons are information messengers a hidden layer is located between the input and output of the algorithm. Backpropagation is an algorithm used in artificial intelligence (AI) to fine-tune mathematical weight functions and improve the accuracy of an artificial neural network's outputs check the episode on various platform https://www.instagram.com/podcasteramit Apple :https://podcasts.apple.com/us/podcast/id1544510362 Huhopper Platform :https://hubhopper.com/podcast/tech-stories/318515 Amazon: https://music.amazon.com/podcasts/2fdb5c45-2016-459e-ba6a-3cbae5a1fa4d Spotify :https://open.spotify.com/show/2GhCrAjQuVMFYBq8GbLbwa

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Predictive Coding has been Unified with Backpropagation, published by lsusr on the LessWrong. Artificial Neural Networks (ANNs) are based around the backpropagation algorithm. The backpropagation algorithm allows you to perform gradient descent on a network of neurons. When we feed training data through an ANNs, we use the backpropagation algorithm to tell us how the weights should change. ANNs are good at inference problems. Biological Neural Networks (BNNs) are good at inference too. ANNs are built out of neurons. BNNs are built out of neurons too. It makes intuitive sense that ANNs and BNNs might be running similar algorithms. There is just one problem: BNNs are physically incapable of running the backpropagation algorithm. We do not know quite enough about biology to say it is impossible for BNNs to run the backpropagation algorithm. However, "a consensus has emerged that the brain cannot directly implement backprop, since to do so would require biologically implausible connection rules"[1]. The backpropagation algorithm has three steps. Flow information forward through a network to compute a prediction. Compute an error by comparing the prediction to a target value. Flow the error backward through the network to update the weights. The backpropagation algorithm requires information to flow forward and backward along the network. But biological neurons are one-directional. An action potential goes from the cell body down the axon to the axon terminals to another cell's dendrites. An axon potential never travels backward from a cell's terminals to its body. Hebbian theory Predictive coding is the idea that BNNs generate a mental model of their environment and then transmit only the information that deviates from this model. Predictive coding considers error and surprise to be the same thing. Hebbian theory is specific mathematical formulation of predictive coding. Predictive coding is biologically plausible. It operates locally. There are no separate prediction and training phases which must be synchronized. Most importantly, it lets you train a neural network without sending axon potentials backwards. Predictive coding is easier to implement in hardware. It is locally-defined; it parallelizes better than backpropagation; it continues to function when you cut its substrate in half. (Corpus callosotomy is used to treat epilepsy.) Digital computers break when you cut them in half. Predictive coding is something evolution could plausibly invent. Unification The paper Predictive Coding Approximates Backprop Along Arbitrary Computation Graphs[1:1] "demonstrate[s] that predictive coding converges asymptotically (and in practice rapidly) to exact backprop gradients on arbitrary computation graphs using only local learning rules." The authors have unified predictive coding and backpropagation into a single theory of neural networks. Predictive coding and backpropagation are separate hardware implementations of what is ultimately the same algorithm. There are two big implications of this. This paper permanently fuses artificial intelligence and neuroscience into a single mathematical field. This paper opens up possibilities for neuromorphic computing hardware. Thanks for listening. To help us out with The Nonlinear Library or to learn more, please visit nonlinear.org.

#deeplearning #backpropagation #simulation More and more systems are made differentiable, which means that accurate gradients of these systems' dynamics can be computed exactly. While this development has led to a lot of advances, there are also distinct situations where backpropagation can be a very bad idea. This paper characterizes a few such systems in the domain of iterated dynamical systems, often including some source of stochasticity, resulting in chaotic behavior. In these systems, it is often better to use black-box estimators for gradients than computing them exactly. OUTLINE: 0:00 - Foreword 1:15 - Intro & Overview 3:40 - Backpropagation through iterated systems 12:10 - Connection to the spectrum of the Jacobian 15:35 - The Reparameterization Trick 21:30 - Problems of reparameterization 26:35 - Example 1: Policy Learning in Simulation 33:05 - Example 2: Meta-Learning Optimizers 36:15 - Example 3: Disk packing 37:45 - Analysis of Jacobians 40:20 - What can be done? 45:40 - Just use Black-Box methods Paper: https://arxiv.org/abs/2111.05803 Abstract: Differentiable programming techniques are widely used in the community and are responsible for the machine learning renaissance of the past several decades. While these methods are powerful, they have limits. In this short report, we discuss a common chaos based failure mode which appears in a variety of differentiable circumstances, ranging from recurrent neural networks and numerical physics simulation to training learned optimizers. We trace this failure to the spectrum of the Jacobian of the system under study, and provide criteria for when a practitioner might expect this failure to spoil their differentiation based optimization algorithms. Authors: Luke Metz, C. Daniel Freeman, Samuel S. Schoenholz, Tal Kachman Links: TabNine Code Completion (Referral): http://bit.ly/tabnine-yannick YouTube: https://www.youtube.com/c/yannickilcher Twitter: https://twitter.com/ykilcher Discord: https://discord.gg/4H8xxDF BitChute: https://www.bitchute.com/channel/yann... LinkedIn: https://www.linkedin.com/in/ykilcher BiliBili: https://space.bilibili.com/2017636191 If you want to support me, the best thing to do is to share out the content :) If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this): SubscribeStar: https://www.subscribestar.com/yannick... Patreon: https://www.patreon.com/yannickilcher Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2 Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

Classics: We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples. 2014: I. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, Yoshua Bengio Generative adversarial networks, Generative model, Discriminative model, Backpropagation, Minimax, Markov chain, Multilayer perceptron, Assignment (computer science), Approximation algorithm, Experiment https://arxiv.org/pdf/1406.2661.pdf

https://astralcodexten.substack.com/p/link-unifying-predictive-coding-with   [epistemic status: I know a little about the predictive coding side of this, but almost nothing about backpropagation or the math behind the unification. I am posting this mostly as a link to people who know more.] This is a link to / ad for a great recent Less Wrong post by lsusr, Predictive Coding Has Been Unified With Backpropagation, itself about a recent paper Predictive Coding Approximates Backprop Along Arbitrary Computation Graphs. Predictive coding is the most plausible current theory of how the brain works. I’ve written about it elsewhere, especially here.

בפרק זה, נארח את אופיר יוקטן - שמתעסק בניבוי התאמה של קורות חיים. אופיר יציג כיצד הוא מתמודד עם מודלים מפלים על רקע מגדריגזעני. נעסוק בשיטות להתמודדות עם הבעיה הזו: העלמת או הינדוס פיצ'רים, אימון עם Database מאוזן, ושיטת Gradient reversal layer. קישורים: "What is Adverse Impact? And Why Measuring It Matters." 26 Mar. 2018, https://www.hirevue.com/blog/hiring/what-is-adverse-impact-and-why-measuring-it-matters. Accessed 15 Dec. 2020. "Bias in Natural Language Processing (NLP): A Dangerous But ...." 1 Sep. 2020, https://towardsdatascience.com/bias-in-natural-language-processing-nlp-a-dangerous-but-fixable-problem-7d01a12cf0f7. Accessed 15 Dec. 2020. "Adversarial Removal of Demographic Attributes from Text Data." 20 Aug. 2018, https://arxiv.org/abs/1808.06640. "Unsupervised Domain Adaptation by Backpropagation." 26 Sep. 2014, https://arxiv.org/abs/1409.7495.

We Have a Real Expert on AI and Big Data And We Take The Opportunity To Ask Him A Number of Questions About The Whole Area of AI This Week Show's Sponsors Kinsta: https://kinsta.com/ LaunchFlows: https://launchflows.com/ Matthew Renze We Interview Data Science Consultant, Author and Public Speaker Matthew Renze is a data science consultant, author, and public speaker. Over the past two decades, he’s taught over 300,000 software developers and IT professionals. He’s delivered over 100 keynotes, presentations, and workshops at conferences on every continent in the world (including Antarctica). His clients range from Fortune 500 companies to small tech startups around the globe. Matthew is a Microsoft MVP in AI, an ASPInsider, and an author for Pluralsight, Udemy, and Skillshare. He’s also an open-source software contributor. His focus includes artificial intelligence, data science, and machine learning https://matthewrenze.com/ The Main Topics For The Interview 1 - What is the main difference between human and machine learning and AI? #2 - What are Neuron Networks? #3 - How does "Weights" work connected to artificial Intelligence neuron networks? #4 - What does the term " BackPropagation" mean connected to AI? #5 - Do we fully understand the mathematics that is the basis of computer "Neuron Networks"? #6 - Did "Game Theory" has a key part in the development of modern AI? #7 - Is it true that most AI networks have about 300 neurons where the human brain has about 100,000,000,000 neurons? #8 - What are the most exciting development connected to AI? #9 - Is one of the real strengths of AI networks is speed compared to the human brain? #10 - Is it true that we don't have a clear understanding connected to how human beings learn and what is real human intelligence? #11 - Can you give some background on the noise problem that a lot of AI systems have connected to recognizing slight chances in images? Main Host Jonathan Denwood https://www.wp-tonic.com Co-Host Steven Sauder https://zipfish.io

We Have a Real Expert on AI and Big Data And We Take The Opportunity To Ask Him A Number of Questions About The Whole Area of AI This Week Show's Sponsors Kinsta: https://kinsta.com/ LaunchFlows: https://launchflows.com/ Matthew Renze We Interview Data Science Consultant, Author and Public Speaker Matthew Renze is a data science consultant, author, and public speaker. Over the past two decades, he’s taught over 300,000 software developers and IT professionals. He’s delivered over 100 keynotes, presentations, and workshops at conferences on every continent in the world (including Antarctica). His clients range from Fortune 500 companies to small tech startups around the globe. Matthew is a Microsoft MVP in AI, an ASPInsider, and an author for Pluralsight, Udemy, and Skillshare. He’s also an open-source software contributor. His focus includes artificial intelligence, data science, and machine learning https://matthewrenze.com/ The Main Topics For The Interview 1 - What is the main difference between human and machine learning and AI? #2 - What are Neuron Networks? #3 - How does "Weights" work connected to artificial Intelligence neuron networks? #4 - What does the term " BackPropagation" mean connected to AI? #5 - Do we fully understand the mathematics that is the basis of computer "Neuron Networks"? #6 - Did "Game Theory" has a key part in the development of modern AI? #7 - Is it true that most AI networks have about 300 neurons where the human brain has about 100,000,000,000 neurons? #8 - What are the most exciting development connected to AI? #9 - Is one of the real strengths of AI networks is speed compared to the human brain? #10 - Is it true that we don't have a clear understanding connected to how human beings learn and what is real human intelligence? #11 - Can you give some background on the noise problem that a lot of AI systems have connected to recognizing slight chances in images? Main Host Jonathan Denwood https://www.wp-tonic.com Co-Host Steven Sauder https://zipfish.io

Today we’re joined by Roland Memisevic, return podcast guest and Co-Founder & CEO of Twenty Billion Neurons.  We last spoke to Roland in 2018, and just earlier this year TwentyBN made a sharp pivot to a surprising use case, a companion app called Fitness Ally, an interactive, personalized fitness coach on your phone.  In our conversation with Roland, we explore the progress TwentyBN has made on their goal of training deep neural networks to understand physical movement and exercise. We also discuss how they’ve taken their research on understanding video context and awareness and applied it in their app, including how recent advancements have allowed them to deploy their neural net locally while preserving privacy, and Roland’s thoughts on the enormous opportunity that lies in the merging of language and video processing. The complete show notes for this episode can be found at twimlai.com/go/427.

Building a sophisticated AI that can evolve to fit our vast and diverse needs is a Herculean challenge. Today we speak with senior engineer Eric Steen about Automata, his experimental Elixir project that uses neuroevolution and cutting edge theory to create a multi-agent behavior tree — or really good AI in the common tongue. But before we tap into that rich topic, we talk with Eric about tech burnout, his background, and why Elixir is an excellent language for writing modern software. He then unpacks AI concepts like the need to develop backpropagation in your system, and the value of “neural diversity,” and Markov decision processes. After Eric gives his take on architecture versus design and the place of domain-driven design, we discuss Automata. A key breakthrough, Eric shares his enthusiasm for ‘novelty search,’ where machines learn from a variety of new behaviors and searches, as opposed to completing one task at a time. We touch on Automata’s progress, Eric’s long-term approach, and what his project might be used for. Near the end of our interview, we chat about CryptoWise, a collaborative analysis platform for cryptocurrency. Todd Resudek then opens with another edition of Pattern Matching, where he interviews Whatsapp engineer Michał Muskała. They talk about Michał’s career, the movies and music that he enjoys, and the projects that excite him. Tune in to hear more about both Michał and neuroevolution in AI. Key Points From This Episode: Experiencing tech burnout and challenges around algorithms rendering you redundant. Hear about Eric’s programming background and shifts in the industry. Backpropagation and using Elixir to build a neural evolutionary system. How Markov decision processes help systems choose between possible actions. Eric’s take on architecture versus design and the place of domain-driven design. Exploring Automata — Eric’s ambitious multi-agent behavior tree. The importance of neurodiversity when building AIs; they need to adapt to many needs. Novelty search; why learn through one task when you can learn through a variety of tasks at the same time? Automata’s practical applications and why Eric sees it as a long-term project. Eric shares a progress report on his work and using design processes like Sprint. What Eric would like people to use Automata for. A sense that Elixir is gaining in popularity within Silicon Valley. Eric gives an elevator-pitch for CryptoWise, a collaborative analysis platform for cryptocurrency. Todd Resudek interviews Michał Muskała on another edition of Pattern Matching. Michał shares his background and his move from Poland to London. Movies and music that Michał enjoys, and details on projects that excite him. Differences between Erlang and Elixir and why both communities would benefit from working together. Links Mentioned in Today’s Episode: SmartLogic — https://smartlogic.io/ Eric Steen on LinkedIn — https://www.linkedin.com/in/ericsteen1/ Eric Steen — https://twitter.com/thesteener Webflow — https://webflow.com/ Automata GitHub — https://github.com/upstarter/automata Automata on Slack — https://join.slack.com/t/automata-project/sharedinvite/zt-e4fqrmo4-7ujuZwzXHNCGVrZb1aVmA CryptoWise — https://www.cryptowise.ai/ Hippo Insurance — https://www.hippo.com/ Carl Hewitt — https://en.wikipedia.org/wiki/CarlHewitt Stanford University — https://www.stanford.edu/ MIT — https://web.mit.edu/ Actor Model — https://en.wikipedia.org/wiki/Actormodel Marvin Minsky — http://web.media.mit.edu/~minsky/ Tensorflex on GitHub— https://github.com/anshuman23/tensorflex Matrex on GitHub — https://github.com/versilov/matrex Handbook of Neuroevolution Through Erlang — https://www.springer.com/gp/book/9781461444626 Markov Decision Process — https://en.wikipedia.org/wiki/Markovdecisionprocess Amazon Web Services — https://aws.amazon.com/ The Little Elixir & OTP Guidebook — https://www.amazon.com/Little-Elixir-OTP-Guidebook/dp/1633430111 Elon Musk — https://www.forbes.com/profile/elon-musk/ Welcome to the Era of Deep Neuroevolution — https://eng.uber.com/deep-neuroevolution/ Kenneth O. Stanley — https://www.cs.ucf.edu/~kstanley/ Why Greatness Cannot Be Planned: The Myth of the Objective — https://www.amazon.com/Why-Greatness-Cannot-Planned-Objective/dp/3319155237/ University of Florida — https://www.ufl.edu/ Uber Air — https://www.uber.com/us/es/elevate/ Jeff Bezos — https://www.forbes.com/profile/jeff-bezos/ Sprint — https://www.thesprintbook.com/ Adobe — https://www.adobe.com/ Horde — https://www.horde.org/development/ Libcluster on GitHub — https://github.com/dsteinberg/libcluster Swift for Tensorflow — https://www.tensorflow.org/swift Triplebyte Blog — https://triplebyte.com/blog EquiTrader — https://coinmarketcap.com/currencies/equitrader/ BloXroute Labs — https://bloxroute.com/ Holochain — https://holochain.org/ Michał Muskała on GitHub — https://github.com/michalmuskala Jason on GitHub — https://github.com/michalmuskala/jason Todd Resudek on LinkedIn — https://www.linkedin.com/in/toddresudek/ Whatsapp — https://www.whatsapp.com/ CERN — https://home.cern/ Ralph Kaminski — https://ralphkaminski.com/ Jayme Edwards — https://jaymeedwards.com/ Special Guest: Eric Steen.

丽莎老师讲机器人之欲在大脑中搜寻AI方法的存在欢迎收听丽莎老师讲机器人，想要孩子参加机器人竞赛、创意编程、创客竞赛的辅导，找丽莎老师！欢迎添加微信号：153 5359 2068，或搜索钉钉群：31532843。直到现在，几乎我们听说过的每一个关于人工智能的进步，仍基于 30 年前的一篇阐述多层神经网络训练方法的论文演变而来。那就是 杰弗里·辛顿Geoffrey Hinton 在 1986 年写下的《Learning representations by back-propagation errors》。这篇论文具有重要的意义，可以说是代表着反向传播算法（Backpropagation algorithm）首次被引入到多层神经网络训练，为后来该算法的推广埋下伏笔，尤其是为人工智能在最近 10 年的发展奠定了基础，但要保持这种进步，还得清醒地面对人工智能的局限性。而作为反向传播的提出者，仍然站在反思（甚至质疑）反向传播的第一线。在 2017 年时，他还曾明确表示，“我们需要放弃掉反向传播算法，重新开辟一条新的路径。”解谜人的大脑将被包含在这种路径之中，近年来，这位 “人工神经网络之父” 频频表现出了其对脑科学研究的浓厚兴趣，并发表了一定数量的有关脑神经科学的论文。反向传播支配下的人工神经网络反向传播在 AI 领域的地位是主导性的，尤其是它在人工神经网络（Neural Networks，NNs）中发挥的作用。要理解这一概念，你可以首先把人工神经网络想象成一块有很多层的三明治。每层都有人工神经元，也就是微小的计算单元。这些神经元在兴奋时会把信号传递给相连的另一个神经元（和真正的神经元传导兴奋的方式一样）。每个神经元的兴奋程度用一个数字代表，例如 0.13 或 32.39。两个神经元的连接处也有一个重要的数字，代表多少兴奋从一个神经元传导至另一个神经元。这个数字是用来模拟人脑神经元之间的连接强度。数值越大，连接越强，从一个神经元传导至另一个神经元的兴奋度就越高。以深度神经网络最成功的应用图像识别为例，正如 HBO 的电视剧《硅谷》中就有这样一个场景：创业团队开发了一款程序，能够辨认图片中有没有热狗。要让它们发挥作用，首先需要一张图片。举一个简单的例子，让神经网络读取一张宽 100 像素、高 100 像素的黑白照片，输入层每一个模拟神经元的兴奋值就是每一个像素的明亮度。那么，在这块三明治的底层，一万个神经元（100x100）代表图片中每个像素的明亮度。然后，将这一层神经元与另一层神经元相连，假如一层上有几千个神经元，它们与另一层上的几千个神经元相连，然后一层一层以此类推。最后，这块三明治的最顶层，即输出层，只有两个神经元，一个代表“热狗”，另一个代表“不是热狗”。这个过程是为了训练神经网络在图片中有热狗时将兴奋仅传导至第一个神经元，而在图片中没有热狗时将兴奋仅传导至第二个神经元。这种训练方法就是 Hinton 开发的反向传播技术。当你刚刚创建一个神经网络时，神经元之间连接的强度是随机的。换句话说，每个连接传导的兴奋值也是随机的，就像人脑中的突触还没有完全成形。反向传播发挥的作用就是通过改变数值，在输入不变的情况下提高输出的敏感度（类似于通过负反馈校准），从而让神经网络实现特定的目标。它是实现人工神经网络中非常重要的技术，作为训练神经网络的基本算法之一，它让神经网络变得更加“智能”。现在来看，反向传播的原理其实并不复杂，但它需要大量的数据才能达到最佳效果。这也是为什么这项技术于 30 年前提出，但直至近年来数据作为最基本 “粮食” 到位之后，才在现实生活中产生巨大价值。问题在于，反向传播是 Hinton 作为计算机科学家设想出来的一种工程方法，它让机器更加智能，但这种机制真实存在于人的大脑中吗？如果要让机器朝着仿生人脑的路径实现更高层次的类人的智能，这个问题或将难以回避，也正是 Hinton 最新论文讨论的核心。人工神经网络中，反向传播试图通过使用对突触权值的微小改变来减少误差。在这篇最新研究中，尽管大脑可能不存在完全依照反向传播的概念运作，但是反向传播为理解大脑皮层如何学习提供了新的线索。已知的是，人脑通过调整神经元之间的突触连接来进行学习，不过由于皮层中的突触被嵌入到多层网络中，这使得很难确定单个突触修饰对系统行为的影响。虽然反馈连接在大脑皮层中无处不在，但很难看到它们如何传递严格的反向传播算法所需的错误信号。在这里，我们在过去和最近的发展基础上，论证了反馈连接可能会诱导神经活动，而这些神经活动的差异可以用来局部近似这些信号，从而驱动大脑中的深度网络的有效学习。最近的工作表明，与灵长目动物视觉皮层腹侧流中表征的其他模型相比，反向传播训练模型与所观察到的神经反应匹配程度更高，而且未使用反向传播训练的模型（如使用 Gabor filter 的生物启发模型，或使用非反向传播优化的网络）的性能不如使用反向传播进行优化的网络。反向传播对比之下体现的实用性和效率，至少暗示了大脑存在利用误差驱动的反馈进行学习，而这正是反向传播的核心思想。Hinton 团队将这种基于活动状态误差驱动突触变化的学习机制称为 NGRAD（neural gradient representation by activity differences）。“反向传播这一概念的引入在神经科学领域引起了轰动，它可能成为深入了解大脑皮层学习的一个开端。但反向传播与大脑皮层的相关性很快遭到质疑——部分原因是它在人工系统中未能产生优秀的表现，且具有明显的生物学不可靠性。随着更强大的计算能力、更大的数据集和一些技术改进的出现，反向传播现在可以训练多层神经网络来与人类的能力竞争。NGRAD 以一种与我们认为的生物回路运作方式一致的方式解决了反向传播的重大不可靠性”，局限性在于，虽然越来越多证据表明，使用反向传播训练的 多层网络有助解释神经数据，但关于如何在皮层中实现类反向传播的学习仍有很多疑问，例如在反向传播中，传递的误差信号不影响前向传播产生的神经元的活动状态，但是在大脑皮层，这些连接会对前馈传播产生的神经活动产生影响，大脑皮层的反馈连接不仅可以调节、激活还可以驱动活动，这比反向传播要复杂得多。在 GATIC 2017 就曾提出一个观点，即人脑的神经网络结构是进化大数据训练的结果。大脑的结构是亿万年‘优胜劣汰'的进化过程造就的，这个过程很可能存在广义的反向传播机制。“这就像深度学习中，反向传播把一个无结构的多层人工神经网络训练成为一个具有特定结构的专用网络，大脑进化也是把不断增生的没有特定结构的皮层，逐渐训练成分工明确、结构基本确定的神经网络。后天学习只是微调，而不能改变大脑的基本结构。反向传播如果发挥作用的话，也主要是在大脑先天结构的形成过程中，而不是在后天学习过程中。包括这篇文章在内，许多学者正在从个体大脑的后天学习中寻找反向传播机制，但这可能没有瞄准方向”，人类视觉系统的神经网络结构在出生时是基本确定的，例如初级视皮层 V1 到 V4 的分区和区间连接关系都是确定的，婴儿期接受真实的视觉刺激，进行突触修改，V1 以及大部分视觉皮层的突触就此固定下来，形成我们的视觉功能，注意，婴儿期只是突触修改，并没有改变 V1 到 V4 的这种基本结构，相比之下，深度学习的起点并不是这样一个先天形成的有基本结构的神经系统，只是一个多层的神经网络，可以被训练成视觉网络，也可以被训练成语言网络，因此，深度学习的训练过程，实际上是在重复大脑亿万年进化要完成的结构生成任务”。从自然环境对人脑的 “训练” 来看，反向传播是可能的训练手段。“亿万年生物进化过程的大数据就是地球环境，训练机制也丰富多样，反向传播可能是其中之一。深度学习算力再强，大数据再大，也都难以望其项背。“因此，模仿生物大脑已经训练好的神经网络结构，而不是从零开始寻找结构，才是实现更强智能的更快捷的道路”，当然，这些诸多可能性仍待计算机科学家、神经科学家们共同推动探索。尽管深度学习这一概念在诞生初期仍有神经科学的影子，但近年随着深度学习本身的快速发展，它也愈发自成一派，几乎与神经科学无关联：研究深度学习的专家们专注于提升算法，而神经科学家们探讨的问题也基本上和人工深度神经网络无关。如 Hinton 团队这样，采用深度学习中发展出的思想来研究大脑并非主流，却是希望在神经科学与现有的人工智能（尤其以深度学习技术为代表）建立更多连接。30 年前人们认为神经科学可能没什么可学的，因为从生物学角度看反向传播算法有些方面是不现实的。随着梯度学习算法在深度神经网络表现出强大的学习能力，我们开始思考大脑的高效学习非常有可能近似计算梯度”。AI 算法与人类大脑会越走越远，还是越走越近？

丽莎老师讲机器人之欲在大脑中搜寻AI方法的存在欢迎收听丽莎老师讲机器人，想要孩子参加机器人竞赛、创意编程、创客竞赛的辅导，找丽莎老师！欢迎添加微信号：153 5359 2068，或搜索钉钉群：31532843。直到现在，几乎我们听说过的每一个关于人工智能的进步，仍基于 30 年前的一篇阐述多层神经网络训练方法的论文演变而来。那就是 杰弗里·辛顿Geoffrey Hinton 在 1986 年写下的《Learning representations by back-propagation errors》。这篇论文具有重要的意义，可以说是代表着反向传播算法（Backpropagation algorithm）首次被引入到多层神经网络训练，为后来该算法的推广埋下伏笔，尤其是为人工智能在最近 10 年的发展奠定了基础，但要保持这种进步，还得清醒地面对人工智能的局限性。而作为反向传播的提出者，仍然站在反思（甚至质疑）反向传播的第一线。在 2017 年时，他还曾明确表示，“我们需要放弃掉反向传播算法，重新开辟一条新的路径。”解谜人的大脑将被包含在这种路径之中，近年来，这位 “人工神经网络之父” 频频表现出了其对脑科学研究的浓厚兴趣，并发表了一定数量的有关脑神经科学的论文。反向传播支配下的人工神经网络反向传播在 AI 领域的地位是主导性的，尤其是它在人工神经网络（Neural Networks，NNs）中发挥的作用。要理解这一概念，你可以首先把人工神经网络想象成一块有很多层的三明治。每层都有人工神经元，也就是微小的计算单元。这些神经元在兴奋时会把信号传递给相连的另一个神经元（和真正的神经元传导兴奋的方式一样）。每个神经元的兴奋程度用一个数字代表，例如 0.13 或 32.39。两个神经元的连接处也有一个重要的数字，代表多少兴奋从一个神经元传导至另一个神经元。这个数字是用来模拟人脑神经元之间的连接强度。数值越大，连接越强，从一个神经元传导至另一个神经元的兴奋度就越高。以深度神经网络最成功的应用图像识别为例，正如 HBO 的电视剧《硅谷》中就有这样一个场景：创业团队开发了一款程序，能够辨认图片中有没有热狗。要让它们发挥作用，首先需要一张图片。举一个简单的例子，让神经网络读取一张宽 100 像素、高 100 像素的黑白照片，输入层每一个模拟神经元的兴奋值就是每一个像素的明亮度。那么，在这块三明治的底层，一万个神经元（100x100）代表图片中每个像素的明亮度。然后，将这一层神经元与另一层神经元相连，假如一层上有几千个神经元，它们与另一层上的几千个神经元相连，然后一层一层以此类推。最后，这块三明治的最顶层，即输出层，只有两个神经元，一个代表“热狗”，另一个代表“不是热狗”。这个过程是为了训练神经网络在图片中有热狗时将兴奋仅传导至第一个神经元，而在图片中没有热狗时将兴奋仅传导至第二个神经元。这种训练方法就是 Hinton 开发的反向传播技术。当你刚刚创建一个神经网络时，神经元之间连接的强度是随机的。换句话说，每个连接传导的兴奋值也是随机的，就像人脑中的突触还没有完全成形。反向传播发挥的作用就是通过改变数值，在输入不变的情况下提高输出的敏感度（类似于通过负反馈校准），从而让神经网络实现特定的目标。它是实现人工神经网络中非常重要的技术，作为训练神经网络的基本算法之一，它让神经网络变得更加“智能”。现在来看，反向传播的原理其实并不复杂，但它需要大量的数据才能达到最佳效果。这也是为什么这项技术于 30 年前提出，但直至近年来数据作为最基本 “粮食” 到位之后，才在现实生活中产生巨大价值。问题在于，反向传播是 Hinton 作为计算机科学家设想出来的一种工程方法，它让机器更加智能，但这种机制真实存在于人的大脑中吗？如果要让机器朝着仿生人脑的路径实现更高层次的类人的智能，这个问题或将难以回避，也正是 Hinton 最新论文讨论的核心。人工神经网络中，反向传播试图通过使用对突触权值的微小改变来减少误差。在这篇最新研究中，尽管大脑可能不存在完全依照反向传播的概念运作，但是反向传播为理解大脑皮层如何学习提供了新的线索。已知的是，人脑通过调整神经元之间的突触连接来进行学习，不过由于皮层中的突触被嵌入到多层网络中，这使得很难确定单个突触修饰对系统行为的影响。虽然反馈连接在大脑皮层中无处不在，但很难看到它们如何传递严格的反向传播算法所需的错误信号。在这里，我们在过去和最近的发展基础上，论证了反馈连接可能会诱导神经活动，而这些神经活动的差异可以用来局部近似这些信号，从而驱动大脑中的深度网络的有效学习。最近的工作表明，与灵长目动物视觉皮层腹侧流中表征的其他模型相比，反向传播训练模型与所观察到的神经反应匹配程度更高，而且未使用反向传播训练的模型（如使用 Gabor filter 的生物启发模型，或使用非反向传播优化的网络）的性能不如使用反向传播进行优化的网络。反向传播对比之下体现的实用性和效率，至少暗示了大脑存在利用误差驱动的反馈进行学习，而这正是反向传播的核心思想。Hinton 团队将这种基于活动状态误差驱动突触变化的学习机制称为 NGRAD（neural gradient representation by activity differences）。“反向传播这一概念的引入在神经科学领域引起了轰动，它可能成为深入了解大脑皮层学习的一个开端。但反向传播与大脑皮层的相关性很快遭到质疑——部分原因是它在人工系统中未能产生优秀的表现，且具有明显的生物学不可靠性。随着更强大的计算能力、更大的数据集和一些技术改进的出现，反向传播现在可以训练多层神经网络来与人类的能力竞争。NGRAD 以一种与我们认为的生物回路运作方式一致的方式解决了反向传播的重大不可靠性”，局限性在于，虽然越来越多证据表明，使用反向传播训练的 多层网络有助解释神经数据，但关于如何在皮层中实现类反向传播的学习仍有很多疑问，例如在反向传播中，传递的误差信号不影响前向传播产生的神经元的活动状态，但是在大脑皮层，这些连接会对前馈传播产生的神经活动产生影响，大脑皮层的反馈连接不仅可以调节、激活还可以驱动活动，这比反向传播要复杂得多。在 GATIC 2017 就曾提出一个观点，即人脑的神经网络结构是进化大数据训练的结果。大脑的结构是亿万年‘优胜劣汰'的进化过程造就的，这个过程很可能存在广义的反向传播机制。“这就像深度学习中，反向传播把一个无结构的多层人工神经网络训练成为一个具有特定结构的专用网络，大脑进化也是把不断增生的没有特定结构的皮层，逐渐训练成分工明确、结构基本确定的神经网络。后天学习只是微调，而不能改变大脑的基本结构。反向传播如果发挥作用的话，也主要是在大脑先天结构的形成过程中，而不是在后天学习过程中。包括这篇文章在内，许多学者正在从个体大脑的后天学习中寻找反向传播机制，但这可能没有瞄准方向”，人类视觉系统的神经网络结构在出生时是基本确定的，例如初级视皮层 V1 到 V4 的分区和区间连接关系都是确定的，婴儿期接受真实的视觉刺激，进行突触修改，V1 以及大部分视觉皮层的突触就此固定下来，形成我们的视觉功能，注意，婴儿期只是突触修改，并没有改变 V1 到 V4 的这种基本结构，相比之下，深度学习的起点并不是这样一个先天形成的有基本结构的神经系统，只是一个多层的神经网络，可以被训练成视觉网络，也可以被训练成语言网络，因此，深度学习的训练过程，实际上是在重复大脑亿万年进化要完成的结构生成任务”。从自然环境对人脑的 “训练” 来看，反向传播是可能的训练手段。“亿万年生物进化过程的大数据就是地球环境，训练机制也丰富多样，反向传播可能是其中之一。深度学习算力再强，大数据再大，也都难以望其项背。“因此，模仿生物大脑已经训练好的神经网络结构，而不是从零开始寻找结构，才是实现更强智能的更快捷的道路”，当然，这些诸多可能性仍待计算机科学家、神经科学家们共同推动探索。尽管深度学习这一概念在诞生初期仍有神经科学的影子，但近年随着深度学习本身的快速发展，它也愈发自成一派，几乎与神经科学无关联：研究深度学习的专家们专注于提升算法，而神经科学家们探讨的问题也基本上和人工深度神经网络无关。如 Hinton 团队这样，采用深度学习中发展出的思想来研究大脑并非主流，却是希望在神经科学与现有的人工智能（尤其以深度学习技术为代表）建立更多连接。30 年前人们认为神经科学可能没什么可学的，因为从生物学角度看反向传播算法有些方面是不现实的。随着梯度学习算法在深度神经网络表现出强大的学习能力，我们开始思考大脑的高效学习非常有可能近似计算梯度”。AI 算法与人类大脑会越走越远，还是越走越近？

Stephan Ajuvo (@ajuvo) vom damals(tm) Podcast, Damon Lee von der Hochschule für Musik und Sebastian Ritterbusch trafen sich zu Gulasch-Programmiernacht 2019 des CCC-Erfakreises Entropia e.V., die wieder im ZKM und der HfG Karlsruhe stattfand. Es geht um Musik, Mathematik und wie es so dazu kam, wie es ist. Damon Lee unterrichtet seit einem Jahr an der Hochschule für Musik und befasst sich mit Musik für Film, Theater, Medien und Videospielen. Im aktuellen Semester verwendet er Unity 3D um mit räumlicher Musik und Klängen virtuelle Räume im Gaming-Umfeld umzusetzen. Auch im Forschungsprojekt Terrain wird untersucht, in wie weit räumliche Klänge eine bessere Orientierungsfähigkeit im urbanen Umfeld unterstützen können. Die Idee zu dieser Folge entstand im Nachgang zur gemeinsamen Aufnahme von Stephan und Sebastian zum Thema Rechenschieber, da die Musik, wie wir sie kennen, auch ein Rechenproblem besitzt, und man dieses an jedem Klavier wiederfinden kann. Dazu spielte Musik auch eine wichtige Rolle in der Technikgeschichte, wie beispielsweise das Theremin und das Trautonium. Die Klaviatur eines herkömmlichen Klaviers erscheint mit den weißen und schwarzen Tasten alle Töne abzubilden, die unser gewöhnliches Tonsystem mit Noten abbilden kann. Der Ursprung dieses Tonsystems entstammt aus recht einfachen physikalischen und mathematischen Eigenschaften: Wird eine Saite halbiert und im Vergleich zu zuvor in Schwingung gebracht, so verdoppelt sich die Frequenz und wir hören den einen gleichartigen höheren Ton, der im Tonsystem auch gleich benannt wird, er ist nur um eine Oktave höher. Aus einem Kammerton a' mit 440Hz ändert sich in der Tonhöhe zu a'' mit 880Hz. Neben einer Verdopplung ergibt auch eine Verdreifachung der Frequenz einen für uns Menschen angenehmen Klang. Da aber der Ton über eine Oktave höher liegt, wird dazu der wieder um eine Oktave tiefere Ton, also der Ton mit 1,5-facher Frequenz betrachtet. Dieses Tonintervall wie beispielsweise von a' mit 440Hz zu e'' mit 660Hz ist eine (reine) Quinte. Entsprechend des Quintenzirkels werden so alle 12 unterschiedlichen Halbtöne des Notensystems innerhalb einer Oktave erreicht. Nur gibt es hier ein grundsätzliches mathematisches Problem: Gemäß des Fundamentalsatzes der Arithmetik hat jede Zahl eine eindeutige Primfaktorzerlegung. Es ist also nicht möglich mit mehreren Multiplikationen mit 2 zur gleichen Zahl zu gelangen, die durch Multiplikationen mit 3 erreicht wird. Somit kann der Quintenzirkel nicht geschlossen sein, sondern ist eigentlich eine niemals endende Quintenspirale und wir müssten unendlich viele unterschiedliche Töne statt nur zwölf in einer Oktave haben. In Zahlen ist . Nach 12 reinen Quinten erreichen wir also nicht genau den ursprünglichen Ton um 7 Oktaven höher, doch der Abstand ist nicht sehr groß. Es ist grundsätzlich unmöglich ein endliches Tonsystem auf der Basis von reinen Oktaven und reinen Quinten zu erzeugen, und es wurden unterschiedliche Strategien entwickelt, mit diesem Problem zurecht zu kommen. Wird das Problem ignoriert und nur die letzte Quinte verkleinert, damit sie auf den ursprünglichen Ton um sieben Oktaven höher trifft, so entsteht eine schlimm klingende Wolfsquinte. Auch im Cello-Bau können durch Wahl der Verhältnisse der Saiten und der Schwingungsfrequenzen des Korpus fast unspielbare Töne entstehen, diese werden Wolfston genannt. In der Musik wird die erforderliche Korrektur von Intervallen auch Komma-Anpassung genannt, die beispielsweise bei Streichinstrumenten automatisch, da hier die Töne nicht auf festen Frequenzen festgelegt sind, sondern durch die Fingerposition auf dem Griffbrett individuell gespielt wird. Bei Tasteninstrumenten müssen die Töne aber im Vorfeld vollständig in ihrer Frequenz festgelegt werden, und hier haben sich historisch verschiedene Stimmungen ergeben: Nach vielen Variationen, die immer durch die Wolfsquinte unspielbare Tonarten beinhalteten, wurde ab 1681 in der Barockzeit von Andreas Werkmeister die Wohltemperierte Stimmung eingeführt, in der zwar jede Tonart spielbar, aber jeweils individuelle Stimmungen und Charaktäre vermittelten. Diese Unterschiede sollen Johann Sebastian Bach bis 1742 zum Werk Das wohltemperierte Klavier inspiriert haben, wo er die jeweiligen Eigenheiten aller Tonarten musikalisch umsetzte. Die heute am häufigsten verwendete Gleichtstufige oder Gleichmäßige Stimmung verkleinert alle Quinten statt 1,5 auf den gleichen Faktor , so dass alle Töne auf die Frequenzen festgelegt sind. Damit sind alle Tonarten absolut gleichberechtigt gut spielbar, sie klingen aber auch alle gleich, und haben alle den gleichen kleinen Fehler. Da aber gerade bei Streichinstrumenten natürlich passendere Frequenzen gewählt werden, klingen gerade synthetisch erzeugte Streicher unrealistisch, wenn sie der exakten gleichstufigen Stimmung folgen. Während bei der Klavierstimmung die Töne durch die Spannung der Saiten eingestellt werden können, so werden metallische Orgelpfeifen mechanisch mit einem Stimmeisen in ihrer Frequenz angepasst. Die Porzellanorgel ist eine ungewöhnliche unter anderem in Meissen hergestellte Form, deren Pfeifen natürlich auch mit Luft und nicht durch Vibration, wie beim Schlaginstrument des Vibraphons klingen. György Ligeti, populär bekannt durch Filmmusiken in 2001: Odyssee im Weltraum und Eyes Wide Shut, hat sich in seinem späteren Schaffenswerk auch mit exotischeren Tonsystemen auf Basis reiner Intervalle mit Streichern befasst. Beispielsweise sollte Continuum, für Cembalo, mit Mitteltöniger Stimmung gespielt werden. Um in der herkömmlichen Notation auf der Basis von 12 Halbtönen auch feinere Tonschritte bezeichnen zu können, wurden die Zeichen Halb-Kreuz und Halb-b eingeführt, die auf die Viertelton-Musik führten. Hier stellt sich die interessante Frage, ob eine Erhöhung auf 24 Tönen pro Oktave bei reinen Intervallen sich der Fehler reduziert. Diese Frage beantwortet die Berechnung des entsprechenden Faktors aus Quinten mit dem nächsten Faktor aus Oktaven und die Berechnung des relativen Fehlers, der korrigiert werden muss. Bis 53 Quinten haben folgende Kombinationen einen Fehler von weniger als 7%: Quinten n 5 7 12 17 24 29 36 41 46 48 53 Oktaven m 3 4 7 10 14 17 21 24 27 28 31 Fehler5,1%6,8%1,4%3,8%2,8%2,5%4,2%1,1%6,6%5,6%0,2% Ein sehr primitives Tonsystem kann also mit 5 Tönen aufgestellt werden, aber offensichtlich treffen 12 Töne deutlich besser. 24 Töne ermöglichen zwar mehr Tonvielfalt, verbessern aber den Fehler nicht. Erst ein Tonsystem mit 29 Tönen würde bei gleichstufiger Stimmung einen exakteren Klang als bei 12 Tönen ermöglichen. Noch besser wäre dann nur noch ein Tonsystem mit 41 Tönen pro Oktave, eine extreme Verbesserung ergibt sich bei 51 Tönen pro Oktave bei entsprechenden Problemen beim Bau einer solchen Klaviatur. Dazu haben Tonsystemerweiterungen in Vielfachen von 12 eine höhere Kompatibilität zum herkömmlichen System, und die Nähe der besseren Tonsysteme mit 29 zu 24 und 53 zu 48 zeigt, dass die Vielfachen in der Aufführung als Näherungen zu den besseren Darstellungen betrachtet werden können. Gérard Grisey (z.B. Les espaces acoustiques) und Tristan Murail sind Vertreter der Spektralisten, die in ihren Partituren erweiterte Tonsysteme verwenden. Hier sind die Tonangaben jedoch harmonisch statt melodisch gedacht, sind also in der Aufführung entsprechend zu interpretieren. YouTube: Gérard Grisey - Vortex Temporum - Ensemble Recherche Natürlich dürfen die Töne von Instrumenten nicht nur mit ihrer Grundfrequenz betrachtet werden, sondern erst das Zusammenspiel aller Harmonischen und Obertöne in Vielfachen der Grundfrequenz machen den charakteristischen Klang eines Instruments aus. Durch eine Fourier-Analyse kann mathematisch ein solches Frequenzspektrum eines Geräusches oder eines Tons berechnet werden. Oft ist hier eine überraschende Anzahl von Obertönen zu sehen, die von Menschen nicht unabhängig vom Grundton gehört werden. In der Ottoman Musik finden sich oft für west-europäische Ohren ungewohnte Harmonien, die aus ihrer langen orientalischen Geschichte andere Formen der Komposition und Tonsysteme entwickelt haben. In der Audioelektronik wurden ab etwa 1912 Röhren für Verstärker und insbesondere in der Musik verwendet, und die exakte Bauform der Bleche und Elektroden hatte deutliche Auswirkungen auf die Übertragung und Erzeugung von Spektren und Audiowellen durch Verzerrungen. Die Hammondorgel war eine sehr beliebte elektromechanische Orgel, wo anstatt von Pfeifen rotierende Zahnräder vor elektrischen Abnehmern die Töne erzeugten. Mit Hilfe von Röhren wurde in der DDR versucht, Silbermann-Orgeln als elektronische Orgeln auf Basis des Prinzips der Hammondorgel nachzubilden. Die Klangfarben der Silbermann-Orgeln wurden hier durch elektronische Rekonstruktion der Obertöne nachempfunden. Was als angenehmer Klang empfunden wird, ist eine persönliche Sache. Jedoch ist auffällig, dass der harmonische Grundklang eines Dur-Akkords einen sehr mathematischen Hintergrund hat: Die Quinte integriert den Faktor 3, bzw. 3/2, also 1.5, die große Terz den Faktor 5, bzw. 5/4 also 1.25, und die Quarte zur nächsten Oktave mit Faktor 2 ist der Faktor 4/3. Ein Zusammenspiel von so kleinen Faktoren wird bei kleinem kleinsten gemeinsamen Vielfachen wieder periodisch und ergibt einen gleichmäßigen Klang. Das persönliche Empfinden kann physiologisch mit dem Aufbau der Hörschnecke zusammenhängen, wird aber auch stark durch Erfahrungen geprägt. Musik besteht aber nicht aus einem Klang, sondern einer zeitlichen Abfolge von Konsonanz und Dissonanz, und das gilt nicht nur für neue Veröffentlichungen alter Meister von Wolfgang Rehm. So spielt Ornette Coleman mit den Erwartungen der Hörenden bis ins Chaos. YouTube: Ornette Coleman Solo - Rare! Im Google-Doodle zu Ehren von Johann Sebastian Bach hingegen versucht aus eine Vorgabe mit einem neuronalen Netz gerade die erwartete Vervollständigung im Stil von Bach zu komponieren. Eine Regelmäßigkeit oder Überraschung in der Musik kann auch im Sinne eines Informationsgehalts interpretiert werden: Sehr regelmäßige Formen sind vorhersagbar und enthalten wenig Information, die unerwartete Wendung hingegen trägt viel Information. Die als algorithmischen Komposition bezeichneten Werkzeuge werden in vielen Programmen und Geräten angeboten, beispielsweise als automatische Begleitung. Die Ergebnisse erscheinen aber nicht sehr kreativ. Bei der Verwendung von künstlichen neuronalen Netzen für die Komposition ist es leider nicht möglich im Nachhinein zu analysieren, warum und wie bestimmte Passagen erzeugt wurden: Auch wenn sie mit existierenden Beispielen mit Backpropagation trainiert wurden, arbeiten dann als Black Box, aus der nicht direkt abstrakte Entscheidungsgrundlagen reproduziert werden können. Alles Lernen setzt voraus, dass es ein Maß für die Güte gibt, was ist demnach die Qualität einer Komposition, was unterscheidet Kreativität vom Zufall und wo stimmt dies zwischen unterschiedlichen Menschen überein? Wie an prähistorischen Instrumenten zu erkennen, ist Klangerzeugung und Musik mit der Stimmbildung eng mit der Evolution des Menschen verknüpft. Recht spät entstanden Techniken zur Kodifizierung von Tonfolgen, wie beispielsweise in der Gregorianik. Es ist anzunehmen, dass der gesellschaftliche Einfluss auf die Kompositionen ihrer Zeit sehr groß war, und es jeweils auch besondere Auswirkungen wie die Blue Notes gegeben hat. Heute wird Komposition in vielen Schritten gelehrt: Angefangen von der Musiktheorie, Erlernen von Instrumenten und Musikgeschichte wird dann in Kompositionstechniken unterschiedlicher Musikepochen eingeführt. Ausgehend von den Techniken von Josquin Desprez im 15. Jahrhundert zur Verwendung des Kontrapunkt im 16. Jahrhundert, oder wie Johann Sebastian Bach den Kontrapunkt im 18. Jahrhundert nutzte. In den Notenblättern von Ludwig van Beethoven ist zu erkennen, wie er von Joseph Haydn das Komponieren auf Basis von Kontrapunkten erlernte, und auch heute mit seinen inzwischen vom Betthoven-Haus umfangreich digitalisierte Werk die Musikforschung begeistert. Ein Lehrkanon kann sich wie Kompositionstechniken über die Zeit ändern, so wie in der Mathematik früher das Riemannsche Integral Standard war, so sehen wir inzwischen den Übergang zum mächtigeren und der Wirklichkeit näheren Integralbegriff nach Lebesgue. So wie heute häufiger der neuere Begriff zum Einsatz kommt, so ist es sinnvoll und gut, auch frühere Techniken, wie auch frühere Kompositionstechniken, zu kennen und daraus lernen zu können. Im Berufsbild einer Komponistin oder eines Komponisten ist es heute meisstens nicht so, dass der Kreativität freien Lauf gelassen wird, sondern die Arbeit erfolgt in interdisziplinärer Zusammenarbeit in einem Team. Besonders für Videospielmusik oder Filmmusik wird die Komposition auf besondere Situationen hin entwickelt und erarbeitet. Wie Kreativität, Teamwork, Künstliche Intelligenz und Programmieren zu neuen Lösungen zusammenwirken kann, war auf der Gulaschprogrammiernacht auch in der Projektion der Schlangenprogrammiernacht zu sehen, wo verschiedene Programme als Schlangen in einer virtuellen Welt miteinander lebten. Der spielerische Umgang mit Algorithmen wie bei Schere, Stein, Papier führt schnell auf Spieltheorie und Herausforderungen im Hochfrequenzhandel. Literatur und weiterführende Informationen C.-Z. A. Huang, C. Hawthorne, A. Roberts, M. Dinculescu, J. Wexler, L. Hong, J. Howcroft: The Bach Doodle: Approachable music composition with machine learning at scale, ISMIR 2019. U. Peil: Die chromatische Tonleiter - Mathematik und Physik, Jahrbuch der Braunschweigischen Wissenschaftlichen Gesellschaft, 2012. M. Schönewolf: Der Wolf in der Musik. Podcasts U. Häse, S. Ajuvo: Theremin, Folge 56 im damals(tm) Podcast, 2018. N. Ranosch, G. Thäter: Klavierstimmung, Gespräch im Modellansatz Podcast, Folge 67, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2015. P. Modler, S. Ritterbusch: Raumklang, Folge 8 im Podcast Neues Terrain, 2019. R. Pollandt, S. Ajuvo, S. Ritterbusch: Rechenschieber, Gespräch im damals(tm) und Modellansatz Podcast, Folge 184, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2018. S. Ajuvo, S. Ritterbusch: Finanzen damalsTM, Gespräch im Modellansatz Podcast, Folge 97, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2016. S. Brill, T. Pritlove: Das Ohr, CRE: Technik, Kultur, Gesellschaft, Folge 206, 2014. C. Conradi: Der erste letzte Ton, Systemfehler Podcast, Folge 26, 12.4.2018. C. Conradi: Elektronische Orgel made in DDR, Zeitfragen, Deutschlandfunk Kultur, 12.6.2019. G. Follmer, H. Klein: WR051 Ortsgespräch, WRINT: Wer redet ist nicht tot, Folge 51, 2012. Audiospuren Tonbeispiele von D. Lee und S. Ritterbusch MuWi: C-g pythagoräischer Wolf, CC-BY-SA, 2007. Mdd4696: WolfTone, Public Domain, 2005. GPN19 Special P. Packmohr, S. Ritterbusch: Neural Networks, Data Science Phil, Episode 16, 2019. P. Packmohr, S. Ritterbusch: Propensity Score Matching, Gespräch im Modellansatz Podcast, Folge 207, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2019. http://modellansatz.de/propensity-score-matching C. Haupt, S. Ritterbusch: Research Software Engineering, Gespräch im Modellansatz Podcast, Folge 208, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2019. http://modellansatz.de/research-software-engineering D. Lee, S. Ajuvo, S. Ritterbusch: Tonsysteme, Gespräch im Modellansatz Podcast, Folge 216, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2019. http://modellansatz.de/tonsysteme GPN18 Special D. Gnad, S. Ritterbusch: FPGA Seitenkanäle, Gespräch im Modellansatz Podcast, Folge 177, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2018. http://modellansatz.de/fpga-seitenkanaele B. Sieker, S. Ritterbusch: Flugunfälle, Gespräch im Modellansatz Podcast, Folge 175, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2018. http://modellansatz.de/flugunfaelle A. Rick, S. Ritterbusch: Erdbebensicheres Bauen, Gespräch im Modellansatz Podcast, Folge 168, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2018. http://modellansatz.de/erdbebensicheres-bauen GPN17 Special Sibyllinische Neuigkeiten: GPN17, Folge 4 im Podcast des CCC Essen, 2017. A. Rick, S. Ritterbusch: Bézier Stabwerke, Gespräch im Modellansatz Podcast, Folge 141, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2017. http://modellansatz.de/bezier-stabwerke F. Magin, S. Ritterbusch: Automated Binary Analysis, Gespräch im Modellansatz Podcast, Folge 137, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2017. http://modellansatz.de/binary-analyis M. Lösch, S. Ritterbusch: Smart Meter Gateway, Gespräch im Modellansatz Podcast, Folge 135, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2017. http://modellansatz.de/smart-meter GPN16 Special A. Krause, S. Ritterbusch: Adiabatische Quantencomputer, Gespräch im Modellansatz Podcast Folge 105, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2016. http://modellansatz.de/adiabatische-quantencomputer S. Ajuvo, S. Ritterbusch: Finanzen damalsTM, Gespräch im Modellansatz Podcast, Folge 97, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2016. http://modellansatz.de/finanzen-damalstm M. Fürst, S. Ritterbusch: Probabilistische Robotik, Gespräch im Modellansatz Podcast, Folge 95, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2016. http://modellansatz.de/probabilistische-robotik J. Breitner, S. Ritterbusch: Incredible Proof Machine, Gespräch im Modellansatz Podcast, Folge 78, Fakultät für Mathematik, Karlsruher Institut für Technologie (KIT), 2016. http://modellansatz.de/incredible-proof-machine

Nel primo speciale Estate 2018 il prof. De Santo ci racconta la storia della ricerca in materia di Intelligenza Artificiale, oltre a qualche consiglio per il cablaggio della rete di casa. Dallo studio distribuito di digitalia: Franco Solerio, Massimo De Santo Produttori esecutivi: Simone Pignatti, Saverio Gravagnola, Nicola Pedonese, Marco Mandia, Massimiliano Saggia, Paolo Boschetti, Vito Astone, Diego Venturin, Christian A Marca, Michele Olivieri, Christian Peretto, Davide Fogliarini, Mario Cervai, Antonio Turdo (Thingyy), Federico Bruno, Stefano Negro, Alessandro Cundari, Matteo Arrighi, Roberto Barison, Daniele Barberi, Salvatore Verrusio, Andrea Plozzer, Annamaria Esposito, Massimo Dalla Motta, Paolo Sartorio, Federico Travaini, Fabio Murolo, Alessio Conforto, Roberto Viola, Alessandro Lazzarini, Davide Capra, Giuliano Arcinotti, Davide Lanza, Raffaele Viero, Christophe Sollami, Renato Battistin, Marco Barabino, Marco De Nadai, Luigi Ricco, Raffaele Marco Della Monica, Diego Arati, Luca Ubiali, Omar Nicoli, Alessandro Morgantini, Antonio Taurisano, Alex Ordiner, Marco Grechi, Massimiliano Casamento, Tommaso Saglietti, Matteo Molinari, Marco Pinna, Movida S.A., Luca Siciliano Viglieri, Cristiano Gariboldo, Marco Caggianese, Antonio Naia (Studio Grafico Padova), Stefano Toldo, Matteo Carpentieri, Paolo Lucciola, Pasquale Maffei, Riccardo Nuti, Davide Ferdinando Precone, Alberto Bravin, Gabriele Serraino, Mirko Fornai, Alessio Pappini, Andrea Dellavia, Paolo Tegoni, Sebastiano Amoddio Sponsor: Squarespace.com - utilizzate il codice coupon "DIGITALIA" per avere il 10% di sconto sul costo dell'abbonamento. Links: Backpropagation Marvin Minsky Rete neurale artificiale Percettrone Gingilli del giorno:

Backpropagation is a common algorithm for training a neural network.  It works by computing the gradient of each weight with respect to the overall error, and using stochastic gradient descent to iteratively fine tune the weights of the network.  In this episode, we compare this concept to finding a location on a map, marble maze games, and golf.

The reason that neural nets are taking over the world right now is because they can be efficiently trained with the backpropagation algorithm. In short, backprop allows you to adjust the weights of the neural net based on how good of a job the neural net is doing at classifying training examples, thereby getting better and better at making predictions. In this episode: we talk backpropagation, and how it makes it possible to train the neural nets we know and love.