Jürgen Schmidhuber
@SchmidhuberAI • 201,750 subscribers
Introduced basics of: P & T in ChatGPT, very deep learning, meta learning, neural distillation, GANs, etc. Co-authored most-cited AI paper of 20th century
Shorts
Videos
![Fukushima's video (1986) shows a CNN that recognises handwritten digits [3], three years before LeCun's video (1989). CNN timeline taken from [5]: ★ 1969: Kunihiko Fukushima published rectified linear units or ReLUs [1] which are now extensively used in CNNs. ★ 1979: Fukushima published the basic CNN architecture with convolution layers and downsampling layers [2]. He called it neocognitron. It was trained by unsupervised learning rules. Compute was 100 times more expensive than in 1989, and a billion times more expensive than today. ★ 1986: Fukushima's video on recognising hand-written digits [3]. ★ 1988: Wei Zhang et al had the first "modern" 2-dimensional CNN trained by backpropagation, and also applied it to character recognition [4]. Compute was about 10 million times more expensive than today. ★ 1989-: later work by others [5]. REFERENCES (more in [5]) [1] K. Fukushima (1969). Visual feature extraction by a multilayered network of analog threshold elements. IEEE Transactions on Systems Science and Cybernetics. 5 (4): 322-333. This work introduced rectified linear units or ReLUs, now widely used in CNNs and other neural nets. [2] K. Fukushima (1979). Neural network model for a mechanism of pattern recognition unaffected by shift in position—Neocognitron. Trans. IECE, vol. J62-A, no. 10, pp. 658-665, 1979. The first deep convolutional neural network architecture, with alternating convolutional layers and downsampling layers. In Japanese. English version: 1980. [3] Movie produced by K. Fukushima, S. Miyake and T. Ito (NHK Science and Technical Research Laboratories), in 1986. YouTube: [4] W. Zhang, J. Tanida, K. Itoh, Y. Ichioka. Shift-invariant pattern recognition neural network and its optical architecture. Proc. Annual Conference of the Japan Society of Applied Physics, 1988. First "modern" backpropagation-trained 2-dimensional CNN, applied to character recognition. [5] J. Schmidhuber (AI Blog, 2025). Who invented convolutional neural networks?](https://image.24vids.com/tw-1995875626092315116/amplify_video_thumb/1995874290957246464/img/eSqWV92Ig7PvvIiH.jpg)
Fukushima's video (1986) shows a CNN that recognises handwritten digits [3], three years before LeCun's video (1989). CNN timeline taken from [5]: ★ 1969: Kunihiko Fukushima published rectified linear units or ReLUs [1] which are now extensively used in CNNs. ★ 1979: Fukushima published the basic CNN architecture with convolution layers and downsampling layers [2]. He called it neocognitron. It was trained by unsupervised learning rules. Compute was 100 times more expensive than in 1989, and a billion times more expensive than today. ★ 1986: Fukushima's video on recognising hand-written digits [3]. ★ 1988: Wei Zhang et al had the first "modern" 2-dimensional CNN trained by backpropagation, and also applied it to character recognition [4]. Compute was about 10 million times more expensive than today. ★ 1989-: later work by others [5]. REFERENCES (more in [5]) [1] K. Fukushima (1969). Visual feature extraction by a multilayered network of analog threshold elements. IEEE Transactions on Systems Science and Cybernetics. 5 (4): 322-333. This work introduced rectified linear units or ReLUs, now widely used in CNNs and other neural nets. [2] K. Fukushima (1979). Neural network model for a mechanism of pattern recognition unaffected by shift in position—Neocognitron. Trans. IECE, vol. J62-A, no. 10, pp. 658-665, 1979. The first deep convolutional neural network architecture, with alternating convolutional layers and downsampling layers. In Japanese. English version: 1980. [3] Movie produced by K. Fukushima, S. Miyake and T. Ito (NHK Science and Technical Research Laboratories), in 1986. YouTube: [4] W. Zhang, J. Tanida, K. Itoh, Y. Ichioka. Shift-invariant pattern recognition neural network and its optical architecture. Proc. Annual Conference of the Japan Society of Applied Physics, 1988. First "modern" backpropagation-trained 2-dimensional CNN, applied to character recognition. [5] J. Schmidhuber (AI Blog, 2025). Who invented convolutional neural networks?
Jürgen Schmidhuber704,247 görüntüleme • 6 ay önce
![Everybody is talking about recursive self-improvement (RSI) and meta learning. Here is my old 2020 talk about this [1]. It has aged well. Example: humans still define the starts & ends of trials of many modern meta learners. My RSI systems since 1994 LEARN to (re)define them [2]! [1] Meta Learning Machines in a Single Lifelong Trial (talk for workshops at ICML 2020 and NeurIPS 2021, based on earlier talks since 1994). Abstract: the most widely used machine learning algorithms were designed by humans and thus are hindered by our cognitive biases and limitations. Can we also construct meta learning algorithms that can learn better learning algorithms so that our self-improving AIs have no limits other than those inherited from computability and physics? This question has been a main driver of my research since I wrote a thesis on it in 1987 [2]. Here I summarize our work on meta reinforcement learning with self-modifying policies in a single lifelong trial (since 1994), and mathematically optimal meta-learning through the self-referential Gödel Machine (since 2003). Many additional publications on meta-learning since 1987 can be found in the RSI overview [2]. [2] J. Schmidhuber (AI Blog, 2020-2025). 1/3 century anniversary of first publication on recursive self-improvement (RSI) and meta learning machines that learn to learn (1987). For its cover I drew a robot that bootstraps itself. 1992-: gradient descent-based neural meta learning. 1994-: meta reinforcement learning with self-modifying policies. 1997: meta RL plus artificial curiosity and intrinsic motivation. 2002-: asymptotically optimal meta learning for curriculum learning. 2003-: mathematically optimal Gödel Machine. 2020-: new stuff!](https://image.24vids.com/tw-2031397968511717584/amplify_video_thumb/2031397324404985860/img/FjgqUEEDb0gBsKbU.jpg)
Everybody is talking about recursive self-improvement (RSI) and meta learning. Here is my old 2020 talk about this [1]. It has aged well. Example: humans still define the starts & ends of trials of many modern meta learners. My RSI systems since 1994 LEARN to (re)define them [2]! [1] Meta Learning Machines in a Single Lifelong Trial (talk for workshops at ICML 2020 and NeurIPS 2021, based on earlier talks since 1994). Abstract: the most widely used machine learning algorithms were designed by humans and thus are hindered by our cognitive biases and limitations. Can we also construct meta learning algorithms that can learn better learning algorithms so that our self-improving AIs have no limits other than those inherited from computability and physics? This question has been a main driver of my research since I wrote a thesis on it in 1987 [2]. Here I summarize our work on meta reinforcement learning with self-modifying policies in a single lifelong trial (since 1994), and mathematically optimal meta-learning through the self-referential Gödel Machine (since 2003). Many additional publications on meta-learning since 1987 can be found in the RSI overview [2]. [2] J. Schmidhuber (AI Blog, 2020-2025). 1/3 century anniversary of first publication on recursive self-improvement (RSI) and meta learning machines that learn to learn (1987). For its cover I drew a robot that bootstraps itself. 1992-: gradient descent-based neural meta learning. 1994-: meta reinforcement learning with self-modifying policies. 1997: meta RL plus artificial curiosity and intrinsic motivation. 2002-: asymptotically optimal meta learning for curriculum learning. 2003-: mathematically optimal Gödel Machine. 2020-: new stuff!
Jürgen Schmidhuber195,893 görüntüleme • 3 ay önce
Video of my opening keynote for the 2026 World Modeling Workshop at Mila, Quebec AI Institute: simple but powerful ways of using world models and their latent space (sorry for sound problem 0:20-0:43). Details and references in: The Neural World Model Boom
Jürgen Schmidhuber112,409 görüntüleme • 1 ay önce
In 2016, at an AI conference in NYC, I explained artificial consciousness, world models, predictive coding, and science as data compression in less than 10 minutes. I happened to be in town, walked in without being announced, and ended up on their panel. It was great fun. Organizer: David Chalmers Distinguished panel members: DanielKahneman Susan Schneider Gary Marcus Jaan Tallin Original YouTube video: 29:40-39:15 Two relevant pages from the AI Blog: 1990: planning & reinforcement learning with recurrent world models and artificial curiosity. This yields a simple explanation of consciousness and self-awareness: 1991: first very deep learning with self-supervised pre-training. A conscious chunker recurrent neural net (RNN) attends to unexpected events that surprise a lower-level subconscious automatiser RNN. The automatiser uses neural network distillation to compress and absorb the formerly conscious insights and behaviours of the chunker, thus making them subconscious: Interview: 2016: J. Carmichael. Artificial Intelligence Gained Consciousness in 1991. Why A.I. pioneer Jürgen Schmidhuber is convinced the ultimate breakthrough already happened. Inverse, 2016.
Jürgen Schmidhuber277,894 görüntüleme • 2 yıl önce
![AGI? One day, but not yet. The only AI that works well right now is the one behind the screen [12-17]. But passing the Turing Test [9] behind a screen is easy compared to Real AI for real robots in the real world. No current AI-driven robot could be certified as a plumber [13-17]. Hence, the Turing Test isn't a good measure of intelligence (and neither is IQ). And AGI without mastery of the physical world is no AGI. That’s why I created the TUM CogBotLab for learning robots in 2004 [5], co-founded a company for AI in the physical world in 2014 [6], and had teams at TUM, IDSIA, and now KAUST work towards baby robots [4,10-11,18]. Such soft robots don't just slavishly imitate humans and they don't work by just downloading the web like LLMs/VLMs. No. Instead, they exploit the principles of Artificial Curiosity to improve their neural World Models (two terms I used back in 1990 [1-4]). These robots work with lots of sensors, but only weak actuators, such that they cannot easily harm themselves [18] when they collect useful data by devising and running their own self-invented experiments. Remarkably, since the 1970s, many have made fun of my old goal to build a self-improving AGI smarter than myself and then retire. Recently, however, many have finally started to take this seriously, and now some of them are suddenly TOO optimistic. These people are often blissfully unaware of the remaining challenges we have to solve to achieve Real AI. My 2024 TED talk [15] summarises some of that. REFERENCES (easy to find on the web): [1] J. Schmidhuber. Making the world differentiable: On using fully recurrent self-supervised neural networks (NNs) for dynamic reinforcement learning and planning in non-stationary environments. TR FKI-126-90, TUM, Feb 1990, revised Nov 1990. This paper also introduced artificial curiosity and intrinsic motivation through generative adversarial networks where a generator NN is fighting a predictor NN in a minimax game. [2] J. S. A possibility for implementing curiosity and boredom in model-building neural controllers. In J. A. Meyer and S. W. Wilson, editors, Proc. of the International Conference on Simulation of Adaptive Behavior: From Animals to Animats, pages 222-227. MIT Press/Bradford Books, 1991. Based on [1]. [3] J.S. AI Blog (2020). 1990: Planning & Reinforcement Learning with Recurrent World Models and Artificial Curiosity. Summarising aspects of [1][2] and lots of later papers including [7][8]. [4] J.S. AI Blog (2021): Artificial Curiosity & Creativity Since 1990. Summarising aspects of [1][2] and lots of later papers including [7][8]. [5] J.S. TU Munich CogBotLab for learning robots (2004-2009) [6] NNAISENSE, founded in 2014, for AI in the physical world [7] J.S. (2015). On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning (RL) Controllers and Recurrent Neural World Models. arXiv 1210.0118. Sec. 5.3 describes an RL prompt engineer which learns to query its model for abstract reasoning and planning and decision making. Today this is called "chain of thought." [8] J.S. (2018). One Big Net For Everything. arXiv 1802.08864. See also patent US11853886B2 and my DeepSeek tweet: DeepSeek uses elements of the 2015 reinforcement learning prompt engineer [7] and its 2018 refinement [8] which collapses the RL machine and world model of [7] into a single net. This uses my neural net distillation procedure of 1991: a distilled chain of thought system. [9] J.S. Turing Oversold. It's not Turing's fault, though. AI Blog (2021, was #1 on Hacker News) [10] J.S. Intelligente Roboter werden vom Leben fasziniert sein. (Intelligent robots will be fascinated by life.) F.A.Z., 2015 [11] J.S. at Falling Walls: The Past, Present and Future of Artificial Intelligence. Scientific American, Observations, 2017. [12] J.S. KI ist eine Riesenchance für Deutschland. (AI is a huge chance for Germany.) F.A.Z., 2018 [13] H. Jones. J.S. Says His Life's Work Won't Lead To Dystopia. Forbes Magazine, 2023. [14] Interview with J.S. Jazzyear, Shanghai, 2024. [15] J.S. TED talk at TED AI Vienna (2024): Why 2042 will be a big year for AI. See the attached video clip. [16] J.S. Baut den KI-gesteuerten Allzweckroboter! (Build the AI-controlled all-purpose robot!) F.A.Z., 2024 [17] J.S. 1995-2025: The Decline of Germany & Japan vs US & China. Can All-Purpose Robots Fuel a Comeback? AI Blog, Jan 2025, based on [16]. [18] M. Alhakami, D. R. Ashley, J. Dunham, Y. Dai, F. Faccio, E. Feron, J. Schmidhuber. Towards an Extremely Robust Baby Robot With Rich Interaction Ability for Advanced Machine Learning Algorithms. Preprint arxiv 2404.08093, 2024.](https://image.24vids.com/tw-1925204946967801861/amplify_video_thumb/1925197727224250368/img/U-5TNxCzlWIlxyYf.jpg)
AGI? One day, but not yet. The only AI that works well right now is the one behind the screen [12-17]. But passing the Turing Test [9] behind a screen is easy compared to Real AI for real robots in the real world. No current AI-driven robot could be certified as a plumber [13-17]. Hence, the Turing Test isn't a good measure of intelligence (and neither is IQ). And AGI without mastery of the physical world is no AGI. That’s why I created the TUM CogBotLab for learning robots in 2004 [5], co-founded a company for AI in the physical world in 2014 [6], and had teams at TUM, IDSIA, and now KAUST work towards baby robots [4,10-11,18]. Such soft robots don't just slavishly imitate humans and they don't work by just downloading the web like LLMs/VLMs. No. Instead, they exploit the principles of Artificial Curiosity to improve their neural World Models (two terms I used back in 1990 [1-4]). These robots work with lots of sensors, but only weak actuators, such that they cannot easily harm themselves [18] when they collect useful data by devising and running their own self-invented experiments. Remarkably, since the 1970s, many have made fun of my old goal to build a self-improving AGI smarter than myself and then retire. Recently, however, many have finally started to take this seriously, and now some of them are suddenly TOO optimistic. These people are often blissfully unaware of the remaining challenges we have to solve to achieve Real AI. My 2024 TED talk [15] summarises some of that. REFERENCES (easy to find on the web): [1] J. Schmidhuber. Making the world differentiable: On using fully recurrent self-supervised neural networks (NNs) for dynamic reinforcement learning and planning in non-stationary environments. TR FKI-126-90, TUM, Feb 1990, revised Nov 1990. This paper also introduced artificial curiosity and intrinsic motivation through generative adversarial networks where a generator NN is fighting a predictor NN in a minimax game. [2] J. S. A possibility for implementing curiosity and boredom in model-building neural controllers. In J. A. Meyer and S. W. Wilson, editors, Proc. of the International Conference on Simulation of Adaptive Behavior: From Animals to Animats, pages 222-227. MIT Press/Bradford Books, 1991. Based on [1]. [3] J.S. AI Blog (2020). 1990: Planning & Reinforcement Learning with Recurrent World Models and Artificial Curiosity. Summarising aspects of [1][2] and lots of later papers including [7][8]. [4] J.S. AI Blog (2021): Artificial Curiosity & Creativity Since 1990. Summarising aspects of [1][2] and lots of later papers including [7][8]. [5] J.S. TU Munich CogBotLab for learning robots (2004-2009) [6] NNAISENSE, founded in 2014, for AI in the physical world [7] J.S. (2015). On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning (RL) Controllers and Recurrent Neural World Models. arXiv 1210.0118. Sec. 5.3 describes an RL prompt engineer which learns to query its model for abstract reasoning and planning and decision making. Today this is called "chain of thought." [8] J.S. (2018). One Big Net For Everything. arXiv 1802.08864. See also patent US11853886B2 and my DeepSeek tweet: DeepSeek uses elements of the 2015 reinforcement learning prompt engineer [7] and its 2018 refinement [8] which collapses the RL machine and world model of [7] into a single net. This uses my neural net distillation procedure of 1991: a distilled chain of thought system. [9] J.S. Turing Oversold. It's not Turing's fault, though. AI Blog (2021, was #1 on Hacker News) [10] J.S. Intelligente Roboter werden vom Leben fasziniert sein. (Intelligent robots will be fascinated by life.) F.A.Z., 2015 [11] J.S. at Falling Walls: The Past, Present and Future of Artificial Intelligence. Scientific American, Observations, 2017. [12] J.S. KI ist eine Riesenchance für Deutschland. (AI is a huge chance for Germany.) F.A.Z., 2018 [13] H. Jones. J.S. Says His Life's Work Won't Lead To Dystopia. Forbes Magazine, 2023. [14] Interview with J.S. Jazzyear, Shanghai, 2024. [15] J.S. TED talk at TED AI Vienna (2024): Why 2042 will be a big year for AI. See the attached video clip. [16] J.S. Baut den KI-gesteuerten Allzweckroboter! (Build the AI-controlled all-purpose robot!) F.A.Z., 2024 [17] J.S. 1995-2025: The Decline of Germany & Japan vs US & China. Can All-Purpose Robots Fuel a Comeback? AI Blog, Jan 2025, based on [16]. [18] M. Alhakami, D. R. Ashley, J. Dunham, Y. Dai, F. Faccio, E. Feron, J. Schmidhuber. Towards an Extremely Robust Baby Robot With Rich Interaction Ability for Advanced Machine Learning Algorithms. Preprint arxiv 2404.08093, 2024.
Jürgen Schmidhuber72,331 görüntüleme • 1 yıl önce
100k people came to Barcelona for MWC in Feb 2025. The lecture hall couldn't accommodate them all, but fortunately there is a video of my keynote and the 1:1 fireside chat with David McClelland 4YFN. I started with the 1st AI pioneer of the 20th century, the Spanish engineer Torres y Quevedo, who built the 1st working chess end game player in 1914. I mentioned the first deep learning (1965) and neural AIs that set themselves their own goals (which have existed in my lab since 1990). Fast-forward to 2025: AI has greatly improved, however, the only AI that works well is AI in the virtual world behind the screen, e.g., LLMs such as ChatGPT. But LLMs are not good enough for AGI: no AGI without mastery of the real world! Future physical AIs won't learn like ChatGPT by downloading the web, no, they'll learn by creating self-invented experiments to improve their world models, like babies and scientists - see our artificial curiosity since 1990. Currently, however, the best robots are humans - a human hand is super-advanced technology, no man-made tech comes close. How long will it take to bridge this gap? My old timeline: over 13 billion years ago, the Big Bang; 13 million years ago, the first hominids; 13 thousand years ago, civilization starts. And now, within the next few years, another revolutionary period of only 13 years may begin, perhaps bringing more subjective change than the past 13 thousand years, or the past 13 million years, or the past 13 billion years. What will drive it? Not just self-replicating and self-improving software, no, self-improving hardware! AI-controlled general-purpose robots that can learn to operate all the machines and tools currently operated by humans will also be able to build/operate/repair the machines required to make more of those robots: the ultimate form of scaling - a new kind of life that can expand where biological life can't. We also talked about human augmentation and cyborgs, the tremendous commercial pressure towards "good AI," and AI for weaponry (no, we can't stop that). At the moment, the only ones who are making lots of money are those who combine hardware & AI in a way that's hard to replicate (Delvitech pitch). Nevertheless, AI is getting cheaper and cheaper, and there will be AI∀ (AI for all): all the trends are working in favor of the little guy, not in favor of the large software companies, who are becoming more like utilities, with enormous capex and greatly shrinking free cash flows. Today, 1 year later, with #MWC26 coming up, many others have started to worry about that, too :-)
Jürgen Schmidhuber19,825 görüntüleme • 4 ay önce
WEF 2019 Davos: Deutsche Bank interviewed me about AI (7 min video). Topics: metalearning & recursive self-improvement, superhuman AI, physical AI, artificial curiosity, AI's simplicity, AI's future, AI∀ (AI for all). 7 years later, I'm still saying more or less the same things🙂
Jürgen Schmidhuber18,712 görüntüleme • 4 ay önce
During the Oxford-style debate at the "Interpreting Europe Conference 2025," I persuaded many professional interpreters to reject the motion: "AI-powered interpretation will never replace human interpretation." Before the debate, the audience was 60-40 in favor of the motion; after, it was about 50-50. The full video is on Youtube. Here is a rough transcript of my speech: Thanks to Gry Gry Hasselbalch for her thoughtful presentation. She argued that AI-supported interpretation will never replace human interpretation. I can't believe that. Today's automatic translators are based on algorithmic methods that my team published at TUM in 1990-1991. Back then, compute was about 10 million times more expensive than it is today, and there wasn't much you could do with our methods - automatic translators at the time could do almost nothing. But since 1941, when the first general-purpose computer was completed, compute has become 10 times cheaper every 5 years. In 2011, compute was already 10,000 times cheaper than in 1991, but the automatic translators still sucked. The Chinese laughed at the automatic translations from Google and other providers. But just five years later, in November 2016, they were no longer laughing, because compute was now 100,000 times cheaper than in 1991, and that was enough: Google suddenly achieved a huge improvement in translation quality through our neural network called LSTM. A year later, everyone was doing it: in 2017, Facebook was using our LSTM to translate 4 billion messages every day. The translations were still not perfect, but now they were getting significantly better every year. Today, a decade after the breakthrough, we have gained another factor of 100, and the translations are already much better than those of 2016. They are certainly better than my own. That's why I also used an AI to translate this text :-) Of course, AI interpreters are already superhuman in some respects, as they can translate for thousands of people simultaneously in dozens of different languages. No human can do that. But when it comes to translation quality, the best human interpreters are still better than the best AIs. But for how much longer? The aforementioned trend will not stop in the coming decades, because the known physical limits of computing are still a long way off. In the next 30 years, compute will again become a million of times cheaper. Everything that seems impressive to us today will seem trivial in retrospect. For this reason alone, no one can seriously believe that AI-supported interpretation will _never_ replace human interpretation. What about the most outstanding of all translations? My favorite example is “The Catcher in the Rye” by Nobel Prize winner J.D. Salinger. Who translated it congenially into German? It was another winner of the Nobel Prize for Literature, Heinrich Böll. One might now hope that such congenial translations are only possible for true people who really know through personal experience what it means to walk barefoot through a field and have all the profoundly human experiences that befall the protagonist of the story. But even in this area, human dominance is not assured. I suspect that the best machine translators will be those who have not only read all the world's literature, who are not just encoders and decoders of sentences, but also have experience in controlling complex humanoid agents that not only have cameras and microphones, but also millions upon millions of touch and pain sensors, and can act in the real world and in principle experience much of what humans experience, interacting with real humans, and learning to better predict their behaviors. These do not yet exist today, but at least simple variants of them have been around for a long time and the prototypes are getting better and better: artificial agents whose main purpose in life is to learn from experience, avoid pain and maximize reward, just like humans. When such agents combine their experience in the real world with the art of expressing themselves like a typical Nobel Prize winner in literature, which can already be acquired today through diligent reading, it seems absurd to assume that AI-supported high quality interpretation will never replace human interpretation. “Never” is such a bold, presumptuous word! Just a few decades ago, many believed that AI would never pass the Turing test, never play chess superhumanly well, never solve difficult math problems formulated in natural language, and so on. All this proved to be a misconception. Consider again that AI gets a million times better per euro every 30 years. It seems inevitable: soon, interpreting AIs will eclipse almost all human interpreters in every important way. Thank you. Final remark: "Unfortunately, Juergen didn't have time to participate in this debate, so he sent me, his avatar." The debate led the charming moderator Andrea Grosso to accept the possibility that AIs will eclipse human interpreters. However, he pointed out that this will take a very long time. My avatar said: "I agree that it will take a long time. This could take months, if not years.”
Jürgen Schmidhuber35,314 görüntüleme • 1 yıl önce
Daha fazla içerik yok.