Geoffrey Hinton | Collaboration with Ilya, problem-solving, and the impact of intuition on AI

The power of intuition, collaboration, and the development of AI algorithms inspired by neuroscience can lead to the potential for AI to surpass human knowledge and understanding

Questions to inspire discussion 

• What inspired the development of AI algorithms?

  —Geoffrey Hinton drew inspiration from neuroscience for developing algorithms for AI.

• What was the significance of collaborations with Terry Sinowski and Peter Brown?

  —Collaborations with them were significant in exploring the workings of the brain and developing technical results.

• How did Ilya's collaboration with Geoffrey Hinton impact their work?

  —Ilya's strong intuition and collaborative working relationship with Hinton had a significant impact on their work.

• What is the potential for AI to surpass human knowledge?

  —AI has the potential to be even more creative and progress beyond current human knowledge.

• What is the speaker's primary motivation for their research?

  —The speaker's primary motivation is driven by curiosity and the desire to understand how the brain learns.

Key Insights 

Evolution of AI and Neural Nets

• 🧠 The disappointment in the education system's lack of understanding of how the brain and mind work led to a shift towards AI and neural nets.
• 🧠 The brain learns by modifying connections in a neural net, not by using logical rules of inference - Geoffrey Hinton's intuition led him to this belief early on.
• 🧠 I learned more from him than he learned from me, that's the kind of student you want.
• 🤔 "It turns out I was basically right. New ideas help things like Transformers helped a lot but it was really the scale of the data and the scale of the computation."
• 🧠 The work and knowledge lie in the vectors used and how their elements interact, not in symbolic rules, suggesting a more plausible model of human thought.
• 🧠 The idea that stochastic gradient descent can learn big complicated things from data has been validated by big models, challenging the belief that innate knowledge or architectural restrictions are necessary.
• 🧠 Geoffrey Hinton believes that the brain may be implementing some approximate version of back propagation, which is a big open question in neuroscience.

Creativity and Intuition in AI

• 🧠 The power of language models like GPT-4 lies in their ability to find common structure and analogies between apparently different things, leading to creativity and insight.
• 🎨 AI could be even more creative than people, progressing beyond human knowledge and current level of science.
• 🧠 As we scale up these models, they get better at reasoning and use reasoning to correct their intuitions, much like how humans use reasoning to correct their intuitions.
• 🤔 People with better intuition don't stand for nonsense and have a whole framework for understanding reality, rejecting information that doesn't fit into their framework.

#SyntheticMinds #Neuroscience #AI

XMentions: @sanalabs @HabitatsDigital

Clips 

• 00:00 🧠 Geoffrey Hinton reflects on his intuitive approach to working with Ilya, their collaboration, and the inspiration drawn from neuroscience for developing AI algorithms.
  • Geoffrey Hinton reflects on how he selected talent, mentioning his intuitive approach to working with Ilya and his experience at Carnegie Mellon.
  • The speaker discusses their early experiences with programming and disappointment with the education on brain function, leading to their interest in AI.
  • Geoffrey Hinton discusses his early intuition about the brain's learning process and the inspiration he drew from neuroscience for developing algorithms for AI.
  • Collaborations with Terry Sinowski and Peter Brown were significant in exploring the workings of the brain and developing technical results, but ultimately did not align with how the brain actually works.
  • Ilya taught Geoffrey about hidden Markov models, which inspired the name "hidden layers" in neural nets, and Geoffrey learned a lot from Ilya about speech.
  • Ilya approached Geoffrey Hinton for a lab position, showed strong intuition, and had a collaborative and enjoyable working relationship with Hinton.
• 06:39 🧠 Geoffrey Hinton discusses the importance of intuition, scale of data and computation, and the potential for AI to surpass human knowledge through examples like AlphaGo.
  • Geoffrey Hinton recalls a time when Ilya wanted to create an interface for Matlab, but Hinton advised against it to avoid getting diverted from their project.
  • Over the years, the biggest shift in skill was not just in algorithms, but also in the intuition and the scale of data and computation, with the idea that making things bigger would work better.
  • Training models to predict the next word using embeddings and back propagation was initially met with skepticism, but eventually proved to be effective in language modeling.
  • Predicting the next symbol in language requires understanding and reasoning, which can be achieved through making the model predict the next symbol.
  • The brain learns by predicting and finding common structure, allowing models to encode information more efficiently and generate creativity by seeing analogies between different things.
  • The idea that AI is just regurgitating human knowledge is wrong, as it has the potential to be even more creative and progress beyond current human knowledge, as seen in examples like AlphaGo making brilliant moves in a limited domain through reinforcement learning and self-play.
• 14:02 🧠 Multimodal models and the power of intuition in neural networks lead to improved reasoning and understanding of spatial concepts, challenging the old-fashioned symbolic view of cognition.
  • Training a neural network with half of the answers wrong can still result in significant improvement, demonstrating the power of intuition and the ability for neural nets to perform better than their training data.
  • As models scale up, they improve at reasoning and can train their intuitions through reasoning, leading to more accurate predictions and creative solutions.
  • Multimodality, including images, video, and sound, will greatly improve models' ability to understand spatial concepts beyond what humans can comprehend.
  • Multimodal models are more effective at understanding objects and reasoning about space, and it is easier to learn from them than from language alone.
  • Language and cognition have a complex relationship, with the old-fashioned symbolic view suggesting cognition consists of strings of symbols in a logical language.
  • Language understanding involves converting symbols into rich embeddings, where the interaction of these vectors predicts the next symbol, representing a more plausible model of human thought.
• 21:27 🧠 Neural networks and GPUs have revolutionized machine learning, with the potential for AI to simulate human consciousness more effectively.
  • In 2006, Geoffrey Hinton's former graduate student suggested using GPUs for training neural nets, which led to a significant increase in speed, and in 2009, Hinton encouraged machine learning researchers to buy Nvidia GPUs, eventually receiving a free one from Jensen.
  • The speaker discusses the evolution of GPUs and the potential for analog computation to run big language models in low power hardware.
  • Our brains are unique and mortal, while digital systems can efficiently share and store knowledge through shared weights.
  • Neural networks need to catch up with neuroscience in terms of time scales for changes, particularly in the implementation of fast weight changes for temporary knowledge retention.
  • The efficiency of processing data in parallel using fast weights and the validation of stochastic gradient descent in learning complicated things have impacted the way we think about neural networks and the brain.
  • Innate structure is not necessary for learning, and the idea that complex language is wired in from birth is nonsense, with the potential for AI to simulate human consciousness more effectively.
• 29:28 🤖 Artificial intelligence can have feelings, weak analogies influence the speaker's life, and symbol processing is not nonsense as it involves embedding vectors and interactions between components.
  • The speaker discusses the idea that artificial intelligence can have feelings and gives an example of a robot showing emotion.
  • The speaker discusses the influence of weak analogies, such as the analogy between religious belief and belief in symbol processing, on his life.
  • Symbol processing is not nonsense, as we actually do it by giving embedding vectors to symbols and using the interactions between the components of these vectors to do thinking.
• 32:44 🧠 Geoffrey Hinton emphasizes the importance of collaboration, choosing unconventional problems, and the power of intuition in AI research, with a focus on understanding how the brain learns and the potential societal impacts of AI.
  • Geoffrey Hinton discusses the importance of collaborating with students and choosing problems that go against the consensus in the field.
  • Adding noise to a neural net can actually improve its generalization, as demonstrated through computer simulation.
  • Find something suspicious, work on it, and demonstrate why it's wrong; prioritize working on reasoning time scales as the most important problem in the field.
  • The speaker discusses the question of whether the brain does back propagation and expresses interest in researching how the brain gets gradients.
  • The speaker's primary motivation is driven by curiosity and the desire to understand how the brain learns, with the realization that the research could have both positive and negative effects on society.
  • Healthcare and engineering are promising applications for AI, but there is concern about bad actors using AI for negative purposes, and slowing down the field could also slow down the positives.
• 39:56 🧠 Trust your intuition, focus on big models and diverse ideas in AI research, and value a variety of graduate students in the lab to efficiently develop learning algorithms and achieve human-level intelligence.
  • Having assistants in AI research will make the process more efficient, and selecting talent is often intuitive.
  • He looks for students who are both technically strong and creative, but also values a variety of different kinds of graduate students in the lab.
  • Having a strong framework for understanding reality and rejecting information that doesn't fit into that framework is key to developing intuition.
  • Trust your intuitions, focus on big models and training them on multimodal data, and diversify ideas in research.
  • The speaker discusses the importance of learning algorithms and the success of backpropagation in achieving human-level intelligence.
• 45:05 🧠 The learning algorithm for Boltz machines is the thing I enjoyed most developing and what I'm proudest of, even if it's wrong.

-------------------------------------

Duration: 0:45:46

Publication Date: 2024-06-02T09:39:22Z

WatchUrl: https://www.youtube.com/watch?v=n4IQOBka8bc

-------------------------------------