Neural Network
A brain-inspired AI architecture made of layers of connected artificial neurons that learns to recognize patterns from many examples and corrections, rather than from explicit programmed rules.
As defined in bigthink_limits_of_rationalism
Other definitions
A computational system modeled after biological brains, consisting of interconnected nodes (parameters) that process information by passing numbers through weighted connections, capable of learning by adjusting these weights.
Generative AI in a Nutshell - how to survive and thrive in the age of AIA computational system inspired by biological brains, consisting of interconnected nodes (neurons) that can learn patterns from data through training
The AI Expert: "Super AI Will Be Unstoppable!" – What’s Coming NEXT | Stephen Wolfram
Where it comes up · 37
- Stephen Wolfram - Where the Computational Paradigm Leads (in Physics, Tech, AI, Biology, Math, ...)13 claims
- The AI Expert: "Super AI Will Be Unstoppable!" – What’s Coming NEXT | Stephen Wolfram11 claims
- How ChatGPT Works Technically For Beginners11 claims
- The Mastermind Behind GPT-4 and the Future of AI | Ilya Sutskever8 claims
- How LLMs Took Over The World8 claims
- The Future of Artificial Intelligence7 claims
- Ray Kurzweil (USA) at Ci2019 - The Future of Intelligence, Artificial and Natural7 claims
- The Wonder of the Emergent Mind (with Gaurav Suri)6 claims
and 29 more
Most important claims · 10 of 129Search all claims for “Neural Network” →
- A neural network (such as Rosenblatt's Perceptron) can be understood precisely as a way of computing with spaces: it takes a vector of values (a point in one space) and transforms it into another vector of values (a point in another space), and adding more layers lets it perform multi-step transformations across successive spaces to solve more complex problems—resolving how to compute when concepts are points in space.343 | Tom Griffiths on The Laws of Thought
- Marvin Minsky, who built a learning neural network as a Princeton doctoral student, abandoned the research because his calculations showed that to learn anything interesting a neural network would have to be ridiculously large; the modern reality is that people simply accepted that ridiculously large neural networks are what is needed and built the technology to create them.343 | Tom Griffiths on The Laws of Thought
- Today's deep learning systems are powerful but inscrutable black boxes that do impressive things without our understanding why, like children whose neural networks learn to speak English in ways we cannot explain; Tegmark's MIT research focuses on AI transparency—'intelligible intelligence'—transforming neural networks into more understandable systems that perform just as well, which is needed to trust everything from judicial rulings to infrastructure-controlling machines.94 - Frontiers Of Intelligence
- Meta-learning trains neural networks by manipulating their initial weights so they can learn from less data; using model-agnostic meta-learning, an outer loop adjusts the initial weights to improve performance across many learning problems (e.g., many languages) while an inner loop learns each individual problem, yielding a starting point that allows rapid learning from limited data and revealing the inductive biases human learners may need.343 | Tom Griffiths on The Laws of Thought
- Minsky and Papert proved that single-layer Perceptrons (with one layer of adjustable weights) are fundamentally limited to computing linear, local functions of their input and so cannot solve interesting problems; this dampened interest until Rumelhart, Hinton, and McClelland developed learning algorithms for multilayer networks that could solve those more complex problems, leading 40 years later to modern AI mainly through scale.343 | Tom Griffiths on The Laws of Thought
- Transformers addressed the fundamental limitations of recurrent neural networks in learning long-term dependencies, and this technical breakthrough immediately made clear that GPT development should switch to transformers as the core architecture.The Mastermind Behind GPT-4 and the Future of AI | Ilya Sutskever
- Frank Rosenblatt's Mark 1 perceptron neural network in 1957 led the Navy to fund neural network research, and the New York Times reported in 1958 that the Navy revealed 'the embryo of an electronic computer today that expects will be able to walk talk see write reproduce itself and be conscious of its existence.'The Future of Artificial Intelligence
- At OpenAI from the earliest days, the fundamental hypothesis was that predicting the next token (word, pixel, etc.) is sufficient for solving unsupervised learning, and prediction is equivalent to compression of data.The Mastermind Behind GPT-4 and the Future of AI | Ilya Sutskever
- In a 1990s course, Michael Jordan argued neural nets were just special, useless versions of regression while Geoff Hinton insisted he only needed more processing power; in hindsight Hinton was more right, as scaling compute vindicated neural networks.138 | Daryl Morey on Analytics, Psychology, and Basketball
- Unlike traditional algorithms which are scrutable and transparent, neural networks are trained by mathematically reconfiguring connections to produce good outputs—analogous to alien neuroscientists rewiring a child's brain to pass a math test—so even though we can make them do impressive things, our ability to understand why they work lags far behind, leaving us epistemically unable to predict their behavior.Erik Hoel on the Threat to Humanity from AI