MIT AGI: Building machines that see, learn, and think like people (Josh Tenenbaum)

A type of directed graphical model that represents probabilistic relationships among a set of variables, generalized by probabilistic programs.

A famous dataset of handwritten digits (0-9) used extensively in deep learning and pattern recognition research.

The understanding that objects continue to exist even when they cannot be seen, heard, or touched, demonstrated by a baby in a video and explained as a strong form of cognitive ability.

A test proposed by Alan Turing to assess a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.

An aspect of visual intelligence where AI systems generate textual descriptions for images, highlighted as an area of apparent success for pattern recognition-based AI but also revealing its limitations.

A type of machine learning that focuses on how intelligent agents should take actions in an environment to maximize cumulative reward.

A concept suggesting that if knowledge is embodied as a program, then learning involves programs that build or modify other programs, analogous to children hacking on their mental code.

A popular metaphor in cognitive development that emphasizes children as active theory builders and their play as a form of casual experimentation.

An algorithm used in deep learning for training neural networks, with origins tracing back to psychological journals.

A model that captures how young children understand other people's actions and goals, using probabilistic programs defined over simple planning and perception programs.

A type of artificial neuron and the earliest feedforward neural network, proposed by Rosenblatt.

A type of stochastic recurrent neural network proposed by Hinton and colleagues.

The idea that infants understand a basic calculus of cost and benefit, where children take costly actions to achieve rewarding goal states, with costs measured in physical terms.

A subset of machine learning, often used in pattern recognition, which has made significant strides but is argued to be insufficient for achieving human-like general intelligence.

The idea that the brain's common-sense knowledge representations are analogous to programs found in modern video game engines, integrating physics engines and simple AI planning models.

A computational model built in the speaker's group that uses probabilistic inference in a game-style physics engine to simulate physical interactions and predict outcomes.

A project from the Probabilistic Computing Group at Koch Men's, aiming to converge AI tools.

A software idea developed by Improbable for very big distributed computing environments, enabling more complex and realistic simulations for immersive videogames.

A Google DeepMind AI program that plays the game of Go, highlighted as an example of narrow AI that excels at a specific task but lacks general intelligence or common sense.

One of the state-of-the-art industry AI captioning systems whose outputs are shared by the Pic Desk Bot to illustrate current AI capabilities and failures.

A computational abstraction used to capture common-sense knowledge, generalizing Bayesian networks and combining symbolic representation, probabilistic inference, and neural networks for expressive knowledge representation and flexible uncertainty handling.

A modern probabilistic programming library that integrates with PyTorch, combining neural networks with Bayesian inference.

A tool that incorporates aspects of Bayesian inference and flow-based models, likely within the deep learning context.

A standard tool in robotics for planning physically efficient movements, used in the speaker's models to simulate human reaches and actions.

A modern probabilistic programming tool that combines neural networks with Bayesian inference.

A Twitter bot that uses a state-of-the-art industry AI captioning system to caption random images from the web and upload the results, serving as a real-world test for AI limitations.

A probabilistic programming language developed in the speaker's group, built on the lambda calculus or Lisp, demonstrating a turing-complete framework for probability models.

An individual associated with MIT and founder of Singular Computing, working on brain-inspired, low-power approximate computing for highly parallel hardware.

A psychologist who, along with Felix Warneken, conducted famous experiments on the spontaneous helping behavior of young children.

Professor at MIT leading the computational cognitive science group and the speaker of this talk, focused on how humans learn efficiently and how to apply this to AI systems.

A prominent researcher in deep learning, associated with the MNIST dataset.

A colleague at CSAIL, known for work in programming languages and automatic code synthesis, collaborating with Kevin Ellis on program learning.

A leading figure in deep learning, formerly at Google and OpenAI, then Director of AI Research at Tesla; his blog post on the state of computer vision highlighted current AI limitations regarding human-like understanding.

Founder of Boston Dynamics, a leading company in humanoid and legged robots; an advisor to the CBMM.

A researcher who developed the models for a study on infants' understanding of goals, showing sensitivity to physical work.

Pioneering researchers in reinforcement learning, credited with early work on temporal difference learning.

A researcher who, along with Jess Hamrick, started the work on the intuitive physics engine in the speaker's group.

A researcher who, along with Pete Battaglia, started the work on the intuitive physics engine in the speaker's group.

Former President of the United States, featured in a popular image used by Andrej Karpathy to illustrate the gap between current computer vision and human understanding.

Proposed simple recurrent networks, an earlier and simpler version of recurrent neural network architectures.

First author of a Science paper on human-level concept learning, which gained significant publicity.

A current PhD student working with the speaker and Armando Solar-Lezama on combining programming language techniques with machine learning for program synthesis.

A psychologist who, along with Michael Tomasello, conducted famous experiments on the spontaneous helping behavior of young children.

A pioneer in AI who, like Turing, proposed the idea of building systems that grow into intelligence much like humans do.

A popular video-sharing platform where several videos (e.g., stacking cups baby, cat and cups baby, mouse vs. cracker, orangutan with Legos) are hosted and used as examples of human and animal intelligence.

A major tech company (big tech) that develops AI technologies; discussed in the context of commercial AI progress and research collaboration with academia.

A London-based startup developing a 'Spatial OS' for very large distributed computing environments, creating more complex and immersive video games.

A major tech company (big tech) that develops AI technologies.

A major industry player in computer vision and AI, specifically mentioned for its image captioning systems and researchers who produced the training datasets.

An automotive and energy company where Andrej Karpathy served as Director of AI Research.

A startup in Kendall Square founded by Joe Bates, developing technology for low-power, approximate, brain-like computing with the goal of building machines with billions of cores and reasonable power consumption.

A British artificial intelligence company acquired by Google, known for its work on AI systems like AlphaGo and a potential destination for researchers as discussed by the speaker.

A major tech company (big tech) that develops AI technologies.

A leading company known for its sophisticated humanoid and legged robots, highlighted for impressive hardware but acknowledged to have limited human-like cognition in its robots.

An AI research and deployment company where Andrej Karpathy was one of the founders.

The funding body for the Center for Brains, Minds, and Machines.

The academic institution where Andrej Karpathy received his PhD and wrote his influential blog post.

A research group led by Josh Tenenbaum at MIT, focusing on cognition and intelligence.

An NSF-funded Science and Technology Center that bridges the science and engineering of intelligence, with affiliations at MIT and Harvard.

The academic institution where Josh Tenenbaum is a professor and where the Center for Brains, Minds, and Machines is hosted.

An academic institution partnered with the Center for Brains, Minds, and Machines (CBMM).

A foundational paper that introduced the backpropagation algorithm for training multi-layer perceptrons (neural networks).

The original paper introducing the perceptron architecture and its learning algorithm.

A leading journal of theoretical and mathematical psychology, where many foundational papers for deep learning and reinforcement learning were originally published.

A general interest science journal where the backpropagation paper was published by researchers affiliated with an Institute for Cognitive Science.

A notable scientific journal where a paper by Brendan Lake and colleagues on human-level concept learning was published and featured on its cover.

A journal focused on cognitive science, also a publication venue for early AI research.