When Machines Have Ideas | Ben Vigoda

When Machines Have Ideas | Ben Vigoda | TEDxBoston

Рет қаралды 36,782

7 жыл бұрын

Astronomers shared stories through the ages, enabling them to contradict and learn from each other. They broke stories apart and put them together in better ways-readjusting and reconsidering beliefs-to present new ideas that are more compelling and more accurate. Machines can do that too. "When we build stories into machine intelligence systems, we should enable them to attach probabilities to their stories" says Ben Vigoda.
Ben Vigoda is Founder, CEO/CTO of Gamalon Machine Intelligence. Before founding Gamalon, Ben was technical co-founder and CEO of Lyric Semiconductor, a startup that created the first integrated circuits and processor architectures for statistical machine learning and signal processing. He currently serves on the DARPA Information Science and Technology (ISAT) steering committee. Ben also co-founded Design That Matters, a not-for-profit that for the past decade has helped solve engineering and design problems in under-served communities and has saved thousands of infant lives by developing low-cost, easy-to-use medical technology such as infant incubators, UV therapy, pulse oximeters, and IV drip systems that have been fielded in 20 countries. Ben completed his PhD at MIT developing circuits for implementing machine learning algorithms natively in hardware.
This talk was given at a TEDx event using the TED conference format but independently organized by a local community. Learn more at ted.com/tedx

Пікірлер: 28

@forestpepper3621 7 жыл бұрын

One component of human thought that is perhaps missing from current artificial intelligence (AI) is the "concept". Perhaps this is similar to what Mr. Vigoda refers to as a "story" in this video. For example, AI can now reliably distinguish between pictures of cats and pictures of dogs. However, this does not necessarily mean that the AI has the concept of a cat or dog; rather it has a large collection of dog and cat photos that it refers to when deciding if a new picture is a cat or dog. I don't know that humans decide how to label a picture by referring to a large collection of other pictures, although we certainly have a large store of pictures in our memories. I think that, in our mind, we judge whether an animal is a cat or dog by considering many characteristics: the animal's size, posture, activity, facial expression, etc. "Concepts", which I won't try to define explicitly, seem to form the basis for human thought. I don't think that a concept in our mind is the same as a large database of similar concrete examples of that concept, though we may initially discover a concept from our experiences. Rather, a concept exists independently from examples of that concept. Examples of concepts (I guess there is the concept of "concept", too) are things like emotions such as love and hate, the concept of movement, the concept of time, etc. Concepts form the basis of communication, and advances in science are largely based on concepts rather than on data. Also, I suspect that the notion of concept is related to the notion of "consciousness", whatever that is. So I wonder if there is any AI research concerning how to realize "concepts" within computer algorithms.

@akashpatel-cb1dq 5 жыл бұрын

That’s pretty cool concept. Bayesian programming , evaluating stories. Couldn’t be more cooler than that.Thanks for your talk.

@genet8408 7 жыл бұрын

Excellent concept idea. Just read yesterday "A team of neuroscientists at the University of Toronto in Canada has discovered a reason why we often struggle to remember small details of past experiences. Morrissey et al found that there are specific groups of neurons in the medial prefrontal cortex of a rat’s brain - the region most associated with long-term memory; these neurons develop codes to help store relevant, general information from multiple experiences while, over time, losing the more irrelevant, minor details unique to each experience." Summarizing, we are developing generalized concepts over time (experience), so we are making next step after being exposed to images and interactions. This is lacking in modern Artificial Neural Networks.

@TheNimaid 6 жыл бұрын

I really do think the "multiple stories", or as I have always said to describe my own biological neural net "multiple mental models" is a major step that has to be taken in machine learning for it to begin truly exibiting human-level intelligence and awareness. Humans think different ways on different days, and it's worked out well for our biocomputers. Having different "personalities" that "know" different things usually allows for a more complete and sound group consensus, moreso than if any one person tried to "learn" the stuff other had. Sometimes, models of reality are incompatible, but useful in different scenarios. Warning: generalized idealism ahead. Indeed, if we want a macine that can act as an "enlightened scientist", we must become enlightened ourselves, and learn from within how our minds and bodies react and interact to form what we call a "conscious experience". There are so many things happening in a human head at any time, and if you can learn to listen well enough to your own, it can give profound insights into both the nature of your own reality and how to enable a machine to experience the same thing. I swear, if DeepMind doesn't have a resident Zen monk, they should find one. Might as well reverse-engineer the hardware in your head intead of reinventing the wheel only to come to the same conclusions.

@aidavoutsas 6 жыл бұрын

The human mind evolved by predicting future events either autonomously or in small groups sharing ideas . Never before has the human mind been so interconnected at a global level answering essentially the same question of how we are going to survive as a species. In order to progress to a higher order of thinking we need to break away from our singularity by developing viral algorithms which constantly feed off and attack preconceived conclusions. Mutation has to occur in order to higher form of intelligence .

@vladimir0700 7 жыл бұрын

I'm not convinced that we're anywhere close to making an intelligent machine if it's even possible with the kind of hardware we're currently utilizing.

@bpunsky 5 жыл бұрын

The speaker from the last TED talk I watched is in the front row...

@michaelhackman3195 7 жыл бұрын

Wow, cool talk!

@aleposada10 7 жыл бұрын

What software did he use to animate the planets?

@benvigoda3950 7 жыл бұрын

Open Frameworks and Java Processing

@benvigoda285 4 жыл бұрын

@@benvigoda3950 But if I had to do it over now I would use p5.js

@lukilooser1 6 жыл бұрын

when talking about the capabilities of micro-chips i think he didn't know about moore's law ending or am i wrong? as far as i know our only shot at getting way more advanced computing power is with quantum processors

@benvigoda285 4 жыл бұрын

quantum processors *may* help, but we do not currently have any evidence that quantum effects are necessary for intelligence. As far as Moore's Law ending, transistors can't get too much smaller, but they also do not need to. To continue Moore's Law, we need to shrink the wires (the "interconnect") in chips and improve GOPS/Joule - we will likely continue progress on these even if progress is lumpy.

@lukilooser1 4 жыл бұрын

@@benvigoda285 haha thanks for the reply man

@vladimir0700 7 жыл бұрын

18:40--jump to conclusions mat--LOL

@cwinhall 7 жыл бұрын

Hi Ben, Random question that popped in to my head while listening to your fascinating talk... Is it possible for machine learning to have a bias in how it learns? Or would a bias only exist due to how the engineer coded it?

@KaplaBen 7 жыл бұрын

Because of the no free lunch theorem, the machine has to have a bias when it learns. It is probably a good idea to have the bias itself be learned. (Those ideas are often referred to as bias learning, and learning to learn)

@benvigoda285 4 жыл бұрын

Each story provides its own bias for how it interprets data. Unfortunately, we can never completely eliminate bias, all of our stories may live within a single paradigm - an overarching story that is such a pervasive prior for how we think that we have great difficulty seeing outside of it. But if we can maintain many stories and always (re)consider them, then we are essentially trying to maintain an open mind for many different ways to view the same events.

@hsjosephlee8746 5 жыл бұрын

God create human beings.....human beings ultimately get out of His control . Now, human beings creaate A.I, it's just natural that Ai will ultimately get out of human's control.

@jawortham8 7 жыл бұрын

A human brain should be simulate for $10million, not $1Billion. For $129,000 you can buy, today, a 170 TeraFlop supercomputer from NVIDIA (DGX1). The human brain should be in the 1 to 10 PetaFlop range.

@benvigoda3950 7 жыл бұрын

Yes, neat, the DGX1 has about ~100B transistors. The brain has order of magnitude ~100B neurons. If we assume 10,000 transistors per neuron, then 10,000 DGX1's would make a brain. The DGX1 costs about $100,000, so 10,000 of them would cost $1B. It might be aggressive to presume less than 10k transistors per neuron, even at ~GHz. But on the other hand, as I point out in the talk, this very well might be 10x or even 100x overkill if we carefully architect how we use the transistors. So this all lines up with the same timeline ~$B now, 20-30 years until a brain = a PC. Certainly there are big error bars on these estimates, but the exponential in Moore's Law covers a lot of mistakes. As long as Moore's Law continues more or less on the same pace, even an order of magnitude mistake in our estimates in neurons or transistors and so forth, only effects our future estimates by +/-5 or so years.

@kichigaisensei 6 жыл бұрын

Playing God. Nothing short.

@tacoscomfygames1282 7 жыл бұрын

first

@arthurrobey4945 7 жыл бұрын

Error! Bad assumption. The Quantum Erasure Experiment shows is that Reality is procedurally generated by consciousness. ( The mind makes the brain) Objective Materialism died in Copenhagen I'm 1927.