Trinity of Artificial Intelligence | Anima Anandkumar | TEDxIndianaUniversity

Summary

In this talk, the presenter explores the concept of artificial intelligence, focusing on the Trinity of AI: data, learning algorithms, and computer infrastructure. They discuss the importance of each component in creating task-oriented intelligence and highlight the rapid advancements in image classification due to the ImageNet dataset and deep learning models. The presenter also touches on the future of AI, including multi-dimensional data processing with tensors and the integration of AI with robotics to create instinctive, deliberative, and multi-agent robots.

Takeaways

• 🤖 The speaker discusses the Trinity of AI, which includes data, learning algorithms, and computer infrastructure as the three main ingredients for AI success.
• 🧠 Intelligence is defined as the ability to acquire and apply knowledge and skills, contrasting the pre-programmed actions of robots with the improvisational skills of animals like dogs.
• 📈 The importance of large datasets, such as ImageNet, is highlighted for training AI systems to recognize and categorize images effectively.
• 💾 The shift from CPU to GPU computing is identified as a key factor in the advancement of AI, with GPUs providing the necessary parallel processing power.
• 📊 Deep learning models with many layers are crucial for processing complex data like images, allowing AI to learn from examples and improve its performance.
• 📉 The speaker points out the rapid progress in image classification, noting how AI systems have reached and even surpassed human-level performance on certain datasets.
• 🔮 Future research directions include multi-dimensional data processing using tensors, which can handle data with more than two dimensions, like videos or text documents.
• 🤝 The integration of AI (the mind) with robotics (the body) is presented as a challenge and opportunity, aiming to create more instinctive, deliberative, and interactive robots.
• 🛸 An example of AI in robotics is given, where drones learn to land more effectively through data-driven learning rather than pre-programmed algorithms.
• 🌪️ The potential for AI to enhance drone performance in adverse conditions, such as different wind patterns, is explored through simulation and data collection.

Q & A

What is the Trinity of AI as discussed in the transcript?

The Trinity of AI consists of three main ingredients: data, learning algorithms, and computer infrastructure. Data provides the examples or information needed for learning, learning algorithms process the data to extract knowledge, and computer infrastructure provides the necessary computational power to perform these tasks.

Why is data considered the most important ingredient in the Trinity of AI?

Data is considered the most important ingredient because it provides the examples or observations that an AI system needs to learn from. Without sufficient data, an AI system cannot acquire the knowledge and skills required to perform tasks effectively.

How does the image classification task exemplify the success of AI?

The image classification task exemplifies the success of AI by demonstrating how AI systems can learn from a large dataset of images, such as ImageNet, to recognize and categorize new images with high accuracy. This progress was made possible by combining data, deep learning models, and powerful computer infrastructure like GPUs.

What is the significance of the ImageNet dataset in AI development?

The ImageNet dataset is significant because it provided a large and diverse collection of categorized images, which allowed AI systems to learn and improve their ability to recognize and classify images. This dataset was crucial in advancing computer vision and deep learning algorithms.

How do deep learning models contribute to AI's ability to perform tasks?

Deep learning models contribute to AI's ability to perform tasks by processing input data through multiple layers, allowing the system to learn complex patterns and features. These models can extract relevant information from vast amounts of data, enabling AI systems to perform specific tasks, such as image recognition, with high accuracy.

What is the role of GPU computing in AI?

GPU computing plays a crucial role in AI by providing the parallel processing capabilities needed to handle the massive computational demands of deep learning algorithms. GPUs enable AI systems to perform billions of operations per image in a highly efficient manner, which is essential for training and deploying large-scale AI models.

Why did the speaker mention the Boston Dynamics Atlas robot and a dog attempting a backflip?

The speaker mentioned the Boston Dynamics Atlas robot and a dog attempting a backflip to illustrate the difference between pre-programmed actions and true intelligence. While the robot executed a perfect backflip, it did so based on a pre-programmed sequence, whereas the dog, despite failing, demonstrated the ability to learn and adapt, which is a sign of intelligence.

What is the challenge with AI systems when it comes to generalizing beyond their training data?

The challenge with AI systems when it comes to generalizing beyond their training data is that they may perform well on specific datasets or tasks but struggle to adapt to new or unseen data. This limitation highlights the need for AI systems to develop more robust learning capabilities that allow them to generalize and apply their knowledge to a wider range of situations.

How does the concept of tensors relate to multi-dimensional data processing in AI?

Tensors are mathematical objects that enable processing in any number of dimensions, generalizing the concepts of vectors and matrices. In AI, tensors are used to handle and analyze multi-dimensional data, such as images, videos, and text, by capturing and processing information across different dimensions, which can lead to more effective learning and understanding of complex data structures.

What is the potential future of integrating AI 'mind' with robotics 'body'?

The potential future of integrating AI 'mind' with robotics 'body' involves creating robots that are instinctive, deliberative, multi-agent, and behavioral. This integration aims to develop robots that can react and control their reactions in a fine-grained manner, make plans and consider consequences, coordinate and work with others, and interact with humans by understanding emotions and acting as partners.

Outlines

00:00

🤖 Introduction to Artificial Intelligence

The speaker begins by introducing the topic of artificial intelligence (AI) and the concept of the 'Trinity of AI'. They discuss the importance of understanding the three main components that make AI work and how they contribute to the future of AI. The speaker contrasts the public's perception of AI, with both its potential and concerns, and poses the question of what constitutes 'intelligence'. They use two videos to illustrate the difference between programmed actions and true intelligence, highlighting that while a robot may perform a complex task like a backflip, it lacks the ability to learn and adapt, which is a characteristic of intelligent beings like dogs. The speaker emphasizes that AI, as it stands, is more about task-oriented intelligence rather than general intelligence, which would mimic human cognitive abilities.

05:03

🔍 The Trinity of AI: Data, Algorithms, and Infrastructure

The speaker delves into the 'Trinity of AI', which consists of data, learning algorithms, and computer infrastructure. They explain that for AI to learn and perform tasks, it requires a vast amount of data or examples. The speaker uses the example of image classification and the ImageNet dataset, which has been pivotal in revolutionizing computer vision by providing a large collection of categorized images. They then discuss the role of deep learning models in processing this data, highlighting the importance of layers of processing to extract knowledge from examples. Lastly, the speaker touches on the necessity of powerful computer infrastructure, like GPUs, to handle the massive computational requirements of deep learning, using NVIDIA's contributions as an example. The speaker concludes by showing how these three elements combined have led to significant progress in AI, particularly in image recognition tasks.

10:03

📈 Progress in AI: Image Classification and Beyond

The speaker continues by discussing the progress made in AI, particularly in the field of image classification, which has been a significant success story for AI technology. They mention how AI systems have surpassed human performance on the ImageNet dataset, but clarify that this does not mean AI is more intelligent than humans, as it is limited to specific tasks and datasets. The speaker then shifts the focus to the future of AI, discussing the importance of learning in multiple dimensions. They introduce the concept of tensors, which are mathematical objects that can process data across various dimensions, and explain how they can be used to improve AI algorithms. The speaker shares their research on using tensors to categorize text documents, highlighting the Amazon Comprehend tool as an example of this technology in action. They emphasize the potential for multidimensional processing to be applied to a wide range of applications.

15:05

🚀 Integrating AI with Robotics: The Future of Autonomous Systems

In the final paragraph, the speaker looks towards the future of AI, focusing on the integration of AI with robotics, or the 'body' of AI. They discuss the potential for drones to learn from data and improve their performance, such as learning to land more effectively. The speaker also mentions ongoing research at Caltech on teaching drones to fly in adverse conditions, like different wind conditions, by using AI algorithms. The speaker envisions a future where robots are instinctive, deliberative, multi-agent, and behavioral, able to interact with humans and understand emotions, suggesting a bright future for AI and robotics working in tandem.

Keywords

💡Artificial Intelligence (AI)

Artificial Intelligence refers to the simulation of human intelligence in machines that are programmed to think like humans and mimic their actions. In the video, AI is discussed as a combination of task-oriented intelligence that can learn from examples and execute specific tasks effectively. The speaker emphasizes the current state of AI, which is far from general intelligence but has made significant progress in specific areas like image classification.

💡Trinity of AI

The Trinity of AI mentioned in the video refers to the three main components necessary for AI to function effectively: data, learning algorithms, and computer infrastructure. The speaker explains that data provides the examples needed for learning, algorithms process this data to extract knowledge, and computer infrastructure provides the necessary processing power. This concept is central to understanding how AI systems are built and operate.

💡Data

Data, in the context of AI, refers to the raw input that AI systems use to learn and make decisions. The video highlights the importance of data by discussing how AI systems require vast amounts of examples to learn tasks, such as image recognition. The speaker uses the example of the ImageNet dataset, which contains millions of categorized images, to illustrate how data is crucial for training AI models.

💡Deep Learning

Deep Learning is a subset of machine learning where algorithms with multiple layers (or 'deep' neural networks) are used to model and understand data with multiple levels of abstraction. The video discusses deep learning models as a key component of AI, where these models process images through multiple layers to extract features and make predictions, such as identifying objects in images.

💡General Intelligence

General intelligence, as mentioned in the video, refers to the ability to understand and learn from a wide range of experiences and apply that knowledge to various tasks. The speaker contrasts this with the current state of AI, which is more task-specific and lacks the broad cognitive abilities of general intelligence. The video suggests that while AI has made strides in specific areas, achieving general intelligence remains a significant challenge.

💡Image Classification

Image classification is the task of labeling images with categories or tags based on their content. In the video, this is presented as one of the most visible successes of AI, where deep learning models have been trained on large datasets like ImageNet to recognize and categorize objects within images. The speaker discusses the progress made in this area and how it exemplifies the application of the Trinity of AI.

💡GPU Computing

GPU Computing refers to the use of Graphics Processing Units (GPUs) for general-purpose processing, beyond just graphics rendering. In the video, the speaker explains how traditional CPUs have reached their limits in terms of processing power for AI tasks, and GPUs have become essential for handling the massive parallel computations required by deep learning models, thus enabling large-scale AI applications.

💡Tensors

Tensors are multi-dimensional arrays of numbers that are used in various areas of mathematics and physics, and have become fundamental in deep learning for representing and processing multi-dimensional data. The video discusses how tensors are used to handle multi-dimensional data efficiently, allowing AI systems to learn from complex data structures and improve their performance in tasks such as text categorization.

💡Autonomous Systems

Autonomous systems are systems that can operate independently without human intervention. The video touches on the integration of AI with robotics to create autonomous systems, such as drones that can learn to land better or navigate in adverse conditions. The speaker's research at Caltech aims to bridge the gap between AI as the 'mind' and robotics as the 'body' to create more capable and intelligent machines.

💡Multi-agent Systems

Multi-agent systems are systems composed of multiple interacting intelligent agents. In the video, the speaker discusses the future of robotics where robots are not only instinctive and deliberative but also capable of coordinating and working together with other robots or humans, similar to how humans function in society. This concept is part of the broader vision for creating more sophisticated and interactive AI systems.

Highlights

The Trinity of AI consists of data, learning algorithms, and computer infrastructure.

Intelligence is defined as the ability to acquire and apply knowledge and skills.

Current robots, like the Boston Dynamics Atlas, lack intelligence as they are pre-programmed.

Image classification is a significant success of AI, demonstrating the Trinity of AI in action.

The ImageNet dataset has been pivotal in revolutionizing computer vision with over 14 million categorized images.

Deep learning models with many layers of processing are essential for image recognition.

GPU computing, led by NVIDIA, enables the vast amounts of parallel processing required for deep learning.

AI systems have reached human-level capabilities in image recognition on the ImageNet dataset.

AI's ability to generalize beyond specific datasets remains a significant challenge.

Research is exploring multi-dimensional data processing using mathematical objects known as tensors.

Amazon Comprehend uses tensor-based algorithms to categorize text documents without predefined topics.

The future of AI involves integrating 'mind' (intelligence) and 'body' (robotics) for more autonomous systems.

Caltech's Center for Autonomous Systems and Technology works on merging AI with robotics.

Drones can learn from data to land more efficiently, showcasing the potential of AI in robotics.

Future robots should be instinctive, deliberative, multi-agent, and behavioral to interact effectively with humans.

The talk concludes with a vision of a future where AI and robotics are seamlessly integrated.

Transcripts

[Music]

today I'm gonna talk about artificial

intelligence more specifically the

Trinity of AI what are the three main

ingredients that make AI happen and

what's the fabulous future we are headed

it when it comes to AI and I'll combine

both the academic perspective I have

from Caltech as well as the industry

perspective from Nvidia to see how there

are so many exciting opportunities for

industry academic collaborations to come

together and realize the dream of AI but

before we get there what is a I write

I'm sure everybody here in the audience

has seen AI in the news has seen both

you know amazing progress that AI has

done but also some dystopian future for

AI that people are worried about but

given that it's so much in the news the

first question we should be asking is

what is intelligence before we get to

artificial intelligence what do we even

mean by intelligence right so for this I

want to show you two videos and I want

you to think which one of the two is

more intelligent so let's see that so

the first one you see is the Boston

Dynamics Atlas robot right that's a

pretty cool backflip now what about the

next one sir game it's our best friend

it's also trying to do a backflip

so in the first case you see the robot

perfectly executing the backflip in the

second case you see it failing pretty

clumsily so what do you think is more

intelligent the robot or the dog

yeah it's the dog that's intelligent and

the robot actually has zero intelligence

the way we have it today

so why is that because how do we define

intelligence the intelligence is the

ability to acquire and apply knowledge

and skills right it was the robot doing

that no right at least the way it is

currently it was completely

pre-programmed even though the bakflip

looks so impressive it was all planned

beforehand and the program was built

into the robot on the other hand the dog

is improvising it's always observing its

environment maybe one day it'll also

learn to do the backflip better and so

the dog is intelligent while our current

robots are mostly not and so there you

see a very big gap of where we are when

it comes to building intelligence into

our robots into our systems and indeed

when we think about general intelligence

the gap is even wider so general

intelligence refers to have a human

level cognitive ability and artificial

systems are so far from it so when I

refer to AI or artificial intelligence I

usually mean some task oriented

intelligence it's the ability to execute

a specific task very well right to learn

from examples and to execute a given

task and that's what I'll show you how

we are making progress on and still the

vast challenges that lie ahead

so given that we got the definitions out

of the way let's now see what is the

Trinity of AI so for AI to be successful

I

see three important ingredients the very

first one and perhaps the most important

one is the data right so when I say data

I mean

examples you know like how can a system

learn it needs information observations

or data and so once it has it there are

learning algorithms which can process

the data and extract knowledge from

those examples but the third important

ingredient is the computer

infrastructure all this processing needs

a substrate how do we do this vast

amounts of processing so we need all

these three to come together for an AI

task to be successful let's now see an

example application and see how the

progress came about by putting this

Trinity together so the image

classification task has been perhaps the

most visible success of AI today as you

see here in the picture there is the

pool there are plans so there's a lot of

things in this image and as human beings

we find it's so natural to absorb and

reason about the entire image right but

it's really hard for a machine to do

that because it has to see lots of

variations of a swimming pool before it

realizes oh this is a swimming pool

right so it needs a lot of examples to

learn what a swimming pool looks like

and that's where we have the first

ingredient in our Trinity that is the

data how do we get lots of examples of

images for different categories so the

image net data set has been now has

revolutionized how we've done computer

vision and that's because for the first

time we had such a large collection of

categorised images this is more than 14

million images and more than thousand

categories and if you even look at one

example category here you see all the

variations of how natural images come

about for instance when you say fish

you know you just see so many different

variations of how a fish can occur in an

image and so the learning algorithm

needs to have access to all these

different variations for it to recognize

what a fish looks like in a new example

so that was the first ingredient the

data the second ingredient are the

algorithms and the models how do we take

these images and extract the information

so that an AI system can automatically

categorize images in a seamless way and

that's where we have the deep learning

models so the term deep refers to the

fact that there are many layers of

processing so you transform your input

image across these layers and in the

output you want an answer

you want to save with how much

confidence do you think there is a dog

in this image and if you have a good

algorithm and you've trained your system

well hopefully you will have a high

confidence that this particular image

has a dog in it right and so you need a

model that's highly flexible that can

learn from millions of examples and

extract the relevant information to

realize what looks like a dog

so those were the two ingredients the

third one is the computer infrastructure

so if you take these current deep

learning systems it is just not scalable

to run it on a normal computer so the

computers of yesteryear so the CPUs had

a wave of progress but now the curve of

their growth is plateauing and so the

new form of computing is known as the

GPU computing it stands for graphics

processing unit because you know it goes

back to how graphics was processed in a

parallel manner but the same form of

technology is also highly relevant for

processing a

because processing in these deep

learning layers requires more than

billion operations for each image but

they can be done in a highly parallel

manner and hence NVIDIA GPUs have been

at the forefront of enabling computation

at scale and have been a crucial

ingredient to realize this dream of

large-scale AI so now you see how all

these three ingredients came together so

we had the data we have the deep

learning model with learning algorithms

and then we had the GPU computing and

when they came together it created the

steep learning revolution we were able

to make progress in such a short length

of time so what you see is the errors of

that an AI system makes in recognizing

the image what are the categories in an

image over this image net dataset and

you see within a span of few years we

could get to human level capability on

this data set and so to have this

capability on such a large dataset in

such a short amount of time was what was

surprising to so many scientists and so

that shows that there is a lot of

promise that a I can still realize in

many other domains I do want to point

out one thing though if you see in 2015

on the image net data set the AI system

is doing better than a human being

so does this mean the AI system got more

intelligent than a human no that's not

the case right because it was it's only

doing better on only this data set and

that's always been a challenge with AI

if you have to now take it beyond to

other data sets or to other tasks so we

need the ability to generalize beyond

what it's seen in a current data and

that still remains in this case as well

so that is great we were able to make

fast progress but where do we go next so

my research looks at how to do learning

in many dimensions so what are

dimensions mean for data so if you think

of an image you have the width and the

height of an image and then you've the

colored image you have the number of

channels now if you took take a video

you also have the time dimension so when

you collect rich forms of data you have

all many dimensions coming together for

instance you may also have texts that is

connected to the video you may have

other forms of data that are all

processed together so how do we merge

all this very efficiently and process

them at scale so the idea is we don't

want to remove or throw away the

dimensions in our data if you do that we

can potentially lose information to help

you visualize this you know you have the

three-dimensional cube here and if you

just look at its projection on the wall

you do not see the complete picture

right and it's the same with data it's

if you're not processing it effectively

and not incorporating the information

from all different dimensions you may

potentially not well learn very well

what my research does is to use

mathematical objects known as tensors to

enable such multi-dimensional processing

so tensors or mathematical objects that

do processing in any dimension and

generalize the concept of vectors and

matrices that you probably see in an

undergraduate class so these may seem

like complicated mathematical objects

but you can realize them into real

algorithms for many applications so one

application I've looked at is how to

automatically categorize text documents

so if you have many millions of

documents

can you quickly say what are the topics

discussed in various documents right and

you feel don't even know what the

underlying topics are you don't have

examples of them when you're learning

it's even a harder task but we have a

tool that I built with my team at Amazon

called Amazon comprehend that precisely

does this and what enables this

algorithm was the concept of tensors so

intuitively what it's looking at are the

relationships of words in a document so

if there are words that are commonly

occurring together that's referred to a

topic you should be able to extract that

information and to do that at scale over

millions of documents requires more

thinking but that's precisely what this

algorithms are able to do they can take

this Corcoran's tensor of Triplets of

words and decompose them to obtain

topics that are present in different

documents and hence you can translate

mathematical concepts into actual

algorithms and then deploy them at scale

to make an impact on various

applications so I see a lot of

possibilities in future to take this

multi-dimensional processing into many

different applications so what's another

step for the future so I want to think

about both the mind and the body so far

I only talked about AI as the mind right

how do we create intelligence how do we

train a an AI system but there's also

the body which is the robotics so I

showed you in the beginning the Atlas

robot that currently is not intelligent

if you want to change that we have to

bring the two together and at Caltech we

have the Center for autonomous systems

and technology that aims to do that so

I'll show you one example we just

started working on which was to ask can

we

have the drones learn to land better so

instead of a pre-programmed algorithm

can the drone learn from lots of data

you know lots of sensors that it's

outfitted with and Bill learning to land

in a better way so let's see what was

the outcome so on the left you see the

draw and when there is no learning it's

a program controller and in the right

there is the learning and if you see on

the right the drone landed whereas it's

was hovering around in a pre-programmed

controller so this means you know there

is a lot of potential that learning can

provide to make drone flights more

efficient and ultimately autonomous

another project we are currently

thinking of is how can we have drones

I'll fly in adverse conditions and here

we can simulate wind conditions of all

different forms in the drone testing

laboratory at Caltech and then you

collect that data to design AI

algorithms that can be more effective in

countering such wind conditions if you

want to bring the mind at the body

together you want to build all different

forms of intelligence and meld them

together with the body so we want the

robots of your future to be instinctive

right so humans have instinct inbuilt in

us which is the ability to react and

control our reactions in a fine-grained

manner how do we do that in a robot

we're also deliberative we make plants

we look at consequences

how do we put do that in a robot it's

also multi-agent we are a society we

coordinate and work together with others

how do we get the robots to do the same

and ultimately the last one perhaps the

most difficult one is behavioral how can

we have robots that can interact

with us that can understand human

emotions ultimately be a partner with us

and that's perhaps you know gonna be a

bright new future for us thanks so much

you