AI
What fields AI cannot impact
handicraft
kreatvt
critical thought
love-empathy-and all abstract ideas..motivation
jokes
AI
Decide
Stock market trading
Military predator drones
loan
insurance
credit
Analyse/Predict
Prices of airlines/hotels
Recognizes tumour better than human
Surveillence cameras/microphones
identify face
recognize speech
translate speech at runtime
identify fraud.
consumer behaviour
habits of humans
Calendar assistant
context aware apps
Writes itself "autonomous"ly
Self-service
Bias in the dataset in identifying
Economic incentives to replace humans with AI will be beyond anything seen since industrial revolution.
Choice of classification of categories introduces a bias, and this is a danger that many are warning us about AI.
This ambiguity is compounded by the datasets that are used to train the network.
Neural networks may optimise profits at the expense of fairness.
Same issues exist with technology earlier attempts as this is not technology issue but human negatives when combined with technology may exponentialize the harm.
Even in brain conclusions are not always based upon logic and there are cognitive biases.
The only thing we need to fear is ourselves.
machines will replace humans at automation tasks in industries as they are faster,cheaper,better. Manufacturing Enterprises will be wiped out of employees.
Jobs that disappear:
driver,cook,motel,house builders,accountants,assembly line workers,warehouse opertors, security guards,QC inspectors, truckers,stock analysts,paralegals, radiologists,farmers
Industries disrupted: Trucking, Manufacturing, Retail,Insurance.
The AI makers thought process will succeed and the naysayers will not have an option but to accept, unless we participare in the AI from the ground up and influence how it shapes.
Humans dont learn from instructions, they practice and experience and learn from that.
SPAM machine learning:
human acts on a task
program watches the action.
program learns from watching.
program predicts new task outcome.
program acts on new task.
human evaluates program performance.
program learns from evaluation.
program improves its learning
program acts applying the new learnings.
In supervised learning we are given a data set and already know what the correct output would look like.
Supervised: regression, classification
In unsupervised learning, only data set is given.
Clustering: google news
cocktail party problem algorithm: Separate 2 users voices whose output is combined in a audio recording.
training set (x(i),y(i))
learning algorithm - h , h(x)
cost function - squared error function
cost function is used to check the performance of the hypothesis.
algorithm that automatically finds the function that minimizes the cost function
This algorithm is called gradient descent
Inputs needed for a good ai: huge data,strong algorithm, narrow domain,concrete goal.
AI Machines radically empower some human beings and violently displace some other human beings.
Machines ability to "LEARN"
reinforcement learning, transfer learning
Implementation or Application of existing research into domains is the cause of hype, rather than research brekthroughs.
Broad availability of huge amounts of data is the cause of hype, compared to earlier lockedin systems.
China Advantage over Silicon valley
O2O startups - Online to offline startups.
Pay with bar codes
CarWash Apps
Beggar accepting AliPay,WeChat Pay.
Armies of food deliverymen
School children pickup vans
On demand electric scooters(IoT devices) with solar powered GPS,accelerator, NFC, bluetooth
Shared Bikes that lockup by scanning barcodes
Super App model- One app to rule them all - "WeCHat", JustDial?
AI World Order
Wang xing cloned facebook, twitter, groupon in China and made a huge thing out of it.
fake it till you make it ideology that initially enabled him to copy and then create local tech capabilities and features that the valley does not comprehend.
Around the Clock work ethic.
Faster, Nimbler, Meaner, Leaner
Relentless iterating.
Corporate America is unprepared for the wave of Chinese enterpreneurship that AI brings.
Hard work
Invention of deep learning means we are moving from the age of expertise to the age of data.
Google left china!
Mass enterpreneurship and Mass innovation
7 giants of AI - Google, Facebook, Amazon, Microsoft, Baidu,Alibaba, Tencent
translate speech to any other language.
amazon buy recommendations.
youtube see suggestions
addiction causing algorithms.
Few AI companies:iFlyTek,Toutiao,rxthinking,VIPKid,Traptic,LipNet
Humans use data to validate their assumptions and create theories based upon that.
Programs use data to generate multiple assumptions and test the assumptions truth and learn from that to create better theories.
Derive predictive power from data points that would seem irrelevant to a human loan officer.. ex: speed at which you typed your name, battery power left on your phone..
Algorithms on par with doctors at diagnosing specific illnesses based upon images,xrays.
AI augments and empowers humans in their decisions.
OMO- Online Merges Offline [order meal speaking from couch,pay with your face]
Perception AI is a hardware enterprise..audio-video streamers.
China has the hardware supply chain advantage.
Autonomous AI will surface first in commercial settings rather than domestic.
Swarm intelligence - drones,satellites.
intelligent superhighway.
World first AI city to accomodate autonomous vehicles: xiong
Didi drove Uber out of China and partnered with local startups like Lyft,Ola, Grab[singapore],taxify[estonia],careem[middle east] to form an anti-uber alliance.
AI needs localisation to serve local markets.
Virtual Therapist
People talking to machines becoming normal.
The second machine age...manufacturing machines replaced by autonomous robots throughout the production line.
Most of today's white collar workforce is paid to take in and process information and then make a decision or recommendation based upon that, which is precisely what AI algorithms do best.
White collar jobs will be replaced with before the Blue collar jobs[through 2nd machine age]
Rather than replacing one to one, AI based startups will fundamentally disrupt through competition to existing industries. ex: Loan decision making company.
It is relatively easy for AI to mimic high level intellectual or computational ability of an adult but harder to give sensory motor skills of a toddler.
AI is a monopolistic centralizing technology.
theano,torch,caffee
The Deep Learning Revolution -Sejnowski Terrence
Recent progress in AI has occured by reverse engineering the brain.
Translate text from one language to another.
Skin Analysis by taking photo.
Detection of cancer through Biopsies.
Deep Sleep Problems EEG analysis.
Cheap sensors that record data continuously.
Deep Learning is getting better at stock transactions compared to algorithmic trading.
Machine learning is being used for credit evaluation for loans.
Technology assited review in law firms.
Games Poker,Go are better played by machines.
Machine learning is used to evaluae huge petabytes of datasets generated by scientific instruments used in astronomy.
Constant learning, learning from home online will help survive the rate of change.
Progress in computer vision was made not by focussing on pixels but on features.
In complex domains like medical , diagnosticians relied on patterns based on experience rather than facts or rules making it difficult to build expert systems.
Initial attempts at AI failed in mimicing programs like the way brain[neurons] worked, but later neural networks took that into consideration.
A fly is so small and weights almost nothing. It can fly, it can see, it can navigate and find food, and what is truly remarkable is that it can reproduce itself. How does our technology compare?
Studying how human functions helps in design.
Eyes see, Ears listen, etc..each does one function well.
We learn to perform difficult tasks through practice.
Brains arent filled with logic or rules.
Reasoning is domain specific and the more familiar with domain, the easier it is to intuit solutions.
Past examples help to solve current problems.
We learn domains electricity/magnetism by solving problems not by memorizing formulae.
For solutions to hard problems , we should look into massively parallel architectures like neurons in our brains.
Support vector machines are programs used to classify multiple sets of data by their boundary points or support vectors.
Relationship of a single data point with respect to others will determine where it stands according to a classifier function.
Neural network equations are nonlinear, noise is Gaussian, and variables non-separable.
Cocktail party problem is solved by Independent Component Analysis which is an unsupervised algorithm, which also led to noise cancelling headphone invention.
Hopfield Net: A particular type of nonlinear network model converges to a stable state called hopfield net, similar to how memories work..if we know a person face, we can recall his name and associated memories.
A variant of hopfield net can be used to solve the travelling salesman problem that is difficult by traditional computer science.
Vase or Face problem
Boltzmann Machine: Find patterns in data by reconstructing the inputs., concept : Human body is bilaterally symmetric.
Optimisation is a key mathematical concept.For many problems a cost function can be found for which the solution is the system with lowest cost.[Gradient Descent.]
Credit Card fraud is detected and credit decisions are made by neural networks.
Deep Networks are core technology for online streaming.
The complete study of mankind is man.
Study of probability distribution in high dimensional spaces was relatively unexplored area of statistics in 1980s.
Neural networks may arrive at a right answer but cannot tell how they arrived at it, on the contrary an experienced doctor might be able to say why to convince a human.[even though it may be faulty at times.]
Convolutional Learning:
ConvNet by Yann LeCun
Convolution can be thought of as a small sliding filter that is passed over an image creating a layer of features over the image.
Output of each filter passed through sigmoid function or Rectified Linear Units(RELUs)
CNN meets Visual Hierarchy.
2015 Kelvin xu LSTM - long short term memory. - unsupervised.
Generative Adversarial Networks produce new input samples when the output is clamped to a category.
Reward or Reinforcement Learning
Reinforcement network takes input makes decisions and takes action.
Much of human learning is based on observation and mimicry.
Neural Information Processing Systems (NIPS) conferences
MOOC Massive Open Online Courses
Big Data made Deep Learning take off.
Our bodies are marvels transforming food into energy and body parts.
Baxter robot used for performing actions like humans.
Facial expressions are window into your soul.
Sensing,Cognition and Emotion are different layers of the being.
Biggest problems in education are not scientific but social and cultural.
Cellular AUtomata- game of life - WOlfram Alpha- Mathematica
Intel purchased Nervana(Naveen Rao),Mobileeye
Google-TensorFlow
CNTK-Microsoft
MVNet-Amazon
handicraft
kreatvt
critical thought
love-empathy-and all abstract ideas..motivation
jokes
AI
Decide
Stock market trading
Military predator drones
loan
insurance
credit
Analyse/Predict
Prices of airlines/hotels
Recognizes tumour better than human
Surveillence cameras/microphones
identify face
recognize speech
translate speech at runtime
identify fraud.
consumer behaviour
habits of humans
Calendar assistant
context aware apps
Writes itself "autonomous"ly
Self-service
Bias in the dataset in identifying
Economic incentives to replace humans with AI will be beyond anything seen since industrial revolution.
Choice of classification of categories introduces a bias, and this is a danger that many are warning us about AI.
This ambiguity is compounded by the datasets that are used to train the network.
Neural networks may optimise profits at the expense of fairness.
Same issues exist with technology earlier attempts as this is not technology issue but human negatives when combined with technology may exponentialize the harm.
Even in brain conclusions are not always based upon logic and there are cognitive biases.
The only thing we need to fear is ourselves.
machines will replace humans at automation tasks in industries as they are faster,cheaper,better. Manufacturing Enterprises will be wiped out of employees.
Jobs that disappear:
driver,cook,motel,house builders,accountants,assembly line workers,warehouse opertors, security guards,QC inspectors, truckers,stock analysts,paralegals, radiologists,farmers
Industries disrupted: Trucking, Manufacturing, Retail,Insurance.
The AI makers thought process will succeed and the naysayers will not have an option but to accept, unless we participare in the AI from the ground up and influence how it shapes.
Humans dont learn from instructions, they practice and experience and learn from that.
SPAM machine learning:
human acts on a task
program watches the action.
program learns from watching.
program predicts new task outcome.
program acts on new task.
human evaluates program performance.
program learns from evaluation.
program improves its learning
program acts applying the new learnings.
In supervised learning we are given a data set and already know what the correct output would look like.
Supervised: regression, classification
In unsupervised learning, only data set is given.
Clustering: google news
cocktail party problem algorithm: Separate 2 users voices whose output is combined in a audio recording.
training set (x(i),y(i))
learning algorithm - h , h(x)
cost function - squared error function
cost function is used to check the performance of the hypothesis.
algorithm that automatically finds the function that minimizes the cost function
This algorithm is called gradient descent
Inputs needed for a good ai: huge data,strong algorithm, narrow domain,concrete goal.
AI Machines radically empower some human beings and violently displace some other human beings.
Machines ability to "LEARN"
reinforcement learning, transfer learning
Implementation or Application of existing research into domains is the cause of hype, rather than research brekthroughs.
Broad availability of huge amounts of data is the cause of hype, compared to earlier lockedin systems.
China Advantage over Silicon valley
O2O startups - Online to offline startups.
Pay with bar codes
CarWash Apps
Beggar accepting AliPay,WeChat Pay.
Armies of food deliverymen
School children pickup vans
On demand electric scooters(IoT devices) with solar powered GPS,accelerator, NFC, bluetooth
Shared Bikes that lockup by scanning barcodes
Super App model- One app to rule them all - "WeCHat", JustDial?
AI World Order
Wang xing cloned facebook, twitter, groupon in China and made a huge thing out of it.
fake it till you make it ideology that initially enabled him to copy and then create local tech capabilities and features that the valley does not comprehend.
Around the Clock work ethic.
Faster, Nimbler, Meaner, Leaner
Relentless iterating.
Corporate America is unprepared for the wave of Chinese enterpreneurship that AI brings.
Hard work
Invention of deep learning means we are moving from the age of expertise to the age of data.
Google left china!
Mass enterpreneurship and Mass innovation
7 giants of AI - Google, Facebook, Amazon, Microsoft, Baidu,Alibaba, Tencent
translate speech to any other language.
amazon buy recommendations.
youtube see suggestions
addiction causing algorithms.
Few AI companies:iFlyTek,Toutiao,rxthinking,VIPKid,Traptic,LipNet
Humans use data to validate their assumptions and create theories based upon that.
Programs use data to generate multiple assumptions and test the assumptions truth and learn from that to create better theories.
Derive predictive power from data points that would seem irrelevant to a human loan officer.. ex: speed at which you typed your name, battery power left on your phone..
Algorithms on par with doctors at diagnosing specific illnesses based upon images,xrays.
AI augments and empowers humans in their decisions.
OMO- Online Merges Offline [order meal speaking from couch,pay with your face]
Perception AI is a hardware enterprise..audio-video streamers.
China has the hardware supply chain advantage.
Autonomous AI will surface first in commercial settings rather than domestic.
Swarm intelligence - drones,satellites.
intelligent superhighway.
World first AI city to accomodate autonomous vehicles: xiong
Didi drove Uber out of China and partnered with local startups like Lyft,Ola, Grab[singapore],taxify[estonia],careem[middle east] to form an anti-uber alliance.
AI needs localisation to serve local markets.
Virtual Therapist
People talking to machines becoming normal.
The second machine age...manufacturing machines replaced by autonomous robots throughout the production line.
Most of today's white collar workforce is paid to take in and process information and then make a decision or recommendation based upon that, which is precisely what AI algorithms do best.
White collar jobs will be replaced with before the Blue collar jobs[through 2nd machine age]
Rather than replacing one to one, AI based startups will fundamentally disrupt through competition to existing industries. ex: Loan decision making company.
It is relatively easy for AI to mimic high level intellectual or computational ability of an adult but harder to give sensory motor skills of a toddler.
AI is a monopolistic centralizing technology.
theano,torch,caffee
The Deep Learning Revolution -Sejnowski Terrence
Recent progress in AI has occured by reverse engineering the brain.
Translate text from one language to another.
Skin Analysis by taking photo.
Detection of cancer through Biopsies.
Deep Sleep Problems EEG analysis.
Cheap sensors that record data continuously.
Deep Learning is getting better at stock transactions compared to algorithmic trading.
Machine learning is being used for credit evaluation for loans.
Technology assited review in law firms.
Games Poker,Go are better played by machines.
Machine learning is used to evaluae huge petabytes of datasets generated by scientific instruments used in astronomy.
Constant learning, learning from home online will help survive the rate of change.
Progress in computer vision was made not by focussing on pixels but on features.
In complex domains like medical , diagnosticians relied on patterns based on experience rather than facts or rules making it difficult to build expert systems.
Initial attempts at AI failed in mimicing programs like the way brain[neurons] worked, but later neural networks took that into consideration.
A fly is so small and weights almost nothing. It can fly, it can see, it can navigate and find food, and what is truly remarkable is that it can reproduce itself. How does our technology compare?
Studying how human functions helps in design.
Eyes see, Ears listen, etc..each does one function well.
We learn to perform difficult tasks through practice.
Brains arent filled with logic or rules.
Reasoning is domain specific and the more familiar with domain, the easier it is to intuit solutions.
Past examples help to solve current problems.
We learn domains electricity/magnetism by solving problems not by memorizing formulae.
For solutions to hard problems , we should look into massively parallel architectures like neurons in our brains.
Support vector machines are programs used to classify multiple sets of data by their boundary points or support vectors.
Relationship of a single data point with respect to others will determine where it stands according to a classifier function.
Neural network equations are nonlinear, noise is Gaussian, and variables non-separable.
Cocktail party problem is solved by Independent Component Analysis which is an unsupervised algorithm, which also led to noise cancelling headphone invention.
Hopfield Net: A particular type of nonlinear network model converges to a stable state called hopfield net, similar to how memories work..if we know a person face, we can recall his name and associated memories.
A variant of hopfield net can be used to solve the travelling salesman problem that is difficult by traditional computer science.
Vase or Face problem
Boltzmann Machine: Find patterns in data by reconstructing the inputs., concept : Human body is bilaterally symmetric.
Optimisation is a key mathematical concept.For many problems a cost function can be found for which the solution is the system with lowest cost.[Gradient Descent.]
Credit Card fraud is detected and credit decisions are made by neural networks.
Deep Networks are core technology for online streaming.
The complete study of mankind is man.
Study of probability distribution in high dimensional spaces was relatively unexplored area of statistics in 1980s.
Neural networks may arrive at a right answer but cannot tell how they arrived at it, on the contrary an experienced doctor might be able to say why to convince a human.[even though it may be faulty at times.]
Convolutional Learning:
ConvNet by Yann LeCun
Convolution can be thought of as a small sliding filter that is passed over an image creating a layer of features over the image.
Output of each filter passed through sigmoid function or Rectified Linear Units(RELUs)
CNN meets Visual Hierarchy.
2015 Kelvin xu LSTM - long short term memory. - unsupervised.
Generative Adversarial Networks produce new input samples when the output is clamped to a category.
Reward or Reinforcement Learning
Reinforcement network takes input makes decisions and takes action.
Much of human learning is based on observation and mimicry.
Neural Information Processing Systems (NIPS) conferences
MOOC Massive Open Online Courses
Big Data made Deep Learning take off.
Our bodies are marvels transforming food into energy and body parts.
Baxter robot used for performing actions like humans.
Facial expressions are window into your soul.
Sensing,Cognition and Emotion are different layers of the being.
Biggest problems in education are not scientific but social and cultural.
Cellular AUtomata- game of life - WOlfram Alpha- Mathematica
Intel purchased Nervana(Naveen Rao),Mobileeye
Google-TensorFlow
CNTK-Microsoft
MVNet-Amazon
Comments