WHAT IS ARTIFICIAL INTELLIGENCE TODAY?
Who to follow
Recognize objects Play chess Medical
IT’S MOSTLY BECAUSE OF DEEP NEURAL
• DNNs consist of artificial neurons
(i.e., mathematical functions)
connected to each other
• Said neurons are arranged in layers,
and those signals — the product of
data, or inputs, fed into the DNN —
travel from layer to layer
NEURAL NETS ARE LOOSELY MODELED AFTER
THE BRAIN’S NEURONS
• Signals can be received from dendrites, and sent
down the axon
• Once enough signals are received, this outgoing
signal can then be used as another input for other
neurons, repeating the process.
• Some signals are more important than others and
can trigger some neurons to fire easier.
DEEP NEURAL NETS ARE TRAINED USING
• The procedure repeatedly adjusts
the weights of the connections in
the network so as to minimize a
measure of the difference between
the actual output vector of the net
and the desired output vector.
• As a result of the weight
adjustments, internal ‘hidden’ units
which are not part of the input or
output come to represent
important features of the task
• Uses gradient descent, to minimize
the loss function (difference
between the predicted vs. desired
DEEP LEARNING WAS INVENTED IN 1943!!!
Teaching these networks was so
computationally expensive that people rarely
used them for machine learning tasks
Compute got orders of
magnitudes faster (Moore’s law)
There was a lot of example
data to come by….
HUGE VICTORY FOR ALEXNET IN 2012
• Classify and detect objects
in Images in a massive
dataset of 14M images that
has over 21K classes
ALEXNET MADE A SIGNIFICANT LEAP...
AlexNet is the name of a convolutional neural network
• AlexNet competed in the ImageNet Large Scale Visual Recognition Challenge on September 30,
• The network achieved a top-5 error of 15.3%, more than 10.8 percentage points lower than that of
the runner up.
MAGIC OF DEEP LEARNING!
No need to hand-craft features
Just give it a bunch of labelled
data and minimize the loss
function, so that the neural
network learns the weights and
biases to make predictions
BIG DIFFERENCES BETWEEN THE BRAIN
AND DEEP LEARNING
• The brain has 86B neurons and 10T connections
• Neurons compute one layer after another, neurons
in the brain can fire asynchronously
• The brain is fault tolerant and self healing.
information is stored redundantly.
BIGGEST DIFFERENCE -
NEURONS THAT FIRE TOGETHER,
• Brain fibers grow and reach out to connect to other neurons,
neuroplasticity allows new connections to be created or
areas to move and change function, and synapses may
strengthen or weaken based on their importance.
• Deep Neural Network learning is rigid. The network is
trained once and then used for inference. It has to be
re-trained whenever there is new data .
DOPAMINE AND REINFORCEMENT
DOPAMINE IS ONE OF THE BRAIN’S
A neurotransmitter is a chemical that carries
information back and forth between neurons.
Glutamate Dopamine GABA Glycine
Acetyl. Norepin. Serotonin Endorphins
DOPAMINE ENABLES US TO TAKE ACTION
AND RECEIVE REWARDS
The dopamine kick that you
get when someone likes your
post is because dopamine is
modifying your neuronal
synapses and contributes to
feelings of pleasure.
DOPAMINE IN THE CONTEXT OF LEARNING
• Because the predictions are often not quite accurate, we need a way to calculate our prediction error so
we don’t make the same mistakes again (hence Reward Prediction Error) 😉.
Generally speaking, learning can be defined as the process of improving predictions of the
Reinforcement Learning Works The
Exact Same Way...
In simple terms, reinforcement learning algorithms use prediction error to
improve the computer’s ability to make better decisions in certain environments
AND NEURONAL LOCALITY
• Neurons interact with each other in
• Use the same concept of locality to train
hidden layers of a neural network to learn
lower level features.
• This results in a similar performance, as a
state-of-the-art supervised algorithm.
Paper: Unsupervised learning by competing hidden units
CAN MACHINES LEARN FROM A FEW
EXAMPLES, LIKE HUMANS DO?
• It takes a child only a few dozen
examples to learn the shapes of
letters like ‘a’ and ‘b’.
• This is because human brains
are very good at generalizing
from a few examples.
• Humans are good at inferring the concepts
conveyed in a pair of images and then applying
them in a completely different setting—for
example, the concept of stacking red and green
objects applied to different settings.
A new computer architecture called Visual cognitive computer (VCC) is proposed.
The components are based on the science of human cognition.
Human concepts are represented as cognitive programs.
VCC is evaluated on how well it can represent and infer visuospatial concepts
that cognitive scientists consider to be the fundamental building blocks.
A ROBOT THAT HAS VCC CAN PERFORM
A WHOLE ARRAY OF TASKS
THE TEAM THAT INVENTED VCC
• (/kæp.tʃə/, an acronym for "Completely
Automated Public Turing test to tell
Computers and Humans Apart") is a
type of challenge–response test used
in computing to determine whether or
not the user is human.
BAYESIAN INFERENCE IS A POPULAR
FRAMEWORK FOR PREDICTIONS
Bayesian inference is a method
of statistical inference in which
Bayes' theorem is used to update
the probability for a hypothesis
as more evidence or information
The basic idea of Bayesian probability is that you update
your beliefs in the light of new evidence
BAYESIAN BRAIN HYPOTHESIS
The hypothesis is meant to
explain several important
brain functions such as
perception, learning and
WHAT IS PREDICTIVE
• During every moment of your life, your brain gathers
statistics to adapt its model of the world, and this
model’s only job is to generate predictions.
• Your brain is a prediction machine.
• Just as the heart’s main function is to pump blood
through the body, so the brain’s main function is to
make predictions about the body.
KEY BRAIN FUNCTIONS EXPLAINED
BY THIS HYPOTHESIS
Learning is the updating of your internal model based on prediction errors so that
your predictions gradually improve. The better your predictions about the causal,
probabilistic structure of the world, the more effectively you can engage with it.
Memory consists of the learned parameters of your internal model, whereas its
non-acquired parameters would be the innate knowledge evolution has
genetically built into your nervous system. Both parts determine your brain’s
Belief is a hyperprior; a systemic prior with a high degree of abstraction; a high-
level prediction that encodes general knowledge about the world. (e.g. Physical
beliefs that apples fall down from a tree. Cultural beliefs that cars slow down
when you reach an intersection).
IN THE NEAR FUTURE
AI systems and devices will recognize, interpret,
process, and simulate human emotions.
MEASURE AND APPLY EMOTIONS
AI models will measure emotional response and
factor that into decision making.
Conversational Chatbots that detect
emotion and react accordingly.
Car software that detects if a driver is
angry and/or is not paying attention
and wants to take control.
Security software that alerts security
when there is fear in traveler’s face.
Chinese schools that monitor children’s
attention levels and alerts moms.
BUT ALGORITHMS STILL CAN’T
In order to ‘FEEL’ emotions, you have to be self aware…
There are other aspects of the human mind besides intelligence that are relevant to the
concept of strong AI which play a major role in science fiction and the ethics of artificial
Consciousness: To have
subjective experience and
Self-awareness: To be aware of oneself
as a separate individual, especially to
be aware of one's own thoughts.
Sentience: The ability to
"feel" perceptions or emotions
Sapience: The capacity for wisdom.