AI Quiz
Artificial narrow intel (ANI)
Focused on a specific task at hand AI that does one task
Generative Adversarial Networks--GANS
Two neural networks that are in competition with each other it provides a means to generate new data that resembles the data that a system may have been trained on Generator produces fake outputs, discriminator tries to decide if it is real or fake--random data sent to the D Discriminator: tries to differentiate real outputs from fake outputs It has two lines and a gate that allows it to switch between one line, the Generator that sends it fake data, or another line that sends it real data Over time, both improve, so generator can produce highly realistic creations across many formats bad examples: deepfakes good: filtering smap emails
Activation
Value of a neuron--the individual number associated with a specific neuron the larger the number, the more activated the neuron is and the more emphasis it will have in the process
Computing Loss
We wanted the value as close to 0 as possible—but we ended up with .5, so the loss is about .5 You would do that for all your output neurons, add them all together, do another equation to compute the overall loss—how bad was your network at computing the values that it should have
the input layer
Whatever data you have coming into the system
Reinforced Learning
You have an agent that assesses an environment, makes a decision and then receives the consequences of that decision--teaching it how to act Trying to have that agent make better decisions through trial and error in order to maximize their overall rewards going to be observing its environment by taking step and seeing what happens--going from there Introducing software known as a ML agent to the environment and tell it how to act Works well in defined environment like Chess
Supervised Learning
A kind of ML takes training data, in which it's given the "right answer," and then has the machine's learning algorithm examine the data and develop a model It tests and refines the model on the training data set, then uses the final version of the model to handle new data It uses labeled data from a human--the system is given the right answer right away and knows if it is right or wrong. It knows if it got the wrong answer that everything is connected wrong. Computer is given three things -A set of data to learn from (labeled) -A model that determines how the computer approaches the problem of assessing the data -The cost function mostly used for regression (predict a continuous numerical value) or classification (assign a label)
Deep Learning
A learning system that uses a cascade of nodes that loosely mimic the neurons in the brain--neural network with many layers Called DEEP because of its hidden layers, it has many layers
Adversarial examples
Adversarial examples: optical/audio "illusions" for AI; inputs deliberately designed to make models fail Adversarial attacks: trying to make machine systems fail; usually small, subtle changes to data, causing systems that examine the data to fail Poisoning attack: poison the data that the system uses-creating a weakness in its training phase that can later be exploited (e.g., if they can manipulate training data, they can influence model behavior) Exploratory attack white box" direct access to model parameters or data black box --gain access to the model parameters (weights, bias term) or alogrhitinm indirectly --ex Oracle attack Image classifier example: They can tweak the system until they see the system misclassify or identify the image. Not discernable to the human eye Triceratops and Georgetown example → trying to make machine learning systems fail
What is the AI Triad
Algorithm, data, and computing power
Alignment challenge
Alignment is the difference between what we humans want the system to do and what the agent or computer does One point is that utility functions are difficult to perfectly align with what humans want, which in turn is difficult to precisely pinpoint. Another point is that AI will seek to succeed at its assigned task. Comes up most in super intel example: emergency button It fills it up, and we do not press the button, we say you get a good job—here is a one If the button is not touched, and the bucket is empty and that is a bad it is 0 Emergency button pressed and AI shut down: 1(good) Otherwise 0 Fails because overflowing and shutting down is easier, it gets a 1 and is happier. Does not do what we wanted--does not just fill the bucket
Gradient descent
Computing how much to tweak each of these weights and bias terms to make the system more accurate computing weights and bias terms
Bias and Variance
High bias: a Machine Learning Algorithm that has high bias under-fits training data--meaning it does not explain or have good predictive value. High error rates and can't predict the outcomes The good news is it learns quickly since it is looking at a few pieces of data and cannot get a good concept of the data it is examining High variance: gets a lot of different answers, MLA over-fits the data meaning it is picking up on a lot of different features in the data and sometimes just noise--picking up on background features and not just the data. Models the data very precisely, answers the question very accurately but then when it gets new data, it fails. real key is finding the balance between the two
Bias term
How much emphasis we want a particular neuron to have as we work through this process. The emphasis piece is important because it is a parameter or number you can change.
Addressing bias in AI
Improve transparency independent test and evaluation of algorithms and their results interdisciplinary teams with domain expertise: By bringing in teams with this, you increase the likelihood that the algorithms that are designed will understand the data sets being used You need to bring people in the criminal justice example with this Machine algorithms should not be done by JUST ML people, need people who understand the underlying data sets
Unsupervised learning
It does not have the right answers, it does not have the labels It looks at the data and tries to draw conclusions, looks at patterns in the data Looks at specific features of the data, like patterns, to find its answer Used for recommender systems--used a lot Also used to spot clusters in a set of data (that the algorithm finds on their own) Unsupervised learning is more useful when there is not a clear outcome of interest
Machine Learning
It is part of ANI and is the process of instructing computers to learn involves programming computers to teach themselves from data rather than instructing them to perform certain tasks in certain ways Leveraging or instructing algorithms to learn from data Using compute, applying an algorithms and set of rules to learn from data Making inferences from the data it is receiving and it is creating and understanding of that data
Output layer
It is the prediction that comes at the end--it is what the system thinks
Weight
It represents the strength of a connection between two neurons A number associated between two connections and that number between them has a weight The closer the number is to 1, the stronger the connection is between the two neurons
Neural Network composed of:
Nodes: mimics neurons in the brain, each artificial neuron has a number associated with it Weights Bias terms Activation of the neuron Activation function: all of the weights from the previous neuron, then squish it down (squishification), move it to a much smaller number often between 0 and 1--Sigmoid is an activation function
How can ML fail?
Non-representative training data: the Amazon resume example: If you start from a set of data that does not represent the population you are trying to examine or make predictions on you are going to have challenges Underlying bias in training data--the feature has an inherent bias--Amazon's resume example lack of transparency in describing conclusions--the black box: we know that it has classified something as a dog or a cat and we do not know what neuron caused that to happen--can't explain-- not knowing how a neural network comes to its conclusion (move 37 by AlphaGO) Irreducible error: the data itself may have underlying noise or features in it that you cannot take away poor tradeoff between bias and variance
Sigmoid
Once the number is all computed which happens inside a particular neuron, it is all pushed into a function called the Sigmoid It is a squish-ification function, it is taking numbers that can get large and is trying to compress them to a much smaller range, making them easier to manage
Overfitting and Underfitting
Over: VERY precisely matches the data, picking up on too much Trained to this specific data but the model fails in the real world when it is given other data Lots of parameters, very complex model tends to overfit Under: does not provide explanatory value, seems random. Very inaccurate because it is not picking up on enough Has fewer parameters, less complex has tendency to do this
Algorithms
Part of the AI triad A set of step-by-step instructions that computers follow to perform a task Role: Gives instructions so tells you what to do with the data and computing power—without the algorithm you do not get the benefits of what we have seen in ML three categories: supervised, unsupervised and reinforced learning
Neural Network Steps
Run through the training data with some initial set of weights Use a cost function to calculate the error Adjust the weights in small ways using backpropagation Run through the training data again and see if the error gets smaller If so, adjust the weights in the same direction If not, adjust the weights in the opposite direction
Agents
Something that acts to achieve an objective Most of the time it is something that sense its environment It is whatever entity is making a decision
AI Superintel
Spring off of AGI, smartest AI that can solve all problems It far exceeds human cognition
Neural Networks
a set of the neurons (the nodes) tied together in layers that connect to the one before and after it partially or fully connected (we deal with fully) Every neuron in the input layer is connected to every neuron in the next layer—same with in the first hidden layer to the second
Backpropagation
back propagation is the process of running backwards through the neural network to recalibrate/adjust the weights and associated bias term. makes the model more accurate
Forms of bias
bias term (see earlier) Bias in training data: Can train on data that isn't representative (e.g., image recognition trained on white males having issues identifying women and people of color) Can accurately predict results, but on an already biased system (e.g., credit algorithms on decisions that adversely impacted women and people of color Can be trained poorly and learn inaccurate or biased prediction rules--high variance or high bias Algorithm can be trained in a way that introduces bias by overlooking small groups
Computing Power
both semiconductors and broader processors that are used Need a lot more power to analyze more data the speed at which a machine can perform an operation
Step Size and Learning rate
he jumps in machine learning improvement, related to learning rate Small step size: need to train many times to converge to the model Large step size: may go quicker, but can miss the optimization point (stepping over the minimum of that model) Learning rate: how quickly an algorithm learns, related to step size A large learning rate--does not get down to the bottom of the cone where the optimized solution is Back and forth without resolving it But down to a learning rate of .000001 It then takes too long of a time—needs way too many epoch But with an epoch of .03 it only takes 31 tries to get to the ideal area
Artificial General Intelligence (AGI)
is the intelligence of a machine that could successfully perform any intellectual task that a human being can Not here yet
Data
part of AI triad Sets of values of quantitative or qualitative variables about one or more persons or objects Data MUST be the right quality and well represented While conceptual breakthroughs in the design of machine learning programs are significant, machine learning still relies on data A subpar algorithm can outperform a fantastic algorithm if given far more relevant data to learn from
Cost function
to calculate the error rate of the NN, or how far an algorithm is from perfect performance
Utility function
what you are trying to get an algorithm to do Have instrumental goals along the way We are asking it to maximize an agent's preferences