deck_15613364 Flashcards
FOL
First Order Logic
In terms of FOL, what is the following:
Objects:
Relations:
Functions:
- Objects: people, houses, numbers, theories, Ronald McDonald, colors,
baseball games, wars, centuries . . . - Relations: red, round, bogus, prime, multistoried . . .,
brother of, bigger than, inside, part of, has color, occurred after, owns,
comes between, . . . - Functions: father of, best friend, third inning of, one more than, end of
Ontological Engineering
General and flexible representations for complex
domains.
Upper ontology:
The general framework of concepts
Categories and Objects
Stuff: a significant portion of reality that seems to defy any obvious
individuation—division into distinct objects
Intrinsic: they belong to the very substance of the object, rather than to the
object as a whole.
Extrinsic: weight, length, shape
Substance: a category of objects that includes in its definition only intrinsic
properties (mass noun).
Count noun: class that includes any extrinsic properties
© 2021 Pearson Education Ltd.
Mental Objects
Mental objects are knowledge in someone’s head (or KB)
Propositional attitudes that an agent can have toward mental objects
* Eg: Believes, Knows, Wants, and Informs
Lois knows that Superman can fly:
Knows(Lois, CanFly(Superman))
Modal Logic
Sentences can sometimes be verbose and clumsy. Regular logic is concerned with
a single modality, the modality of truth.
Modal logic addresses this, with special modal operators that take sentences
(rather than terms) as arguments
Semantic networks
- convenient to perform inheritance reasoning
- Eg: Mary inherits the property of having two legs. Thus, to find out how many
legs Mary has, the inheritance algorithm follows the MemberOf link from Mary
to the category she belongs to and then follows SubsetOf links up the hierarchy
until it finds a category for which there is a boxed Legs link
Description logics
- notations that are designed to make it easier to describe definitions and
properties of categories - evolved from semantic networks in response to pressure to formalize what the
networks mean while retaining the emphasis on taxonomic structure as an
organizing principle - Principal inference tasks:
- Subsumption: checking if one category is a subset of another by
comparing their definitions - Classification: checking whether an object belongs to a category
- The CLASSIC language (Borgida et al., 1989) is a typical description logic
- Eg: bachelors are unmarried adult males
- Bachelor = And(Unmarried, Adult, Male)
© 2021 Pearson Education Ltd.
Belief revision:
inferred facts will turn out to be wrong and will have to be
retracted in
Truth maintenance systems
or TMSs, are designed to handle complications of
any additional sentences that inferred from a wrong sentence.
Justification-based truth maintenance system (JTMS)
- Each sentence in the knowledge base is annotated with a justification
consisting of the set of sentences from which it was inferred - Justifications make retraction efficient
- Assumes that sentences that are considered once will probably be
considered again
fol basic elemets
Events
NLP
Natural Language Processing
N-GRAMS
Smoothing n-gram models
Atomic Model in N-grams
Part-of-speech (POS) tagging
Part-of-speech (POS) tagging
* way to categorize words (lexical category/tag)
* POS allows language models to capture
generalizations such as “adjectives generally come
before nouns in English”
* useful first step in many other NLP tasks, such as
question answering or translation
Big Libraries for this are NLTK and spaCy. Other libraries exist
in other languages but Python generally regarded as main NLP
language these days.
Grammar
A grammar is a set of rules that defines the tree structure of allowable phrases
* A language is the set of sentences that follow those rules.
* Syntactic categories such as noun phrase or verb phrase help to constrain the
probable words at each point within a sentence
* The phrase structure provides a framework for the meaning or semantics of the
sentence
Probabilistic context-free grammar (PCFG)
* A probabilistic grammar assigns a probability to each string
* “context-free” means that any rule can be used in any context
Parsing
Parsing is the process of analyzing a string of words to uncover its phrase
structure, according to the rules of a grammar.
* Search for a valid parse tree whose leaves are the words of the string
* can start with the S symbol and search top down, or we can start with the
words and search bottom up.
* Inefficiency: If the algorithm guesses wrong, it will have to backtrack all
the way to the first word and reanalyze the whole sentence under the
other interpretation.
* every time we analyze a substring, store the results so we won’t have to
reanalyze it later.
Parsing Tree
dependency grammar:
- assumes that syntactic structure is formed by binary relations between lexical
items, without a need for syntactic constituents - phrase structure tree is annotated with the head of each phrase
- recover dependency tree
- Convert dependency tree to phrase structure with arbitrary categories
Learn Parser From Examples
Learning semantic grammars
Pragmatics:
- resolving the meaning of indexicals, which are phrases
that refer directly to the current situation - Example sentence: “I am in Boston today,” both “I” and
“today” are indexicals. The word “I” would be represented
by Speaker, a fluent that refers to different objects at
different times - interpreting the speaker’s intent
- The speaker’s utterance is considered a speech act, and it
is up to the hearer to decipher what type of action it is
(question, a statement, a promise, a warning, a command,
etc.)
Time and tense:
Ambiguity:
Disambiguation
Deep Learning
Feedforward network
- connections only in one direction (input to output)
- directed acyclic graph with designated input and output nodes
- No loops
Recurrent network
- its intermediate or final outputs back into its own inputs.
- signal values within the network form a dynamical system that has internal
state or memory
Networks as complex functions
Different activation functions
Input encoding
Input Encoding For Image
back-propagation
back-propagation: the way that the error at the output is passed back
through the network – not going to get into this in a ton of detail today
In practice when we train a neural network with see that a loss function
is minimized
Recurrent neural networks (RNNs)
Recurrent neural networks (RNNs) allow cycles in the computation graph
each cycle has a delay
* units may take as input a value computed from their own output at an
earlier step
* RNN has an internal state/memory
* RNNs add expressive power compared to feedforward networks,
Training a basic RNN
* input layer x, a hidden layer z with recurrent connections, and an output
layer y,
Generalization
Choosing a network architecture
Some neural network architectures are explicitly designed to generalize well on
particular types of data
When comparing two networks with similar numbers of weights, the deeper network
usually gives better generalization performance.
Deep learning systems perform better than any other pure machine learning
approaches for high dimensional inputs (images, video, speech signals, etc)
Deep learning models lack the compositional and quantificational expressive power
May also produce unintuitive errors. Tend to produce input–output mappings that
are discontinuous
Neural architecture search
Neural architecture search
* neural architecture search to explore the state space of possible network
architectures.
* Some options to do this:
* Evolutionary algorithms: recombination (joining parts of two networks
together) and mutation (adding or removing a layer or changing a
parameter value)
* train one big network, search for subgraphs of the network that perform
better
Transfer learning and multitask learning
For transfer learning experience with one learning task helps an agent learn
better on another task
freeze the first few layers of the pretrained model that serve as feature
detectors
modify the parameters of the higher levels only
* problem-specific features and do classification
Common to start with pretrained model such as the ROBERTA model
Followed by fine-tuning the model in two ways:
* giving it examples of the specialized vocabulary used in the desired domain
* training the model on the task it is to perform
Multitask learning is a form of transfer learning in which we simultaneously train
a model on multiple objectives
Vision – CNNs (didn’t cover these today)
- success of the AlexNet deep learning system in the 2012 ImageNet competition
that propelled deep learning into the limelight - supervised learning task with 1,200,000 images in 1,000 different categories
- top-5 error rate has been reduced to less than 2%— below error rate of trained
human (5%)
Natural language processing - RNNs
- machine translation and speech recognition
- end-to-end learning, the automatic generation of internal representations for the
meanings of words, and the interchangeability of learned encoders and decoders - end-to-end learning outperforms classical pipelines
- re-representing individual words as vectors in a high-dimensional space—so-
called word embeddings
What is a nlm
Neural Language Model
What is a Corpus
What is a Token
Vocabulary
AI and the Cloud
- Need video memory
- Need time
- Difficult to run computers for extended
periods (power supply, hardware) - Training of big neural networks has moved to
the cloud
Rant on GPUs
- Deep learning can be done on the CPU but it is
slow - Generally need a GPU (4060Ti, 4080, 3090,
4090) - Memory biggest factor for training models
- 16GB+
FLAN-T5:
works if you ask it questions
directly
What is a hidden layer in the computation graph in deep learning
Intermediate computations before producing the output y.
Different representation for the input x.
Each layer transforms the representation produced by the preceding layer to produce
a new representation
In the process of forming all these internal transformations, deep networks often
discover meaningful intermediate representations of the data
The hidden layers of neural networks are typically less diverse than the output layers.
Back Propagation
back-propagation: the way that the error at the output is passed back
through the network – not going to get into this in a ton of detail today
In practice when we train a neural network with see that a loss function
is minimized
Neural network inputs
Properties of Neural Networks
Activation Function
Tensor operations in CNNs
tensors, which (in deep learning terminology) are simply multidimensional
arrays
of any dimension
Vectors and matrices are one-dimensional and two-dimensional special
cases of tensors
Computational efficiency of tensor operations: given a description of a
network as a sequence of tensor operations, deep learning software package
can generate compiled code that is highly optimized
Are run on GPUs (graphics processing units) or TPUs (tensor processing
units), which make available a high degree of parallelism
Parameters (Weights)
tunable values learned during training
process
Epoch
full pass of your training data
Batch Size
N samples from your training data
Iteration
updating your model every batch
Hidden Layer
intermediate layer
Loss function
we want to minimize this to zero
Frameworks
Keras, PyTorch
Image classification with convolutional neural networks
With enough training data and enough training ingenuity,
CNNs produce very successful classification systems
Images can have small alterations, patterns can be quite
informative.
Convolution followed by a ReLU activation function—as a
local pattern detector
composite patterns can be detected by applying another
layer to the output of the first layer.
Convolution Filter
Data set augmentation
training examples are copied and modified slightly
Images can have small alterations without changing the identity
* randomly shift, rotate,or stretch an image by a small amount, or randomly shift
the hue of the pixels by a small amount local patterns
* CNN-based classifiers are good at ignoring patterns that aren’t discriminative
* Context: patterns that lie off the object might be discriminative
* e.g., a cat toy, a collar with a little bell, or a dish of cat food might actually
help tell that we are looking at a cat
Detecting Objects
Object detectors find multiple objects in an image, report what class each object
is and also report where each object is by giving a bounding box around the
object
Building an object detector:
* looking at a small sliding window onto the larger image—a rectangle.
* At each spot, we classify what we see in the window, using a CNN classifier
Details:
* Decide on a window shape
* Build a classifier for windows
* Decide which windows to look at
* Choose which windows to report
* Report precise locations of objects using these windows
Robot Overview
- physical agents that perform tasks by manipulating the physical
world - equipped with effectors such as legs, wheels, joints, and grippers
- equipped with sensors, which enable them to perceive their
environment - Maximizing expected utility for a robot means choosing how to
actuate its effectors to assert the right physical forces - Robotic learning is constrained because the real world operates at
real time
Applications of Robots
Home care: Robots have started to enter the home to care for older adults and
people with motor impairments
Health care: Robots assist and augment surgeons, enabling more precise,
minimally invasive, safer procedures with better patient outcomes
Services: Mobile robots help out in office buildings, hotels, and hospitals
Autonomous cars
Entertainment: Disney has been using robots (under the name animatronics) in
their parks since 1963.
Exploration and hazardous environments: Robots have gone where no human
has gone before, including the surface of Mars.
Industry: The majority of robots today are deployed in factories, automating
tasks that are difficult, dangerous, or dull for humans.
Deployment of AI: Considerations
Deploying ai Low stakes
Deploying ai high stakes
Digital Content
No Free Lunch Therome
AGI ethics
Transparency
Weak Ai
- weak AI: the idea that machines could act as if they were intelligent
Strong Ai
he assertion that machines that do so are actually consciously
thinking (not just simulating thinking)
Good Old-Fashioned AI (GOFAI
- simplest logical agent design
- qualification problem: difficult to capture every contingency of appropriate
behavior in a set of necessary and sufficient logical rules
The argument from disability
Measuring AI
Representing the state of the world
The mathematical objection
Deciding what we want
Resources
AI engineering
AI engineering
* The AI industry has not yet reached that level of maturity.
* We do have a variety of powerful tools and frameworks, such as
TensorFlow, Keras, PyTorch, CAFFE, Scikit-Learn and SCIPY.
* But many of the most promising approaches, such as GANs and deep
reinforcement learning, have proven to be difficult to work with—they
require experience and a degree of fiddling to get them to train properly in
a new domain
* start with a single huge system and, for each new task, extract from it the
parts that are relevant to the task
The future
