Cogsci 1 UC Berkeley Summer 2025 Final
UPDATED ACTUAL Exam Questions and
CORRECT Answers
Expert Systems - CORRECT ANSWER Computerized advisory programs that imitate the
reasoning processes of experts in solving difficult problems. (like MYCIN at Stanford in the
early 1970's). It was more accurate than human experts.
decision trees - CORRECT ANSWER Diagrams where answers to yes or no questions
lead decision makers to address additional questions until they reach the end of the tree. Basic
IF-THEN rules.
terminal leaf or node - CORRECT ANSWER the leaf at the end of a branch - the decision
/ outcome in a decision tree
Classification algorithm - CORRECT ANSWER A decision tree is like a classification
algorithm in that it divides down the population - it classifies (classifying system).
Machine Learning - CORRECT ANSWER the extraction of knowledge from data based
on algorithms created from training data (sub-field of expert systems)
ID3 algorithm (for machine learning) - CORRECT ANSWER The ID3 algorithm works
by assigning attributes to nodes. It identifies, for each node in the decision tree, which attribute
would be most informative at that point, called information gain. Information gain measures how
well a particular attribute classifies a set of examples. At each node, the algorithm chooses the
remaining attribute with the highest information gain.
target attribute - CORRECT ANSWER attribute you are most interested in when
classifying examples (objects in your database) - for example if Loan is the target attribute the
two possible values are yes and no.
,information gain - CORRECT ANSWER measures how well a given attribute separates
the training examples according to their target classification
entropy - CORRECT ANSWER a measure of uncertainty
feature engineering - CORRECT ANSWER Applying a machine learning algorithm such
as ID3 to a database depends crucially on how the examples in the database are labeled and
categorized. And that is an activity that is typically performed by human experts.
representation learning - CORRECT ANSWER Standard machine learning algorithms
such as ID3 can only get to work once the feature engineering has been carried out, typically by
programmers and other human experts. And this feature engineering is often the most
complicated and time-consuming part of the process. It is not hard to see the advantages of
designing programs and algorithms that would be able to do this feature engineering for us. The
technical name for this is representation learning or feature learning.
deep learning - CORRECT ANSWER The AI uses methods like unsupervised learning
which allows it to characterize the problem itself. It is a process that employs specialized
algorithms to model and study complex datasets; the method is also used to establish
relationships among data and datasets.
lateral geniculate nucleus (LGN) - CORRECT ANSWER The first station is the lateral
geniculate nucleus (LGN), which receives input directly from the retina. Because neurons in the
LGN have similar receptive fields to neurons in the retina, the LGN functions as a relay. It
projects to area V1 (the primary visual cortex, which maps more or less onto the anatomically
defined area known as the striate cortex because of its conspicuous stripe). This is where
"information processing" proper begins.
primary visual cortex (V1) - CORRECT ANSWER V1 takes the retinotopic map coming
from the LGN and filters it in a way that accentuates the edges and contours. This is the first step
in parsing the raw data arriving at the retina into objects.
secondary visual cortex (area V2) - CORRECT ANSWER The neurons in V2 are tuned to
the same features as the neurons in V1, as well as to more complex features, such as shape and
, depth. V2 is much less well understood than V1, but the neurons there are sensitive to multiedge
features, such as complexes of edges with different orientations.
Area V4 - CORRECT ANSWER Area V4 takes the representations from V2, which
contain primitive representations of shapes and some information about depth, and then uses
those representations to construct further representations that incorporate more information about
figure/ground segmentation, as well as about colors.
fusiform face area - CORRECT ANSWER a region in the temporal lobe (inferior temporal
cortex) of the brain that helps us recognize faces
fusiform body area - CORRECT ANSWER Brain region located in the inferior temporal
cortex (within the ventral visual-processing stream) that responds preferentially to human bodies
and body parts.
Selectivity/Invariance Problem - CORRECT ANSWER how to identify something
independent of the way it is represented (angle, movement, etc) White wolf vs Samoyed (dog)
autoencoder - CORRECT ANSWER a type of feedforward neural network that is trained
to compress and then reconstruct input data. Both encoding and decoding are done by hidden
layers.
convolutional neural network - CORRECT ANSWER A convolutional neural network is a
class of deep neural networks, most commonly applied to visual recognition. They often have
fewer connections then other neural networks. These are feedforward neural networks.
dimensionality reduction - CORRECT ANSWER data encoding schemes are applied so as
to obtain a reduced/compressed representation of the original data
ConvNet - CORRECT ANSWER short for convolutional neural networks - they help
make computer vision possible. they have these characteristics: Sparse connectivity, Shared
weights, Invariance under translation. they are feed-forward. They have at least one
convolutional layer.
UPDATED ACTUAL Exam Questions and
CORRECT Answers
Expert Systems - CORRECT ANSWER Computerized advisory programs that imitate the
reasoning processes of experts in solving difficult problems. (like MYCIN at Stanford in the
early 1970's). It was more accurate than human experts.
decision trees - CORRECT ANSWER Diagrams where answers to yes or no questions
lead decision makers to address additional questions until they reach the end of the tree. Basic
IF-THEN rules.
terminal leaf or node - CORRECT ANSWER the leaf at the end of a branch - the decision
/ outcome in a decision tree
Classification algorithm - CORRECT ANSWER A decision tree is like a classification
algorithm in that it divides down the population - it classifies (classifying system).
Machine Learning - CORRECT ANSWER the extraction of knowledge from data based
on algorithms created from training data (sub-field of expert systems)
ID3 algorithm (for machine learning) - CORRECT ANSWER The ID3 algorithm works
by assigning attributes to nodes. It identifies, for each node in the decision tree, which attribute
would be most informative at that point, called information gain. Information gain measures how
well a particular attribute classifies a set of examples. At each node, the algorithm chooses the
remaining attribute with the highest information gain.
target attribute - CORRECT ANSWER attribute you are most interested in when
classifying examples (objects in your database) - for example if Loan is the target attribute the
two possible values are yes and no.
,information gain - CORRECT ANSWER measures how well a given attribute separates
the training examples according to their target classification
entropy - CORRECT ANSWER a measure of uncertainty
feature engineering - CORRECT ANSWER Applying a machine learning algorithm such
as ID3 to a database depends crucially on how the examples in the database are labeled and
categorized. And that is an activity that is typically performed by human experts.
representation learning - CORRECT ANSWER Standard machine learning algorithms
such as ID3 can only get to work once the feature engineering has been carried out, typically by
programmers and other human experts. And this feature engineering is often the most
complicated and time-consuming part of the process. It is not hard to see the advantages of
designing programs and algorithms that would be able to do this feature engineering for us. The
technical name for this is representation learning or feature learning.
deep learning - CORRECT ANSWER The AI uses methods like unsupervised learning
which allows it to characterize the problem itself. It is a process that employs specialized
algorithms to model and study complex datasets; the method is also used to establish
relationships among data and datasets.
lateral geniculate nucleus (LGN) - CORRECT ANSWER The first station is the lateral
geniculate nucleus (LGN), which receives input directly from the retina. Because neurons in the
LGN have similar receptive fields to neurons in the retina, the LGN functions as a relay. It
projects to area V1 (the primary visual cortex, which maps more or less onto the anatomically
defined area known as the striate cortex because of its conspicuous stripe). This is where
"information processing" proper begins.
primary visual cortex (V1) - CORRECT ANSWER V1 takes the retinotopic map coming
from the LGN and filters it in a way that accentuates the edges and contours. This is the first step
in parsing the raw data arriving at the retina into objects.
secondary visual cortex (area V2) - CORRECT ANSWER The neurons in V2 are tuned to
the same features as the neurons in V1, as well as to more complex features, such as shape and
, depth. V2 is much less well understood than V1, but the neurons there are sensitive to multiedge
features, such as complexes of edges with different orientations.
Area V4 - CORRECT ANSWER Area V4 takes the representations from V2, which
contain primitive representations of shapes and some information about depth, and then uses
those representations to construct further representations that incorporate more information about
figure/ground segmentation, as well as about colors.
fusiform face area - CORRECT ANSWER a region in the temporal lobe (inferior temporal
cortex) of the brain that helps us recognize faces
fusiform body area - CORRECT ANSWER Brain region located in the inferior temporal
cortex (within the ventral visual-processing stream) that responds preferentially to human bodies
and body parts.
Selectivity/Invariance Problem - CORRECT ANSWER how to identify something
independent of the way it is represented (angle, movement, etc) White wolf vs Samoyed (dog)
autoencoder - CORRECT ANSWER a type of feedforward neural network that is trained
to compress and then reconstruct input data. Both encoding and decoding are done by hidden
layers.
convolutional neural network - CORRECT ANSWER A convolutional neural network is a
class of deep neural networks, most commonly applied to visual recognition. They often have
fewer connections then other neural networks. These are feedforward neural networks.
dimensionality reduction - CORRECT ANSWER data encoding schemes are applied so as
to obtain a reduced/compressed representation of the original data
ConvNet - CORRECT ANSWER short for convolutional neural networks - they help
make computer vision possible. they have these characteristics: Sparse connectivity, Shared
weights, Invariance under translation. they are feed-forward. They have at least one
convolutional layer.
