Midterm 3 Flashcards
What are the patterns of laminar interconnections?
feedforward connections from LGN to V1, feedforward connections to V2, feedback connections from V2 to V1
Information and processing in the eye
rod and cone receptors -> horizontal cells -> bipolar cells -> ganglion cell…out to optic nerve fibers and then to LGN in thalamus.
Process through neural spikes via light
What does the superior colliculus do?
sends info to motor/cerebellum area. It is the ‘inner eyes’
What are the five steps of visual computation?
1) rod and cone cells (light detection)
2) horizontal and bipolar cells (preprocessing)
3) ganglion cells (preprocessing)
4) Lateral geniculate nucleus (relay station)
5) visual cortex (conscious perception)
Rods vs cones
rods: discriminate between brightness in low illumination
. contribute to peripheral vision
cones: discriminate colors, contribute to central vision
What is our visual spectrum?
400nm (violet) - 700nm (red)
What are the three types of cone cells (color detectors)
S-cones (blue)
M-cones (green)
L-cones (red)
Why did we develop the system of trichromacy (three cone cells)?
evolutionarily advantageous to be able to distinguish between a wide range of colors. Since there is overlap of the wavelengths for the cones, they combine to create different colors
What is the sensory binding problem?
how does the brain combine different sensory features into one unified, coherent object
What are the elements of visual computation?
- receptive field
- on-center/off-surround RF
- edge detection
- orientation detector
- location-invarient orientation detector
What is the receptive field of a neuron?
the area of retina cells that trigger activity of that neuron
What is on-center/off-surround receptive field?
on-center cells - the surround ganglion cells are inhibitory and dampen the signal of the center ganglion cell. Most excitatory activation (biggest response) when light is concentrated in the center of the receptive field
off-surround cells have the opposite pattern -> most activation when light concentrated in surround illumination
How do the ganglion cells detect edges?
the on-center/off-surround RF is able to detect the change in light intensity which creates an edge
How do orientation detectors work?
a neuron has a preferred orientation
selectivity - respond to (detect) only a relatively narrow range of orientations
graded responses for nearby orientations (neighboring cells share similar orientations)
How were the first orientation cells discovered?
first detected in a cat where the excitatory vs inhibitory area of a selective neuron corresponds to the orientation preference
David Hubel and Torsten Wiesel used single-cell recordings in primary visual cortex of cats
Got nobel prize in physiology/medicine
How do orientation detector networks work?
‘Boolean AND’ neuron (all LGN cells need to be activated to activate cortical cell). selective cells in the retina are organized according to their orientation.
How do location-invarient orientation detectors work?
detecting motion of the same object. Many receptive fields of ganglion cells work together and go to LGN neurons which send signals to cortical simple cells and then activate the cortical complex cell. Boolean OR computation
What is the visual computational pathway?
Retina -> LGN -> VA -> V2…-> IT -> anterior IT
in anterior IT learning generalizes over orientation, location, and form
connectionist model approach
Brains are parallel, distributed, analog computers, not based on symbolic logic
father of neural network modeling (modern AI)
david rumelhart
Is a neural network a universal turing machine?
yes, it can solve any computable problem
Further, if equipped with appropriate algorithms, the neural
network can be made into an intelligent computing
machine that solves problems in fast and finite time.
Recent learning algorithms for neural networks
– Backpropagation learning rule (D. Rumelhart et al, 1986)
– Hebbian learning rule (1949)
– Kohonen’s self-organizing feature map (1982)
– Boltzman machine (1986)
– Deep learning network (G. Hinton et al, 2006)
Who won the 2024 physics nobel prize?
geoffrey hinton & john hopfield
content-addressable memory and deep belief nets
When was traditional vs modern AI?
traditional: 1950-2008 (symbolic system models)
modern: 2008-present (connectionist/neural networks)
mathematics of single neuron computation
Simplifying assumptions:
*Ignore:
– ion-current dynamics
– spike timing
*Discrete time “cycles” (~10 ms?).
*“Activation” models firing rate (0-100
spikes/s)
*“Weights” model synaptic efficacies.
if sum of weights x inputs exceeds threshold, binary output activation
Neural network model of word identification
- Word superiority effect:
– TAKE (faster)
– DAKE (slower) - Context sensitivity
- Pattern completion
examples of pattern completion by a hopfield network
noise tolerant memory: denoising
pattern completion (content addressable memory)
Simple Neural Networks:
McCulloch-Pitts (M-P) Nets
Binary output neurons, S = 1 (fire) or 0 (silent), with an all-or-none step function f
- Model of artificial neurons that computes Boolean
logical functions
where Y, X 1 , X 2 take on binary values of 0 or 1, W 1 , W 2 , Q
take on continuous values, and f is the step-function.
Boolean AND computation
Knowledge representation, learning, and acquisition of new knowledge in M-P network
Hiring rule problem
Knowledge is in the connection weights (and thresholds)
learning through weight modification
Acquisition of new knowledge through recruiting other neurons
McCulloch & Pitts
1943
– First neural network model
– Binary “neurons”
– Boolean logic functions
– No learning
D. Hebb
1949
– McGill psychobiologist
– Proposed a learning rule (“Hebb Learning Rule”)
* Unsupervised learning (i.e., learning w/o teacher)
(e.g., learning natural categories)
F. Rosenblatt
1957
– Perceptron: 2-layer network with Delta Learning Rule
Can the perceptron solve the convex problem?
No, it is nonlinear. Need third (hidden) layer.
First AI revolution
Backpropagation Learning Rule (1986)
* Led by David Rumelhart & James McClelland
* Discovered Backpropagation (BP) Learning Rule for multi-
layer networks
* “Any problems” can now be solved!
* Universal Turing Machines (“can compute whatever is
computable”)
* Biological plausibility of BP not established.
second AI revolution
Deep Neural Networks (DNN) (2006)
* Led by G. Hinton (U. Toronto)
* Possibility of realizing Strong AI
Deep neural net
A deep neural/learning/belief network is a multi-layer net with many,
many hidden layers (e.g., 5-100).
DNNs represent the most successful modeling framework to date for
solving real-world problems in speech recognition, text-based image
retrieval, and language translation.
what launched the second revolution?
A DNN For Hand-written Character Recognition
NN-review 26
Hinton et al. (Neural Computation, 2006)
Hierarchical feature representations in DNN
Object features are learned and presented in a hierarchical manner such
that each hidden layer represents a different level of features: the higher
the layer, the higher level the feature.
How do the two DNNs in AlphaGO interact with each other?
DNN1: To select the
next “best” move
DNN2: To evaluate the
odds of winning the match
C2 Consciousness (self-reflection, meta-cognition)??
Google deep mind competition against lee sedol
Generative Adversarial Networks (GAN)
a type of deep learning model that use an adversarial process, pitting a generator against a discriminator, to learn to generate realistic synthetic data. The generator creates fake data samples, and the discriminator attempts to distinguish them from real data.
face generative AI
godfather of modern AI
Geoffrey Hinton
Deep dreams
DeepDream is a computer vision program created by Google engineer Alexander Mordvintsev that uses a convolutional neural network to find and enhance patterns in images via algorithmic pareidolia, thus creating a dream-like appearance reminiscent of a psychedelic experience in the deliberately overprocessed images.
Three approaches to robotics AI
1) Model-based (intelligence with representation; symbolic processing)
v ACT-R (John Anderson, CMU)
2) Model-free/behavior-based (intelligence without representation;
reflexive processing)
v BigDog, Roomba (Rodney Brooks, formerly MIT)
3) Neural networks
Model-based Robotics/AI Systems:
Intelligence with Representation
ACT-R
SHAKEY the robot
What is ACT-R
ACT-R (Adaptive Control of Thought – Rational) is a
computational modeling framework developed in the
early 1990s by John Anderson at CMU (refined multiple
times since then).
John Anderson
(CMU)
ACT-R is a symbolic processing
approach to the study of cognition in the
tradition of the General Problem Solver
(GPS: Newell & Simon, 1959), as
opposed to the connectionist (i.e., neural
network) approach.
What is the current ACT-R?
6.0 brain mapping & instructable production system
organization of ACT-R
modular architecture
Modules connected serially & activated one after another
sensors -> cognition layer -> ACT-R buffers -> perception layer -> environment ->actuators
SHAKEY the robot
developed by SRI team in 1960/1970s
the world’s first mobile robot capable of perceiving and reasoning about its environment
could perform tasks requiring planning, route-finding, and rearranging simple objects.
symbol manipulation
does ACT-R have an explainability problem like AI does?
no, we know exactly what and why it’s doing what it’s doing
Cogito ergo sum
“I think, therefore I am.”
descartes
– The world is not present, but instead is
re-presented in an internal model of
the external world (“virtual world”)
– Cognition (information processing)
involves only the manipulation of
internal (symbolic) representations.
– Although an external world exists, it is
irrelevant for understanding cognition.
Behavior-based Robotics/AI Systems:
Intelligence without Representation
cricket robot
ALLEN the robot
BigDog
roomba
Atlas humanoid robot
father of behavior based robotics
Rodney Brooks
“The world outside is its own best model.”
Cricket robot
- Sensory inputs from
microphones - Motor outputs to
wheels - Neurons mediate
- Behavior is jointly
determined by:
– Environment
– Neuronal connectivity
The speed of the left wheel is proportional to the
intensity of the sound in the right microphone.
The speed of the right wheel is proportional to the
intensity of the sound in the left microphone.
- Note the tight coupling between the robot and its
environment. The robot does not build an internal
model of the world. It just reacts to sound
intensities.
a better cricket robot
- Additional inputs from
photoreceptors - Excitatory auditory
neurons - Inhibitory visual
neurons – only allow
motion in the dark - Motor neurons
integrate and “decide”
subsumption architecture
Made of layers of autonomous sub-systems that operate
simultaneously in parallel, unlike a modular architecture (e.g.,
ACT-R) that is made of functional sub-systems called modules
that operates serially.
Parallel Decomposition by Behaviors
ALLEN the robot
(developed by Rodney Brooks in 1980s)
Three levels of behavior:
Layer 1: Avoid contact with
other object (including object
coming at you).
Layer 2: Wander around
aimlessly without hitting
obstacles
Layer 3: Explore “interesting”
places to visit
allen’s subsumption architecture
- Each layer adds
new functionality - Lower layers serve
as foundations for
upper layers
Where is there subsumption architecture in the brain?
Perception-Action Cycles (PACs) in Basal Ganglia
Who made BigDog and when?
Boston Dynamics 2005-2010
Who made atlas and when?
Boston Dynamics 2025
