Cognitive Science 1 - Lec 7

Computational Neuroscience

Based on modeling biological neurons and populations of neurons

Constructs models by abstracting away from some biological details while preserving others

Connectionist Modelers

Aim to stimulate and reproduce well-documented psychological phenomena like a child’s pattern of development or ways cognitive processes break down in brain-damaged patients

86 billion

How many neurons the human brain has

Artificial Neural Networks (ANN)

a computer program that mimics the human brain's network of neurons.

used to solve problems, make decisions, and learn from experience with an organized layers of nodes/neurons

Weights

Determine signal strength of node (positive = excitatory, negative = inhibitory)

The Part of ANNs that are attached to the connections between pairs of units in adjacent layers to determine the behavior of the network

Similar to how excitatory and inhibitory neural connection of various strengths connect to a particular neuron in human neural networks

Bias term

Adjusts activation threshold/Indicates what weighted sum needs to be before node/neuron activates

Similar to threshold necessary for activation of a neuron in human neural networks

Input Layer

First part of ANN processing of an image

Pixels have values between 0 (back) and 1 (white) for its intensity in a grayscale

Each pixel corresponds to a neuron that holds that grayscale value as its activation level

First hidden layer (edge detection)

Neurons recognize edges or basic shapes & calculates a weighted sum with the assigned weights of the input layer pixel

Bias is added to adjust when a neuron activates (ex: only if the sum exceeds 10)

Second hidden layer (feature assembly)

Neurons recognize complex patterns by combining edges

Ex: detect lines or curves that distinguish “9” from “3”

Output layer (digit prediction)

Last ANN layer

Each neuron outputs a number 0-1 for the confidence level of each digit and the highest activation neuron determines the predicted digit

Backpropagation

Learning algorithm that fine-tunes weights and biases

• compares predictions with actual labels, calculates errors, and updates weights accordingly

• Overtime network learns to correctly classify different images

Multilayer neural network

Network’s knowledge is distributed across the relative strengths of the connections between different units and is constantly changing rather than being encoded in discrete symbol structures

Flow of information is determined by what happens in all of the units in a given layer

Suitable for pattern recognition tasks like image classification

Single-layer network

Simplest form of a neural network w/ one layer of artificial neurons

Works for linearly separable data (ex: given sentence A and B, the only circumstance which A and B is true is if both A and B are true) and can use traditional programming

Hebbian Learning

Donald Hebb (1949) speculated that learning takes place in the brain through an associative process

“neurons that fire together wire together”

If two neurons fire simultaneously, their interconnecting weights will be reinforced

Can engage in unsupervised learning

Perception Convergence Rule

Frank Rosenblatt (1958) came up with learning rule that would allow a single layer network or perceptron w/ random weights and thresholds that would allow it to produce the right output for every input.

Involved supervised learning

Underfitting

Network tries to fit a function to the data but the function is not complex enough to correctly represent the data

Optimum

Network fit function that has the appropriate complexity to accurately represent the data and generalize since it learned the trend the data follows

Overfitting

Network’s function fits data but hasn’t learned the trend and it unable to generalize to new data

Autoencoder

Learns to produce a simplified representation of the input

Ex: in computer visual processing, it learns to encode images as complexes of edges

• Can improve performance on different types of tasks like classification

• Compresses then decompresses an input signal

Convolutional neural network (AKA CNNs or ConvNets)

An artificial system that performs complex visual information processing that humans do every second

Used for natural language processing, facial recognition software, self-driving cars, SuperVision program to identify things in a box

Creates filter that functions as a localized feature detector

• filters out everything except feature its trained to detect

Small-world network

Any network where one can get from any single point to any other point only in a small number of steps even though the total number of elements may be huge

“Six-degrees of separation”

Allows efficient information transfer and neuroplasticity

Six-world network

All people are six, or fewer, social connections away from each other

Supported by research from Duncan Watt

Bayesian Probability

Probability is interpreted as reasonable expectation representing a state of knowledge or as quantification of a personal belief

Updates probability of an event based on new evidence

Fuzzy Logic

Probabilistic approach to logic where truth values are on a spectrum (0 to 1) instead of binary (True/False)

• Ex: 0.7 = mostly true; 0.3 = mostly false

• Used in AI, robotics, and expert systems to model uncertain environments

Ex: operation consists of taking the lowest rating among all variables:

− Job 1 quality = 0.4

− Job 2 quality = 0.2

− Job 3 quality = 0.3

➜ Choice of best job = Job 1