Flashcards
LH - Classification
Flashcards do conjunto (21)
- 1Whats a neuron? Name the five different parts.Dendrites, Axon, Soma (cell body), myelin sheath, synaptic terminals.
Neuron is a specialized cell found within the nervous system that trasmits impulses - moves information around the body at pace. - 2"Neuronal Selectivity"? What's an example?If a neuron is selective it fires more to certain incoming signals (if sensory neuron) than it does to other incoming signals.
If it is a motor neuron then it fires more in association to certain motor behaviors, for example the movement of my hand differs from movement of hand in different direction. High firing rate is associated with some specific incoming sensory stimulus or specific outgoing voluntary behaviour. - 3Sceanrio of Neuronal selectivity - auditory..Hypothetical neuron in primary auditary area of brain - responded 10 action potential/second at acoustic stimulus. If that changed- went up to 600Hz, didnt fire - band pass filter - respond selectivley to a band of stimulus.
- 4Is motor cortex rostral or caudal of visual cortex?Rostral - towards beak Caudal-towards tail (cord). Leftside lateral views. It was rostral.
- 5Classification definitionClassification is a statistical process whereby observations
are organized into categories (or classes).
Classification can also be used for prediction/decoding —
to assign previously unseen observations to categories - 6The typical pattern recognition process has 5 stages, name them.
- 7feature, (extraction)In machine learning, a feature is a computational outcome.
A feature characterizes an input by deriving a measure. - 8Q. In assignment 1, what’s the input? How is that input sensed? And then how are the data segmented?Input EEG singals (biopotential on scalp), measure using electrodes on the head, segment into constituent 5 sec blocks, each traial/segment having an eyes open or closed label.
- 9Q. In assignment 1, how many features are we computing? What are those features? Give me an alternative — what else could we have computed?2 features - for each 5 sec long trials - perform spectral analysis, get power in alpha and beta bands (2 features). Alternative would be total power (across spectrum). Point is theres an infinite number of things to compute -
- 10Problem: It’s rarely clear which features should be computed.Feature extraction is usually domain-dependent, relying on
specialist knowledge about what’s worth computing. - 11Q. In assignment 1, how do we know what features to compute?Previous research/literature.. recall they made observations of alpha waves..something to do with attention, eyes open/closed. Develop domain expertise informs us what features to compute.
- 12Q. In assignment 1, what’s an ‘irrelevant transformation of the input’?- We normalize the data - doesnt effective the computation but could say normalization is a transformation (of the input) that doesnt effective the behaviour of the classifier.
- 13Ideally, computed features are .... for inputs of the same category and .... for those of different categories.similar, different
We
want distinguishing features that are invariant to irrelevant
transformations of the input. - 142-dimensional feature vector, xx = [P8 P11.2]
- 15The abstraction provided by the feature extraction allowsthe
development of classifiers that are usually domainin dependent. - 16Q.. Why is classification described as a statistical process?Measurements are subject to noise - need to make classifiers robust to noise.
- 17Classifiers rarely perform perfectly (i.e., 100% accuracy). Why?That’s partly due to feature variability, and partly due to the
choice of classifier (e.g., linear versus quadratic). - 18Feature variability:Is the difference between features within
a category small, and the difference between features
across categories large? If so, expect high-accuracy
classification. - 19decision boundaryA decision boundary is a mathematically defined criterion:
Points falling on one side of the boundary are classified as
belonging to one category, and those falling on the other
side are classified as belonging to the other category.
Obviously, the boundary takes the familiar form:
ax + by + c = 0
where a, b and c are optimized parameters, and x and y are
the axes of our 2-dim. feature space. - 20Calculating the decision boundary is a 2-step process. (1st step)First, we the project the data onto a line defined by vector w.
We vary the direction of w to find the angle that maximizes
the separation of the projected category means (m1 and m2)
and minimizes the within-category variance of the projected
data. - 21(2nd step) ijn “Fisher’s linear discriminant”.Second, we choose a point on the line
defined by w. The line perpendicular to w
which intersects this point is the decision
boundary.
Informações
- Descrição
- BIOMECH - PART 2 - LUKE FlashCards sobre LH - Classification, criado por Amber Turner em 08-11-2020.
- Criado por
- Amber Turner
- Visualizações
- 0
- Avaliação