Machine Learning- Ritter Flashcards
which of the following statements on computer learning is/are correct?
a. Machine learning uses rule-based learning whereas in AI learning occurs without being explicitly programmed
b. Deep-learning is a part of AI which is a part of machine learning
c. AI uses rule-based learning, whereas in machine learning, learning occurs without explicit programming.
d. AI is the base for machine learning which is the base for deep learning
e. deep-learning is based on end-to-end learning
c, d, e:
c- AI uses rule-based learning, whereas in machine learning, learning occurs without explicit programming.
d- AI is the base for machine learning which is the base for deep learning
e- deep-learning is based on end-to-end learning
what are the differences between supervised, unsupervised and reinforcement learning?
All are different frameworks of ML;
supervised learning - learning with labelled training set
unsupervised learning - finding patterns in unlabelled data
reinforcement learning - learning based on feedback
which if the following statements is/are correct? Why?
a. if one wants to classify different types of cats without specific characteristics of each cat type but only the general classification of cats one should use unsupervised learning as it is a fast and efficient approach for detecting variability within data classes
b. if one wants to classify different types of cats one should use supervised learning, and train the network to recognise all cat species (with labels) as it is a more specific approach for identifying different structures
a.
Unsupervised learning is trained to find patterns in unlabelled data.
For classification of different types of cats, the network should not necessarily be trained to identify the exact type of cat, but rather it should recognise the general class (= cats) and detect patterns and variability within the class. Therefore unsupervised learning is more efficient in this case.
what type of models are neural networks?
a. discriminative models
b. generative models
c. could be either discriminative or generative (depending on the research question)
d. neither discriminative nor generative
a. discriminative models
what are discriminative models?
a. modeling of p(x,y) with x = features and y = labels
b. modeling of p(x,y) with x = labels and y = features
c. modeling of p(x|y) with x = features and y = labels
d. modeling of p(x|y) with x = labels and y = features
c. modeling of p(x|y) with x = features and y = labels
what is the difference between deep-learning (DL) and machine-learning (ML) in terms of features?
a. ML predefines hand-crafted features and extracts them from existing data whereas DL does not extract prior features but rather uses the data to create its own representation of data
b. DL predefines hand-crafted features and extracts the from existing data whereas ML does not extract prior features but rather uses the data to create its own representation of data
c. ML can either predefine and extract features (supervised learning) or not (unsupervised learning) whereas DL never predefines features
d. ML can either predefine and extract features (supervised learning) or not (unsupervised learning) whereas DL uses reinforcement learning
a. ML predefines hand-crafted features and extracts the from existing data whereas DL does not extract prior features but rather uses the data to create its own representation of data
Define an artificial neural network according to a biologocal neuron model - what are the neuronal features mutual for both?
a neuron can receive one or multiple inputs from different neurons in its dendrites –> the information is then summated and processed in the soma, and different inputs receive different weights based on anatomical and physiological characteristics which sends a signal along the axon and the output/response is generated at the synapse
In a neural network, different inputs (x1….xm) receive ‘synaptic weights’ which refer to the strength of a connection between two nodes. the inputs and the weights are all summated and processed in a joining node which determines the bias of the future output towards some inputs. The whole signal is then sent to the activation function which defines the output of the node given an input or set of inputs. the activation function can be compared to a synapse, which generates an output based on the input processed in the soma. Finally, the activation function generates an output.
a multilayer perceptron (MLP) consists of an input layer, and output layer and ….
a. hidden layer(s)
b. additional layer(s)
c. multiprocessing layer(s)
d. sub-input layer(s)
e. a and c are correct
a. hidden layer(s)
what is the role of hidden layers in MLPs?
perform nonlinear transformations of the inputs entered into the network
what type of supervised learning do MLPs utilise and what is it good for?
backpropagation; used for training feedforward neural networks/ MLPs
potential solutions for the sample size problem….
a. data augmentation
b. generation of heatmaps based on learning
c. sub-group specific analysis
d. feature extraction
a. data augmentation
potential solutions for the black box problem….
a. data augmentation
b. generation of heatmaps based on learning
c. sub-group specific analysis
d. feature extraction
b. generation of heatmaps based on learning
reproducibility/benchmarking is a problem in machine learning and neuroimagong because of…
a. the use of non-representitive data in addition to small sample size and publication bias leads to publication of biased results that are hard to replicate
b. it is hard to know whether a neural network recognises the features it is trained to detect which makes it hard to determine actual characteristics of pathologies
c. data quality is variable which makes it difficult to replicate studies and findings
d. reproducibility is not a problem of machine learning in neuroimaging
a. the use of non-representitive data in addition to small sample size and publication bias leads to publication of biased results that are hard to replicate
