Week 1

Machine Learning and Limitations

Question 1

Artificial Intelligence (AI) is only concerned with the designing and building of intelligent agents that receive precepts from the environment and discover patterns in that environment.

True
False

Question 2

Classification is what kind of machine learning?

Regression
Reinforcement learning
Supervised learning
Unsupervised learning

Question 3

During the inference phase, the parameters of the rules (or “model”) are updated automatically using one or more mathematical algorithms.

True
False

Question 4

Machine learning models are prone to biases in the data.

True
False

Question 5

Which of the following are good ways to build societal trust into an artificial intelligence (AI) system? Select all that apply.

Robust
Efficiency
Ethical
Lawful

Embedded Machine Learning

Question 1

A microcontroller is a powerful processing unit that relies on off-chip flash and RAM for computations and is often capable of running general-purpose operating systems, like Windows, macOS, and Linux.

True
False

Question 2

If a system is 99% accurate at detecting a spoken word, it is deterministic.

True
False

Question 3

Machine learning can be used to detect anomalies.

True
False

Question 4

All microcontrollers are capable of running machine learning.

True
False

Question 5

TinyML software frameworks optimize machine learning operations to allow embedded devices to run machine learning algorithms more quickly and efficiently.

True
False

Data Collection

Question 1

Which portion of the dataset is used to evaluate the model in order to fine-tune the model’s hyperparameters?

Validation set
Test set
Common set
Training set

Question 2

Which portion of the dataset is put aside until the very end to evaluate the model after training and hyperparameter tuning?

Test set
Training set
Validation set
Common set

Question 3

Which portion of the dataset is used to automatically update the model’s parameters?

Common set
Training set
Test set
Validation set

Question 4

A “naïve classifier” always predicts the same class every time it is run.

True
False

Question 5

What is a “balanced dataset?”

A dataset where the input dimensions of the model are equal to the output dimensions.
A dataset that contains approximately equal number of samples in the training, validation, and test sets.
A dataset that contains samples with the same number of dimensions.
A dataset that contains approximately equal number of samples in each class.

Feature Extraction

Question 1

What is a “feature” with regards to machine learning?

A measurable property of something that’s observed
A plot of the data over time
A spectral analysis used to break apart the data into its various frequency components
A description of the overall shape and trends in the data

Question 2

Feature extraction can be used to reduce the computational complexity of a machine learning model.

True
False

Question 3

What is “inference” with regards to machine learning?

Automatically updating a model’s parameters with data
Analyzing the data to determine which features to use
Reducing the number of dimensions in features to reduce computational complexity.
Using a trained model to make predictions with unseen data

Question 4

Hyperparameters are automatically updated during model training.

True
False

Question 5

You generally want to use different feature extraction methods between training and inference.

True
False

Machine Learning Overview

Question 1

All artificial intelligence (AI) is considered machine learning (ML).

True
False

Question 2

During the training phase, the parameters of the rules (“model”) are updated automatically using one or more mathematical algorithms.

True
False

Question 3

Machine learning can be used to identify objects in images.

True
False

Question 4

Ensuring data integrity and privacy is a key factor in building societal trust of AI systems.

True
False

Question 5

Which of the following would be appropriate uses of embedded machine learning? Select all that apply.

Detecting anomalies in an online banking application
Detecting anomalies in the servomotor of a satellite
Performing an action based on a spoken keyword
Identifying objects in an image

Question 6

The test set is used to evaluate the model in order to fine-tune the model’s hyperparameters.

True
False

Question 7

Using a balanced dataset can help you avoid training a naïve classifier.

True
False

Question 8

Which of the following would be considered a “feature” with regards to machine learning? Select all that apply.

How long to train the model
Raw data value
The root mean square (RMS) value of all the data
Prominent frequency values found in the data

Question 9

The process of extracting features from unseen data is known as “inference.”

True
False

Question 10

Machine learning models can learn biases found in data.

True
False

Week 2

Neural Networks and Training

Question 1

The activation function allows a perceptron (node) to act as a linear regression function.

True
False

Question 2

A single node in a neural network performs the following actions in order:

Multiplies each input value by a weight

Sums the products of step 1

Adds a bias term to the sum computed in step 2

Applies an activation function to the sum from step 3 to produce a single value output

True
False

Question 3

“Backpropagation” is the process of computing a prediction given an input sample and comparing the model’s prediction to that sample’s known label to calculate a loss value.

True
False

Question 4

Model parameters (weights and bias terms) are updated manually by the programmer during training.

True
False

Question 5

Performing forward and backward passes on the model with the entire data set is called what?

An iteration
A mini-batch
Regression
An epoch

Evaluation, Underfitting, and Overfitting

Question 1

You plot the training and validation loss over the training period of the model, and you see the following. What can we conclude about the model?

The model’s input dimensions do not match the data’s dimensions
The model has underfit to the training data
The model has overfit to the training data
The model model has a good fit with the training data

Question 2

Which of the following are good ways to try to reduce overfitting? Select all that apply.

Add regularization terms
Early stopping
Use a more complex model
Gather more data

Deploy Model to Embedded System

Question 1

The features you extracted during the training process must have the same number of dimensions as the features you extract during inference.

True
False

Question 2

Your machine learning model is a neural network that ends with the softmax function, and you have trained it to predict the label among 5 different classes. How many outputs does your model have?

Question 3

After performing inference, the outputs of the softmax function should always add up to 1.0.

True
False

Question 4

After performing inference, your model outputs 5 probabilities, each pertaining to the probability or confidence rating of each class. Finding the lowest probability will give you which class the model is predicting that the data belongs to.

True
False

Anomaly Detection

Question 1

You wish to identify the spoken keyword “Alexa.” An anomaly detection system is a good way to accomplish this task.

True
False

Motion Classification and Anomaly Detection

Question 1

The activation function provides a non-linear output of a single node in a neural network.

True
False

Question 2

Hyperparameters are updated automatically during the training process.

True
False

Question 3

The process of automatically updating a neural network’s parameters based on the calculated loss between the predicted label and actual label is known as a forward pass.

True
False

Question 4

You plot the training and validation accuracy over the training period of the model, and you see the following. What can we conclude about the model?

q6-tt_350_308

The model has overfit to the training data
The model has a good fit with the training data
The model’s input dimensions do not match the data’s dimensions
The model has underfit to the training data

Question 5

You are creating a motion classification system, and your feature extraction method includes calculating the root mean square (RMS) and the frequency value of the 3 highest peaks of the power spectral density (PSD) of each axis on a 3-axis accelerometer from a 10 millisecond sample period. How many dimensions should the input to your model be?

Question 6

You deploy your machine learning model to an embedded system. The model ends with the softmax function, giving you 4 outputs, one for each class. How do you find the predicted class?

Sum the results, which will give you an index to the array of class labels
Choose the class with the highest output value
Average the results together, which will give you an index to the array of class labels
Choose the class with the lowest output value

Question 7

The identification of rare events is known as anomaly detection.

True
False

Question 8

Gathering more data is one way to reduce overfitting.

- True
- False

Week 3

Audio Classification and Sampling Audio Signals

Question 1

Which of the following are good uses of audio classification? Select all that apply.

Translating handwriting to computer-encoded text
Tracking wildlife
Identifying anomalous noises in a home or office
Translating spoken words to computer-encoded text

Question 2

In a digital system, audio signals are sampled at a regular interval. A sampling period of 125 microseconds translates to a sampling rate of what?

44.1 kHz
48 kHz
8 kHz
16 kHz

Question 3

A sampled audio signal with a bit depth of 8 bits has more resolution than an audio signal with a bit depth of 16 bits.

True
False

MFCCs and CNNs

Question 1

You are creating a system to classify audio events whose maximum frequency is 10 kHz. Your sample rate needs to be at least what?

20 kHz
10 kHz
5 kHz
40 kHz

Question 2

Why do we use the Mel-frequency cepstral coefficients (MFCCs) as features for audio data?

They provide data augmentation to help create a more robust audio classifier.
They mimic how the human ear perceives sound.
They mimic how the human vocal cords and mouth produce sound.
They help prevent overfitting.

Question 3

What does a single node in a convolution layer of a convolutional neural network do?

It prevents a random selection of outputs of the previous layer from reaching the next layer.
It combines the outputs of the previous layer such that they each have a value between 0 and 1 and all sum to 1.
It filters the image using a kernel.
It slides a window over the image, selecting the highest pixel value in that window.

Question 4

Dropout layers are used to reduce underfitting.

True
False

Question 5

You plot the training loss over the entire training period, and you discover the graph appears as follows:

5a_185_166

What is the most likely issue for not having a convergent (or near-maximally low) training loss?

Not enough training time
Overfit model
Underfit model
Too much training time

Implementation Strategies

Question 1

You deploy a trained model to your embedded system. The model gives us 4 outputs, each corresponding to the probability of each class. How do you identify the predicted class?

Choose the class with probability over a threshold
Sum the probabilities together and divide by 4
Choose the class with the lowest probability score
Choose the class with the highest probability score

Question 2

You wish to choose a threshold for your classification system, so you create 2 histograms on the same plot: one for the output probabilities of your target class when the input is your target class and another when the input class is not the target class. You get the following:

Which threshold should you choose if you want to minimize false negatives?

2a_424_216

Question 3

You use the same validation dataset to evaluate 4 different models, which produce the ROC curves given below. Which model is likely to give you the best accuracy for your application?

Question 4

A neural network (the ones we’ve looked at so far in the course) is probabilistic.

True
False

Question 5

In sensor fusion, all of the sensors must be the same type.

True
False

Implementation Strategies

Question 1

Audio classification can be used to detect anomalies in machinery.

True
False

Question 2

Your audio classification system samples at a rate of 16 kHz. Your target sounds should be at most what frequency?

16 kHz
4 kHz
8 kHz
32 kHz

Question 3

The collection of slices of Mel-frequency cepstral coefficients (MFCCs) describe how the frequencies in audio data change over time.

True
False

Question 4

A convolutional neural network (CNN) uses a fully-connected neural network to perform classification before performing convolution.

True
False

Question 5

Convolution layers in a convolutional neural network (CNN) perform feature extraction.

True
False

Question 6

You plot the training loss over the entire training period, and you discover the graph appears as follows:

7a_320_293

The model is not as accurate as you require (i.e. it is underfit to the data). Which of the following are likely to help you get a more accurate model? Select all that apply.

Increase model complexity
Gather more data
Train for longer
Early stopping

Question 7

You plot the training loss over the entire training period, and you discover the graph appears as follows:

8b_320_288

What is the most likely issue for having a divergent training loss over time?

Overfit model
Underfit model
Learning rate is too low
Learning rate is too high

Question 8

9c_431_220

Which threshold should you choose if you want to minimize both false positives and false negatives equally (as best as possible)?

Question 9

Sensor fusion is a form of machine learning.

True
False