Introduction to Embedded Machine Learning Week 3 Quiz Answers

Question 1) Which of the following are good uses of audio classification? Select all that apply.

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?

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.

MFCCs and CNNs Quiz Answers

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?

Question 2) Why do we use the Mel-frequency cepstral coefficients (MFCCs) as features for audio data?

Question 3) What does a single node in a convolution layer of a convolutional neural network do?

Question 4) Dropout layers are used to reduce underfitting.

Question 5) You plot the training loss over the entire training period, and you discover the graph appears as follows:

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

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?

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?

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.

Question 5) In sensor fusion, all of the sensors must be the same type.

Question 1) Audio classification can be used to detect anomalies in machinery.

Question 2) A sample rate of 20 kHz would require a sampling period of what? Please give your answer in microseconds rounded to the nearest whole number.

Question 3) Your audio classification system samples at a rate of 16 kHz. Your target sounds should be at most what frequency?

Question 4) The collection of slices of Mel-frequency cepstral coefficients (MFCCs) describe how the frequencies in audio data change over time.

Question 5) A convolutional neural network (CNN) uses a fully-connected neural network to perform classification before performing convolution.

Question 6) Convolution layers in a convolutional neural network (CNN) perform feature extraction.

Question 7) You plot the training loss over the entire training period, and you discover the graph appears as follows:

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.

Question 8) You plot the training loss over the entire training period, and you discover the graph appears as follows:

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

Question 9) 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 both false positives and false negatives equally (as best as possible)?

Question 10) Sensor fusion is a form of machine learning.