If you are wondering why you should take a deep learning test, the answer is simple: skills matter. By taking the test, you will:
- know where you stand in the community of AI practitioners
- review your performance to learn about your strengths and weaknesses
- access personalized study plan to prepare for interviews
- fast track to job opportunities within our network
- get a certificate
And, it’s free. Your results are only ever shared with your permission to refer you to a company. Let’s go over the deep learning test.
I What is the deep learning test
The deep learning test is one of six standardized tests that were developed by a team of AI and assessment experts at Workera to evaluate the skills of people working as a Deep Learning Engineer (DLE), Deep Learning Researcher (DLR) or Software Engineer-Deep Learning (SE-DL). It is comprised of multiple choice questions selected from a large database, so that different test takers get different questions, and takes 8 minutes to complete.
II What to expect in the deep learning test
Before taking a test, it is important to understand what it evaluates and how it is graded. The grading rubric for the deep learning test includes four categories:
- Understanding deep learning foundations, which covers concepts useful to implement a simple neural network including neurons, fully-connected layers, deep neural networks, forward propagation and backpropagation, and activations functions.
- Understanding techniques to improve neural networks, which encompasses tactics to initialize, optimize, regularize, and select deep learning models. This includes techniques such as dropout, Xavier initialization, transfer learning, vanishing and exploding gradients as well as optimizers commonly used in deep learning (e.g., stochastic gradient descent, Momentum, RMSprop, or Adam).
- Understanding convolutional neural networks, which covers objects and methods used to implement convolutional neural networks (CNNs). This includes convolutional and maxpool layers, classic CNNs (e.g., LeNet, AlexNet, ResNet, inception networks, etc.), and miscellaneous computer vision techniques.
- Understanding sequence models, which spans objects and methods used to implement sequence models. This includes recurrent neural networks (RNNs), their cells (e.g., LSTM and GRU), and miscellaneous techniques used in natural language processing, speech recognition, or time-series analysis.
You will be evaluated and assigned to a skill level in each category: beginning, developing, or accomplished, depending on your mastery of the skill at hand. Your skill level in deep learning will be determined using a combination of your scores across all four categories.
You can learn about the categories and performance levels in the table below.
|Understanding deep learning foundations||Demonstrates limited understanding foundational deep learning concepts.||Demonstrates ability to implement fully-connected neural networks to solve a classification or regression problem.||Demonstrates ability to implement fully-connect neural networks, and identify their use cases and shortcomings.|
|Understanding techniques to improve neural networks||Demonstrates limited understanding of strategies to improve neural networks.||Demonstrates ability to understand techniques used to improve neural networks with some effectiveness. This includes dropout, stochastic gradient descent, weight initialization, and transfer learning.||Demonstrates mastery of techniques to improve neural networks and solve problems such as overfitting, vanishing/exploding gradients, lack of convergence, and slow training.|
|Understanding convolutional neural networks||Demonstrates limited understanding of how CNNs work.||Demonstrates ability to understand the components of a CNN and how to train it.||Demonstrates ability to set up, train and apply CNNs to computer vision problems.|
|Understanding sequence models||Demonstrates limited understanding of how sequence models work.||Demonstrates ability to understand RNNs and how to train them.||Demonstrates ability to set up, train and apply RNNs to use cases such as time-series modeling, natural language processing, or speech recognition.|
At the end of the test, you’ll see your overall skill category in deep learning.
You will also receive feedback for every skills evaluated (e.g., Backward propagating through a ReLU function or Understanding the forward propagation in a convolution layer).
III Deep learning practice questions
Nothing beats practice! Here are examples of questions you might encounter in the deep learning test. Think carefully before selecting your answer. Then, click submit to see the answer and get feedback.
Question 1: fully-connected neural networks
Question 2: activation functions
Question 3: batch size
Question 4: overfitting
Question 5: sequence tagging
IV Tips for the deep learning test
Now that you know what to expect in our deep learning test, it’s time to take it! You can take the test up to three times in a 90-day period (unless the test is being administered to you by a company for a job) and your results are only ever shared with your permission. The first test is simply meant to act as a baseline to show you where to start studying. So why wait? Sign up here to take the deep learning test.
Which, if any, of the following propositions is true about fully-connected neural networks (FCNN)?
In a FCNN, there are connections between neurons of a same layer.
In a FCNN, the most common weight initialization scheme is the zero initialization, because it leads to faster and more robust training.
A FCNN with only linear activations is a linear network.
None of the above.
You’re building a fully connected network to classify all animals on images taken in a zoo. Here are some examples of images in your dataset$:$
If there are bears and iguanas in the image, your network should classify the image as containing two classes$:$ “bear” and “iguana”, no matter how many animals from each class there is. What is a good choice for the last activation of your neural network?
Assume that your machine has a large enough RAM dedicated to training neural networks. Compared to using stochastic gradient descent for your optimization, choosing a batch size that fits your RAM will lead to$:$
a more precise but slower update.
a more precise and faster update.
a less precise but faster update.
a less precise and slower update.
Which of the following methods DOES NOT prevent a model from overfitting to the training set?
Text sequence tagging is the task to output a class (or "tag") for each word of an input text sequence. For instance, the input "I want to go to Burkina Faso" can result in the following prediction$:$ "O O O O O B-LOC I-LOC" where O indicates that the word is not a location, B-LOC (res. I-LOC) indicates that the word is the beginning (resp. inside) word of a location.
You are discussing three possible approaches with your teammates$:$ fully-connected neural networks (FCNN), recurrent neural networks (RNN) and 1-D convolutional neural networks (CNN).
Which of the following is FALSE?
At test time, the CNN will probably be faster than the RNN because it can process the input sequence in parallel.
If you are using GPUs, the CNN will probably be faster than the RNN because GPUs optimize convolution operations.
If the window size of the CNN is small (let's say 3), the FCNN will likely perform better than the CNN on long sequences such as "I am not sure I am available this summer, but I hope I could go to Kuala Lampur".
During training, CNN will probably be faster than the RNN because it can process the input sequence in parallel.