Silverpond AI Internship Test
Hello! Hola! Kon’nichiwa! Ni-hao! Hallå ! Aloha! !مرحبا
We'd like you to prepare some code in a Jupyter Notebook that explains a concept in machine learning/AI/deep learning that is of interest to you. You may use TensorFlow, PyTorch, or the framework of your choice; as long as it is in notebook form. We're assuming you know some Python and how to make your way around in a notebook; if not, it's worth learning!
As a team we're particularly interested in, and will be looking for:
- Good explanations,
- Modular and Idiomatic code,
- Clear goals,
- Tying-back final output to original goals,
- Different viewpoints,
- Fundamental understanding,
- Visual explanations,
- Awareness of knowledge boundaries ("What do I know about how much I know!").
As a reference, your notebook need not be any longer than the one that we use for the prerequisite for the deep learning workshop:
We would like this notebook to be hosted on your own GitHub (or other public git source control system; we want to know you can use Git!). Don't worry if it's the only thing on your account!
Please try not to spend more than 6 hours (we'd aim for 2-4 hours); so either try and explain well a concept you're already familiar with, or perform a brief investigation into the open areas of an interesting technique! If you get stuck, documenting your current understanding and future directions of investigation would be a nice way to conclude. Also, feel free to get in contact at any time to chat through thoughts; it's encouraged!
Potential example projects:
Here are some ideas that should indicate the general direction we're interested in. Please feel free to come up with your own!
-
Playing with convolutions: A notebook that looks at convolutions, purely independently of deep learning, and tests a few different varieties on images.
-
Algebra on auto-encoders: A notebook that loads up an existing auto-encoder model and performs some algebra, on the latent vectors, within it, to demonstrate what it's capabilities are.
-
t-SNE: A notebook that takes a model that produces some latent vector
z, and performs t-SNE on it, then explains what the t-SNE is doing. -
Bayesian probability: A notebook that demonstrates how Bayesian probability works, and some examples of prior-updates, and computations of the posteriors in a few different scenarios.
-
Bias in ML: A notebook that explains how bias can be found in ML models, and highlights a few ways that people are attempting to approach this problem.
-
Precision scores: An overview of the different measures that people use to assess the capability of different ML models.
-
An Exploration of the TensorFlow Tensor: A review of what Tensors are, in the context of TensorFlow, and the various operations that can be performed on them, and what their affect is.
-
Learning about Fully Connected Networks: A tour through your own learning of how to build a few fully-connected networks in PyTorch or TensorFlow; perhaps with some investigation into how many parameters they have as the input grows, or similar visual explanation of what is going on under the hood.
-
TensorBoard Visualisation: An example of using TensorBoard for visualising different graphs, images, and perhaps even some potential embeddings! Doesn't neccessarily use a real neural network, just demonstrates using a few features of TensorBoard itself.
Notes
If you wish, you could configure your repo so that it can be executed in the MyBinder.org environment; but this isn't necessary!
Fun Datasets
I personally quite like "Fashion-MNIST", if you wanted to play around with that:
Magenta is a another interesting place to look, but can be quite involved to get up and going with:
The TensorFlow Playground also contains some fun inspiration:
CIFAR-10 is a classic image dataset that could be fun to explore:
And if that's not enough, consider the Computer Vision Online list of datasets:
If you find other interesting ones; let us know!