Some suggestions about learning machine learning algorithms (advice)

• Do not waste a lot of time in one-time get to know the theoretical understanding
• Do not try to stop and understand all the knowledge points
• Do not waste time: please learn quickly acquire knowledge points, 4-5 day of school.
• Please run the code, "real" to run the code. Do not go in-depth learning theory. Break the code to see if they "eat what spit out what"
• Select an item, do a good job, so awesome!
• ** If you get stuck, do not stop to dig deep, move on!

 

• If you are unsure what better learning rate, each try.
• When you deploy the model into a production environment, you will choose to make predictions with CPU, unless massive scale service request will be used GPU.
• Most companies waste a lot of time collecting more data. The correct approach is to use a handful of data that running watch, and then see if the problem is not enough data.
• If you think you are "born not good at math," please take a look at the TED Talks Rachel:  There's NO SUCH AS Thing "not the Math the Person A" 8 My own the INPUT:. Only 6 minutes, the Everyone Should Watch IT!
• When you use a data set, be sure to give honor and thanks to the data creator
• Remember to use the mean and variance when pretrained training model used.
• To the question "reason to try to stop us from 64x64 to 128x128 to 256x256 size of images and other training to fine-tune the model?" The answer: "Yes, well worth a try good results, try it!!"
• If you do that NLP, there is no reason not to participate in training validation set.
• "In 10% of what you will not use neural networks"? "You may have tried, random forest and neural networks."
• Accustomed to using these words (parameters parameters, layer, layers, activation activations ... etc) and try to use accurately.
• "A great potential for research in various areas of research data to enhance almost no attention in this direction, and I think this is tremendous opportunity to help save 5-10 times more data needs."
• If you take the time to run the entire Nb, including the convolution kernel and the heatmap section, try the experiment code to see their reaction. The most important thing to remember: shape is here that the tensor rank or dimensions. Try to think: "Why?" Look back at the model summary, setting each layer, various maps made, imagine the following mechanism behind it.
• Watching classic paper, thesis all about derivations / theorems / lemmas (part mathematical formula) can be ignored, because they can not help increase the depth of understanding of the practical learning. Read about why and how to define and solve problems. Write a summary to help those people and your status level six months ago similar.
• Perhaps the most important thing is to learn together and small partners, so the effect is often better. Set up a book club, study groups meet regularly, do hands-on projects. What does not need to stick something special, as long as it is to make the world a better place, or even let your 2-year-old children happy. Accomplish one thing, and then to further improve it.
• Learning deep learning, the most important is the code, stop writing code to see your input value, output value, you try to output a mini-batch.