I started studying a study machine for more than six years, the field was between money. In 2018-Issh, when I took my first university studies to ancient machine learning, after the scenes, the main methods were already inventions that would lead to AWA's earning early 2020. GPT models are published, and some companies are followed by a suit, press the restrictions, both in the size of the Performance and Parameter, with its models. For me, it was a good time to start reading a machine study, because the territory was so fast that something new.
Occasionally, usually every 6 to 12 months, looking back years, I look forward to the university's schooling talks. Looking back, I often find new principles accompanied during ML study. In the review, I found that working deeply was one small article into a key goal of my advancement over the years. Without a deep work, I have identified the rest of the three principles. It is not actually understanding of technology, but instead of the mindset patterns and methods.
The Importance of Deep Work
Winston Churchill is not only crying but also his good mind. It is a popular story about the words conflict between Him and Lady Ator, the first woman in the British parliament. Trying to finish the argument with her, turned off:
If I were your wife, I put the poison in your tea.
Churchill, by his sharp mark, answered:
And if I were your husband, I would drink it.
Giving the Witty Reporte like that is popular because it is an extraordinary skill, and not everyone is born with such intelligent intelligence. Fortunately, in our domain, you are doing ML research and engineering, a fast wit is not a big power to find away. What is the strength to focus on.
Mechanical learning, especially the research side, does not automatically keep in a native sense. It requires a long-term stretch of thought that is not disturbed, much. To enter coding algorithms codes, material debugging conversations, designing hypothesis – everything requires a deep work.
With a deep work, “I mean:
- Focus Deeply for a long time
- Nature that allows and encourage such focus
Three years ago, I saw a serious job in making a reasonable development. The hours I have used in the focus of focused on – a few times a week – they produce more than the most disturbed production. Also, thanks, deeply work can be read, and your environment is set to support them.
To me, the most fulfillment seasons will always lead to finding papers. These are times where you can focus on Laser: The world crosses from your project, and you are flowing. Richard Feynman said planned:
Making a good physics, you need a strong length of time … requires a lot of focus.
Replace “herght“With”Machine reading“And the point is still holding.
Should (mainly) ignore styles
Have you ever heard of big languages of languages? Of course, No – words like Llama, Gemini, Claude, or Bard Fill in the Tech news cycle. They are the cool kids of the productive AI, or “Genai,” as it is now known as stylight.
But here is a catch: If you just start, chasing a tendency can make a hard break.
I once worked with a researcher, and both of us were just beginning to “do ML”. We will call my colleague who had been a former employee. According to his research, he is first looking at the new Retral-Activated Generatisi (RAG) meeting, hoping to improve the language model to combine foreign documents. He also sought analyzing of evolving skills of llms – things that these types can do or are not clearly trained – and clearly distinguish those small models.
The problem in John? His work-based models appeared as soon as possible. Just finding a new Running Model working on weeks. By the time he did, the new, best model was published. The speed of change, combined with unclear test procedures for his niche, made it possible to be aware of their research. Especially for a young person to research, like John and me back then.
This is not a john critic (likely to have failed too much). Instead, I tell the story to make you think: Does your progress depend on the processing and using the best wave of the latest culture?
To make boring data analysis (over and above)
Every time I get to train the model, I breathe murmuring the relief.
Why? Because it means I'm done with a hidden part: Data analysis.
Here is the standard sequence:
- You have a project.
- You get some data (real world).
- You want to train ML models.
- But first … you need to fix data.
A lot It can be fine in that last time.
Let me show this accident when I worked while working with the ERA5 weather data – a large, colored dataset from the European weather forecast. I wanted to predict NTHVI (Index of the general variants, which indicates the plants of vegetation, using historical climate patterns from eRA5 data.
In my project, I had to meet the ERA5 weather data with NDVI satellite Data I received from NOAA, US Weather Agency. I am translating NDVI data to the ERA5 decision, added as another layer, and, unacceptable, gladly began to train Vision Transformer view.
A few days later, I saw in paper forecasts and … It is surprising! The thought of the world was down. Literally – My installation data displayed the most familiar world, but my vegetable data became packed.
What went wrong? I ignored the translation of the decision to underline NTHVI data direction.
Why do I miss that? Simple: I don't want to make data engineering, but right forward to the center learning. But the truth is: In the real ML ML activity, to find the right data by work.
Yes, educational research often allow you to work with selected information such as ImageNet, CIFAR, or Squad. But in real projects? You will need:
- Clean, mix, select normal, and confirm
- Charges on the edge of the edge
- In view of checking the middle data
Then repeat this until it is really ready
I read this in a difficult way by skip the steps I think is not necessary for my data. Don't do the same.
(Machine-reading) some temptation of temptation and error
From the External, Science Progress Always looks smoothly smooth:
Problem → Hypothesis → Test → Solution
But in work, it is a great miracle. You will make mistakes – some small, right for others. (eg the world is missing down.) OK. What is important is how you treat those mistakes.
Bad mistakes just appear. But the decisions of understanding teach something.
To help me learn quickly from the thought-out failure, now I keep the writing book for simple lab. Before conducting examination, writing down:
- My hypothesis
- What I expect is to happen
- Why am I expecting
Then, when the test results return to (often as “Nope, it did not work”), I could imagine why it is said why that means why it means my thinking.
This changes the errors into an answer, and the answer in reading. As the saying moves:
The scholar is someone who made every mistakes in the smallest field.
That study.
The last thoughts
After 6 years, I have realized that making a well-therapy is so small to do with burning conditions or models (a large language). By looking back, I think it's more about:
- Creating time and place of deep work
- Choosing the depth over the hype
- Taking data analysis seriously
- Accept the correction and error
As long as you start – or even a few years in – these courses are worth additions. They will not show in conference departments, but they will appear with your true advances.
- Feynman quote from a book Deep workWith CAL Newport
- For Churchill's average, many different variations are, some have a coffee, some have tea, poisonous
