Artificial Intelligence #11: Is maths needed for AI? What many developers miss about maths and AI
Welcome to issue #11
We crossed 13K subscribers in about two months!
Thanks for your support
This week, I saw a post from Prof. Dr. Patrick Glauner where he disagreed with a view posted by a developer which said “You don’t need maths for machine learning”
I agree with Dr Glauner in that ‘not needing maths for AI’ is a dangerously short-sighted approach.
While we see such blogs trying to get traction, they create a false dichotomy (coding vs maths)
Courses at Oxford, Cambridge, IIT, Technion, MIT, Tsinghua typically emphasise maths.
However, a vast majority of courses – especially industry led courses – do not.
This gives many developers a myopic view that maths is not needed for AI
To put this in perspective, many AI (ML or DL) applications can indeed be built from development skills alone and do not need you to know maths.
Hence, if you define your job in AI as purely writing code, data analysis etc – you do not need maths knowledge.
But therein lies the limitations of this approach – because a broader vision for AI exists
Here are some reasons why you need to know maths if you want to learn AI
1)????IPR: If your role involves creating intellectual property, you often need to understand the mechanics of how things work. These may include maths. For example, you may take an open-source version of an algorithm and fork it to create your own customised version. That would need you to understand the workings of the algorithm (instead of just the API).
领英推荐
2)????Design of Experiments: Most business applications start with existing data and from subject matter experts. But scientific and research applications start from first principles i.e. Design of experiments. PhD level research would also start from first principles. But business experts rely on data that they can get hold of and / or a subject matter expert i.e they start at a much higher level. Design of experiments needs understanding of stats and maths
3)????Seduced by big data not aware of small data: Most business applications are ‘big data’ oriented. That’s fine – but there are many domains where datasets are much smaller (ex: bioinformatics, cybersecurity, medicine etc).?This situation would need you to understand additional statistical techniques to handle small data.
4)????Interviews: while I personally do not agree with this, some interviews could include maths
5)????Learn new algorithms: AI is expanding very fast. If you want to learn new algorithms, you won’t get far without the data.?My favourite book on AI is heavily based on maths?
6)????Research papers:?If you are working with complex problems in AI, you need to understand research papers. That may need maths. I wrote about a tool I found called labml.ai which helps in analysing AI papers
7)????To remain relevant: finally, data scientist jobs will be automated using a variety of strategies automl, cloud, low code etc. The deeper your understanding, the more commercially valuable you can be. Most data scientists are very familiar with one algorithm based on their job. For example, if you are working for a fintech firm involved in a price comparison site, you are likely to be working with recommendation algorithms or a classifier like xgboost. Knowing maths, allows you to understand common elements of many algorithms and to reapply knowledge in different contexts. As an example, while not directly related, Nir Regev posted about how he used the CRT (Chinese Reminder Theorem) – created from the times of Sun Tzu – to studying doppler effects in radar.
I use maths in my teaching at #universityofoxford – Artificial intelligence: cloud and edge implementations
The good news is: There are only four main ideas you need from maths to understand machine learning and deep learning:
Hence, its actually possible to teach the maths of data science which I teach as a set
of modules. I am working on a book for the same. But its taken me a long time!
Jobs
Finally, the UK serial entrepreneur Nick Halstead is looking for a few positions in his new company infosum
Image source: pixabay
Founding Data Scientist & Lead @ Minfy | Plaksha University | Indian Insititute of Information Technology
3 年The underlying and fundamental principles of science are from mathematics. I am a student switching from mechanical engineering to data science and ML and as a part of the preparation, I am currently learning and revisiting concepts from probability, linear algebra, statistics and multivariant calculus. It is tough sometimes, but there is joy when you start from the ground up. Much of the courses available today in this domain do not focus on math. Also glad to hear you are preparing a course on maths to bundle all these domains. I think that is absolutely amazing. God bless you for your contributions to the world!
PhD | Professor | Data Science | Machine Learning | Deputy Dean (Research)
3 年In addition to the four main ideas you need from maths, I would like to include a bit of calculus, especially differential calculus.
Global Technology & Business Executive | Digitalization & Transformation Expert Across Multiple Verticals | Talent/D&I Leadership, Mentor & Coach | Board and C-Suite Tech Advisor | Trusted Advisor & Board Member |
3 年As usual a great/informative and balanced exposition . I couldn't agree more. Alan Kay said ' context adds 80 IQ points'. I think trying to be at the deep end of the AI pool without serious math depth leaves you without valuable context. Having said that over time more and more AI will be leogized but that won't be at the leading edge or the real creative deep end of the pool. I do think you will also see the occasional infusion of anthropology into this space. The value of AI creating new questions from scale data not just answers will be huge. That may require an outside-in contextual and cultural view...that's the domain of anthropologists.
@Knowledge Engineering is Evolutionary&
3 年Mathematics is the foundation upon which Science & Technology is founded from the beginning of time to now such as the launching of a rocket to Space requires the escape velocity to effectively leave the earth's gravitational field, the table that is at upon to eat is ergonomically designed with measurements in design. With Mathematics, there is precision and accuracy synergized with metrics. Metrics allows you to measure the performance of systems such as in Big O Notation, etc. false positive rates, pixels, and so on. Machine Learning is built on mathematical foundations which enables designers, engineers, scientists to optimize applications.