I think in a social system, it is good that individual can think independently. Good that some outliers that think differently from the main group, as this will prevent groupthink, can discover the problems better, earlier, as the leaders can be wrong, as most people thinking can be wrong.
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://drive.google.com/drive/folders/1BL0HwVcEmVPscuAbdJzvtuhsdpMk2c-J

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

Alex Honnold climbed the Taipei 101 skyscraper. How long will AI humanoid robot takes to achieve this level of skills?
AI robot is advancing at a very fast rate, but certain level of dexterity is hard to achieve. Will it hit a cornerstone, or it will take many years of incremental improvements and a series of major breakthrough to achieve this?
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://drive.google.com/drive/folders/1BL0HwVcEmVPscuAbdJzvtuhsdpMk2c-J

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

In signal processing, the impulse function initially was a finite width and height pulse, to compute the impulse response of a linear time invariant system. In theory this is not beautiful, did not fit well with other signal math, theoretical guys invented the zero width infinite height impulse. Practical experiments discover things that theory cannot, but math theory can discover things that experiments cannot. My deep learning BSnet is convex and has beautiful theory.
https://vixra.org/author/sing_kuang_tan

Crelu activation function breaks an input into 2 channels, one is the positive part, one is the negative part. This is like the split functions in decision tree for regression or classification. So the linear programming formulation of relu function bridges the theory of decision tree or even fuzzy logic with deep learning. Deep learning network can be converted to and from decision tree. kaka.
https://vixra.org/author/sing_kuang_tan
https://singkuangtan.github.io/

https://drive.google.com/drive/folders/1BL0HwVcEmVPscuAbdJzvtuhsdpMk2c-J

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

Vladimir Vapnik said that his Support Vector Machine can be trained with extremely small datapoints of 10 points, 5 for each class. It works because of the max margin property and slack variables.
My BS-Pidnet can be trained on 525 images where Pidnet model needs at least a 1000+ images with many model tuning, because BS-Pidnet is convex. Challenge me if I'm wrong.
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://singkuangtan.github.io/

https://drive.google.com/drive/folders/1BL0HwVcEmVPscuAbdJzvtuhsdpMk2c-J

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

When you look at the poor results of semantic image segmentation model such as DeepLabV3, you notice that it is patchy (small error segments). It is because the model during training learn patch by patch, then slowing merge the patches. If there is not enough data, then it is patchy because the model is non-convex. My BS-Pidnet will learn holistic segments, from blur boundaries to sharp boundaries because it is convex.
https://singkuangtan.github.io/
https://vixra.org/author/sing_kuang_tan
https://drive.google.com/drive/folders/1BL0HwVcEmVPscuAbdJzvtuhsdpMk2c-J

There were some dark histories in United States where the workers went strikes, protests and riots in large numbers and out of control. The US government and the bosses said that they need a discipline workforce, they need to discipline these people.
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://drive.google.com/drive/folders/1BL0HwVcEmVPscuAbdJzvtuhsdpMk2c-J

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

Einstein IQ 160, me IQ 110. But if Einstein worked 4 hours, I did 8 hours, 4*160<8*110, then 640<880. That means I am smarter than Einstein if I worked doubly hard or have better mental stamina. lol.
That means if a company hires average people, but very committed and united, they would have beaten those from Google or Microsoft. And they can use ChatGPT to make up for the lack of skillsets. lol
https://drive.google.com/drive/folders/1BL0HwVcEmVPscuAbdJzvtuhsdpMk2c-J

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

https://singkuangtan.github.io/

邦有道则仕，邦无道则隐。没事我就在家闲着。
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://drive.google.com/drive/folders/1BL0HwVcEmVPscuAbdJzvtuhsdpMk2c-J

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

Vladimir Naumovich Vapnik published his Support Vector Machine paper at age 59. It is never too old to invent! Many said that an invention impact depends on how it is widely used in applications, but theoretical inventions is general enough to last many years or even forever. Look at the VC dim in his Statistical Learning Theory. Compared to Yann Lecun's work doesn't seem to have much theory. LOL.
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://singkuangtan.github.io/

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

One of my greatest achievements is that I developed and tested my BSnet family series of deep learning models without any supercomputer and GPU server. I did my experiments on my GPU laptop and Google Colab, with modifications to do loseless reformat to train on my laptop. I use Math to study the properties of my models, which in many cases experiments cannot tell us. LOL
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

Stalin management series: Stalin once said, one death is a tragedy; a million is a statistic. We see that he had an aggregate economic loss function he tried to optimize, did not care whether each individual was happy, optimised the global most important ones and sacrificed the small part. Russian are good in system science and Stalin showed strong system thinking where I think many modern leaders lack of.

https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://singkuangtan.github.io/

古人说，先立己而立人。LOL
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

My thoughts on underachiever and even low aptitude person. I think there is nothing wrong to be one, as not everyone is so lucky to make it in life. They should have a place and say in this world. As most people are chasing to be like the same superstar achiever, there are no diversity in the talking and thinking in the workplaces and in this world.
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

The world has become more divided. People wanted freedom, liberty and democracy, which will create more self interests, create a wedge that further divides us. We need more love and empathy to bring us together. 💏👩‍❤️‍💋‍👨👩‍❤️‍💋‍👩
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

A technologist need to be good at drawing, imaginative like Leonardo Davinci the artist. He observed, imagined and problem solving with his drawing pen.
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg

The Square–cube law states that as a spider grew beyond a certain size is unable to walk. As weight varies with the volume of the spider (cubic), where as strength is the cross section area (quadratic) of the spider leg muscles. Cubic rate is much faster than quadratic rate, that's why we don't see very big spiders in nature. Similar for Humanoid robots. There is a limit for the strength and reaction speed of servo and bigger robot has higher momentum as in pole balancing problem.

There is never a very large printed analog circuit. But there is a very large digital circuit like our CPU or GPU. So I think the design of Deep Learning network is the same. Since it is so large, it should be Boolean like operations in nature. I think we should not spend too much time designing complex activation functions such as Swish or Softplus functions, but use simple Relu function and make the training convex.
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://singkuangtan.github.io/

Recently when I applied for jobs, when some recruiters knew my age, they ignored me. Mind you, we are in the age of longevity. There are 90+ years old Japanese programmers and 70+ years old wrestling coaches. The modern medical science, sport science and nutrition knowledge are so powerful that most of us have a very long active life that ends close to the day we die. In the past, soccer players retired at 30, but now some elite players at even 37.
https://vixra.org/author/sing_kuang_tan
https://singkuangtan.github.io/

Let me make a futuristic prediction. Large language model becomes so smart that they can create their own hypotheses and design scientific experiments. But it is not good with their hands. So it will hire and instruct humans to do these physical experiments and they key in the results to the model for next moves. LOL
https://www.slideshare.net/SingKuangTan
https://vixra.org/author/sing_kuang_tan
https://github.com/singkuangtan
https://singkuangtan.github.io/

https://www.linkedin.com/in/sing-kuang-tan-b189279/?originalSubdomain=sg