A computer program is said to learn from experience with respect to a class of tasks and a performance measure
P, if its performance at tasks as measured by
P improves with experience beyond a baseline of accuracy defined by: "Guessing the most frequently occurring outcome."
3Blue1Brown, how trained neural networks simulate human abilities: https://www.youtube.com/watch?v=aircAruvnKk
CGP Grey, how the genetic algorithm can simulate human abilities: https://www.youtube.com/watch?v=R9OHn5ZF4Uo
Machine Learning is 90% just a matter of being good at programming and being able to reason about Calculus and recreating those equations in software. Machine learning principles have been well-known since the 1960's. The difference today is that now a budget computer with $2500 worth of NVidia GPU's (or cheap EC2 instance) gets you more FLOPS: 32-bit floating point operations per second than ten million dollars worth of computer in 1995. In the last 20 years, computers are outperforming humans (`1*10^18` signal-processing neurons plus 100TB of bio-memory) at narrowly defined tasks. Today in 2019, a $2700 macbook can outperform the best humans at Chess/Go, do wordplay understanding such as Jeopardy and even chaos-combat Strategy games like Starcraft. Soon image recognition tasks such as Piloting vehicles will be something $2700 worth of computer can do better than any human. Musk has Level 5 autonomy in his flagship product (https://www.youtube.com/watch?v=tlThdr3O5Qo), but it only works on certain kinds of roads in ideal conditions. It works, but he can't ship it to customers, because the world isn't ideal.
The mathematical part of Machine Learning boils down to "Calculus 3 of multiple variables" and "Being good at translating math to code, and code to math". You remember studying
y=mx+b in high school, that is the beginning of machine learning. To excel in machine learning: make sure to really understand all math themed classes:
Discrete and continuous mathematics,
Calculus I through III and get good at transforming the theory to illustrative code and back again to Math.
The one concept in Calculus you must understand both backwards and forwards (assuming you want to understand neural networks) is how to derive the derivative and integral of an equation with multiple variables. That 'finding of slope of a point of a tangent on a curve' is the operating principle behind machine learning algorithms: the knowing of which way to tease the
m multiple and
b bias in the
y=mx+b, to reduce error from what you have to what you want.
The reason machine learning is good for everyone is because it offloads real work from a human, who would have had to allocate neurons and bio-memory with the expenditure of glucose and time, onto a computer that is able to do the same work but at a tenth of the overall cost.
Member for 2 years, 3 months
10 profile views
Last seen Apr 5 '19 at 18:01