I believe it's more proper, clear coding to make a model for your input, thereby making your code readable to anyone, whether they know the brain.js software or not. The abstraction modeled here is that a neural network gets an input node, and yields an output node. This is also how the parameters for the train functino are set up.

The problem is probably in the test code, and not in your code. However, it probably is caused by your code. You need to check all your brackets and braces and see if they match.

Change grammar:
"You are given an array of ages of a group of people. Find the youngest person in the list."

Does 'you did good' mean you actually found the answer, or does it mean the chance you have found an answer is increased? In other words, is the data set ambiguous and do we need to find the most likely candidate function based on statistics?

Passing the tests does not equal 'working'. Your solution is not a Huffman encoding, and it not useful in real life compression. Still, nice. My first own solution was also lacking, and yet it passed the tests. I think the tests for this kata can definitely be improved.

frequesncies?? replace went wrong?

If you omit semicolons in order to get a smaller footprint you might as well remove all whitespace and add semicolons.

toString and parseInt can convert to and from binary

I really really stumbled until I checked the comments and saw people saying "don't modify the input". I think this should really be required text in the kata description!

it's spelled 'symmetric'

nice utility functions

Yes, clearly.

it's spelled 'hundred'