Deep Learning from scratch, a Journey of Discovery

Today, I started a garden, a garden foray unlike any other, into the roots of deep learning. My tools were that of curiosity and determination, and I used them to excavate the mechanics of chaining functions, dissecting Numpy arrays, and refining my python development skills. This was not simply a technical exercise-it was an opportunity to refine my understanding of how neural networks work from the roots and up! I will share, in the upcoming sections, key lessons and insights I gained that can inspire anyone delving into deep learning and python programming 🤓

🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀

Blast off! Here we go! To the realm of deep learning!

One of the most profound (and coolest) concepts I practiced today was the idea of chaining functions-executing multiple operations on a dataset sequentially. This directly mirrors how data flows through layers of a neural network, with each function using a composite (or transforming) the input before it passes to the next function.

Here is the optimized version of a function chaining pipeline I crafted:

from typing import Callable, List
import numpy as np
from numpy import ndarray

#Define a type alias for a function that takes and returns an ndarray

Array_function = Callable[[ndarray], ndarray]

#function to chain

def chain_any_length(chain: List[Array_function], a: ndarray) -> ndarray:
'''
evaluates all functions in a chain, sequentially
'''

for func in chain:
a = func(a)
return a

The true power of this function lies in its flexibility. Using this flexibility, I then created a chain of trigonometric transformations (sin, cos, and tan) and then applied it to a numpy array:

a = np.array([1, 2, 3])
result = chain_any_length([np.sin, np.cos, np.tan], a)
print(result)

#the output in my notebook is: [0.78635739 0.70533911 1.52387302]

I then used the data wrangler in vs code, an awesome feature for debugging arrays in tabular form!

The journey waas not just about Python, but about undeerstanding data flow through a network. Each function in the chain mirrors a corresponding layer in a neural network:

np.sin as the activation function,
np.cos as a normalization layer,
np.tan as an additional transformation

By chaining these operations, I could see how intermediate transformations shape the final output. This lays the foundation for back-propogation and other concepts.

Conclusion:

Today was a reminder that true mastery comes from love of the subject, curiosity, experimentation, and most of all, determination. There are days when I don’t want to do this, but after I start feeling my fingers on the keyboard, typing away, I feel an intense surge of curiosity and love engulf me. No matter what, never give up, and you will find your Sadhana, your calling. This is mine.