Neural Networks 101
A PM Guide to Neural Networks
Before this, find out our Recent Article -
Imagine you are a Senior Product Manager at Netflix.
You are in charge of the Home Screen.
You have a massive problem: you have 200 million users and 10,000 movies.
How do you decide which 10 movies to show to a specific person in India at 9 PM on a Saturday?
Possible options to solve this problem?
Option 1: Rule-Based Systems (Heuristics)
This is the most basic approach. You write explicit logical rules. If a user watches a Horror movie, show them the most popular Horror movie in your catalogue.
The Problem with this approach is that it is rigid and does not scale. You would need to write millions of rules to cover every user type. It cannot handle discovery. If you only show Horror to Horror fans, they will eventually get bored and churn.
Option 2: Content-Based Filtering
This is a step up where you focus on the attributes of the product. You tag movies with keywords like Oscar Winner, Tom Cruise, or Space. If a user likes Space movies, the system finds other movies with the Space tag.
The Problem with this approach is that it requires massive manual effort. Someone has to tag every single movie accurately.
Option 3: Collaborative Filtering
This is the most common traditional method. It looks at other users.
If User A and User B both liked Movie 1 and Movie 2, and User A then watches Movie 3, the system recommends Movie 3 to User B.
The Problem with this approach is the Cold Start and Sparsity problems. If a movie is brand new and no one has watched it, the system cannot recommend it. If a user is new, the system has no data to compare them to others. It struggles to find complex patterns. It might know two people who like the same genre, but it doesn’t know why.
Now let’s see how Neural Networks can solve the Problem ( But before that let’s understand what is Neural Network?
What is a Neural Network
A neural network is a computational model inspired by the structure of the human brain.
In simple terms, it is a mathematical function that maps a set of inputs to a specific output. A neural network does not need these manual rules. It learns the relationship between variables by looking at millions of examples.
Why do we need Neural Networks?
We need them because human behaviour is multidimensional and non-linear.
High Dimensionality: To predict what a user wants, we might look at 500 different factors like their location, time of day, device, past 50 movies watched, and how long they stayed on a thumbnail. A human cannot write a rule for every combination of these 500 signals.
Non-Linearity: A user might like action movies, but only if they are under two hours long and only on weekdays. Simple models struggle with these complex overlaps, but neural networks excel at them.
I hope you have gotten some jist around the Neural Network but let’s understand Neural Network in slightly more detail.
How Neural Network works?
A neural network consists of three main parts: the input layer, hidden layers, and the output layer.
1. The Inputs Xn
Every piece of data is an input.
x1: User age
x2: Percentage of the last movie watched
x3: Current time
Likewise we haveXn : nth Variable
2. Weights Wn and Bias Bn
Each input is assigned a weight. This represents the importance of that signal. If the model finds that what a user watched 5 minutes ago is more important than their age, that input gets a higher weight.
The bias is an extra constant added to the calculation to allow the model to shift its decision boundary.
The basic calculation inside a single unit or neuron is:
$z = (w1 * x1) + (w2 * x2) + (w3 * x3) + b$
3. Hidden Layers
These are the layers between the input and the final result.
Each layer takes information from the previous one and extracts more complex features. The first layer might look at basic genres. The deeper layers might identify very specific moods or themes that the user prefers but has never explicitly stated.
4. The Output $y$
The final layer gives the prediction. In our Netflix case, it is a number between 0 and 1, representing the probability that the user will click on the movie.
How do we train the Neural Network?
The magic of a neural network is that it corrects itself. This happens in three steps:
Forward Propagation
The model takes the inputs, passes them through the weights and hidden layers, and makes a guess. For example, it predicts an 80% chance that the user will watch a specific documentary.
Loss Function
The model compares its guess to what actually happened.
If the user did not click the documentary, the reality is 0. The difference between the guess (0.8) and reality (0) is the loss. A high loss means the model is performing poorly.
Backpropagation and Optimisation
The model uses a process called backpropagation to go backward from the error to the starting point. It identifies which weights were responsible for the wrong guess and adjusts them slightly.
It uses an algorithm called Gradient Descent to make these adjustments. You can think of this as the model taking small steps to find the lowest possible error.
Challenges with Neural Network?
I know Neural Networks are good but still there are some challenges with our Problem Statement.
The coldstart problem - When a user signs up for Netflix, the model has no "x" variables (inputs) for them. It has no idea what they like. You often need a fallback system (like showing trending movies) until the network has enough data to take over.
Running backpropagation on billions of parameters requires expensive GPUs and massive amounts of electricity. This can cost thousands or even millions of dollars per month.
Key Metrics for a Product Manager
When you manage a product powered by a neural network, you must track these technical metrics:
Precision: Out of all the recommendations the model made, how many were actually clicked?
Recall: Out of all the movies the user eventually watched on the platform, how many did the model correctly identify beforehand?
Inference Latency: How many milliseconds does it take for the network to produce a result? If the model is too deep and slow, the user will see a loading screen and leave.
Training Cost: How much money are we spending on cloud servers to run the backpropagation process?
If you like this article, you will absolutely love our AI Product Management Course ( having real AI PM Interview Questions from Google, OpenAI, Anthropic etc) - ( 35+ Videos ) & ( Extra 25+ Real Case studies as well )
About Author
Shailesh Sharma! I help PMs and business leaders excel in Product, Strategy, and AI using First Principles Thinking. For more, check out my AI Product Management Course, PM Interview Mastery Course, Cracking Strategy, and other Resources
 | A guest post by
|