Great question—this is a common point of confusion when working with categorical features in neural networks, so let’s break this down clearly.

First: Your "weight update limitation" concern is unfounded #

You’re right that for male samples (encoded as 0), the product gender * weight is 0, but this doesn’t actually restrict the model’s ability to learn. Here’s why:

• Neural networks learn using gradients, and the update rule for a weight w is w = w - learning_rate * dL/dw, where dL/dw is the derivative of the loss function with respect to w. For male samples, dL/dw does end up being 0—but that’s intentional!
• Think about what the weight represents here: if your neuron is output = w * gender + b, then for males (gender=0), the output is just b (the baseline bias). For females (gender=1), it’s b + w—so w directly captures the difference in model output between females and males. The weight only needs to be updated using female samples, because that’s where the gender-specific signal lives. Male samples still contribute to updating the bias b and other feature weights, so the model doesn’t lose any ability to generalize.

0/1 encoding vs 1/2 encoding: Not equivalent, and 0/1 is better #

These two encoding schemes are not interchangeable, and 0/1 (a simplified one-hot encoding for binary variables) is the industry standard for good reason:

• 0/1 encoding preserves categorical meaning: Gender is a nominal variable (no inherent order between male/female), and 0/1 treats each category as a distinct, independent signal. The weight w clearly represents the "female-specific adjustment" relative to the male baseline.
• 1/2 encoding introduces false numerical relationships: Using 1 and 2 implies there’s a linear difference between the two categories (like "2 is twice as much as 1"), but gender has no such ordered relationship. This can confuse the model—instead of learning a simple categorical difference, it might waste capacity trying to interpret the arbitrary numerical gap between 1 and 2, which has no real-world meaning.

Why most code uses 0-starting encoding? #

It’s not a cognitive mistake—it’s rooted in how categorical variables are best represented for machine learning:

• One-hot encoding (which 0/1 is for binary cases) is the gold standard for nominal variables because it avoids imposing false order. For multi-class variables, you’d use a vector like [1,0,0] for class A, [0,1,0] for class B, etc.—0/1 is just the binary version of this.
• Starting at 0 also aligns with how many programming languages handle indexes, making it easier to map categories to arrays or tensors.

Final takeaway #

Stick with 0/1 encoding for your gender feature. It’s intuitive, mathematically sound, and ensures your model learns the correct categorical relationship without unnecessary complexity. The 1/2 encoding approach doesn’t solve any real problem and introduces avoidable ambiguity.

内容的提问来源于stack exchange，提问作者edn