Let's assume Neural Network Classifier X->model->embedding vector-4 scores ->saftmax -> APD of 4 Taking a closer look at the process of obtaining scores embedding vector-4 scores din (2) Task 1: Instead of calculating the scores based on dot product, we want to do so based on the Euclidean distance between e vector and Wi. How will you do it? Task 2: dot =]] [1] pridely di (mance) dit ) Currently sl = dot(wl,e vector) a score Si is calculated based on the dot product with one weight vector Wi. This means the model needs to learn a representative embedding for class I (WI). We now want the model to be able to leam representative embeddings per class (Wil, Wi2, Wis), but still give 4 scores, one foe each class. How will you do it? Embedding vector 4 Scores W,T S, Will Wy ] Sg w dimension (wa) - Limension (e) o. Si - dot (wi,) Task wil dot Wel Wy products 2 E! Sy dimension (wit) - dimension (e) So = dot (wise)