TrAct: Making First-layer Pre-Activations Trainable | Read Paper on Bytez