From 8d615866beb266e169061aa1e9733efe3de2618a Mon Sep 17 00:00:00 2001
From: Atharva Sehgal Symbolic Labelling: Concaten
The symbolic labelling process is repeated for each entity in the scene. The resultant symbolic - vectors are stacked to form a symbolic embedding of the scene. + vectors are stacked to form a symbolic encoding of the scene. +
++ Following NPS, we will break down the transition function into two parts: Learning to select a module and learning to specialize the module to a task. However, how should we employ the neural and symbolic encodings? +
++ The symbolic encoding will help the selection module be robust to attribute compositions. However, if we just use the symbolic encoding, we will risk bottlenecking the model's ability to learn fine-grained dynamics-relevant attributes that may not be known ahead of time. +
++ The neural encoding, on the other hand, captures rich dynamics-relevant attributes which will enable good reconstruction. However, we will risk overfitting to attribute compositions seen during training. +
++ We solve this problem by employing a hybrid approach. We'll use the symbolic encoding to select a module and the neural encoding to predict the next state. +
++ We'll now describe the rest of the architecture.
- The symbolic embedding is concatenated with the action vector after the embedding and action are + The symbolic encoding is concatenated with the action vector after the encoding and action are reordered to match the canonical ordering of the symbols. The concatenated vector is select a learnable module which is used to predict the next state.
- Note that the symbolic embedding is - only selecting the module, which leaves room for the neural embedding to learn fine-grained + Note that the symbolic encoding is + only selecting the module, which leaves room for the neural encoding to learn fine-grained dynamics-relevant attributes that may not be known ahead of time.
- The selected module is applied to the neural embedding to predict the next state. + The selected module is applied to the neural encoding to predict the next state.