Generative artificial intelligence to pre-train and fine-tune models for multiple autonomous vehicle future trajectories

An apparatus comprising:

• one or more hardware processors to:
• train a first generative artificial intelligence (AI) model on a first set of training data;
• provide a second generative AI model with at least one distillation of the first generative AI model, the at least one distillation comprising transformer weights, embeddings, or predictions labels of the first generative AI model;
• train the second generative AI model that is initialized with the at least one distillation of the first generative AI model, wherein the second generative AI model is trained using a second set of training data that is smaller than the first set of training data;
• deploy the second generative AI model to an autonomous vehicle;
• generating a prediction using student generative AI model;
• and
• causing the autonomous vehicle to maneuver using the prediction from the student generative AI model,
• wherein the first generative AI model comprises an encoder-decoder transformer comprising a first transformer encoder and an autoregressive transformer decoder, and further comprises a one-shot decoder, wherein the autoregressive transformer decoder outputs a sequence of predictions using an autoregressive model, and
• wherein the one or more hardware processors to provide the second generative AI model with the at least one distillation further comprises the one or more hardware processors to:
• provide the transformer weights of the first generative AI model to the second generative AI model by:
• providing first weights of the first transformer encoder as initial transformer encoder weights to a second transformer encoder of the second generative AI model;
• and
• providing second weights of the one-shot decoder as initial decoder weights to a second decoder of the second generative AI model;
• and
• train the second generative AI model using the initial transformer encoder weights and the initial decoder weights.

A non-transitory computer-readable medium having stored thereon instructions that, when executed by one or more processors, cause the one or more processors to:

• train a teacher generative artificial intelligence (AI) model on a first set of training data;
• provide a student generative AI model with at least one distillation of the teacher generative AI model, the at least one distillation comprising transformer weights, embeddings, or predictions labels of the teacher generative AI model;
• train the student generative AI model that is initialized with the at least one distillation of the teacher generative AI model, wherein the student generative AI model is trained using a second set of training data that is smaller than the first set of training data;
• deploy the student generative AI model to an autonomous vehicle;
• generate a prediction using student generative AI model;
• and
• cause the autonomous vehicle to maneuver using the prediction from the student generative AI model,
• wherein the teacher generative AI model comprises an encoder-decoder transformer comprising a transformer encoder and an autoregressive transformer decoder, and further comprises a one-shot decoder, wherein the autoregressive transformer decoder outputs a sequence of predictions using an autoregressive model, and
• wherein the one or more hardware processors to provide the student generative AI model with the at least one distillation further comprises the one or more hardware processors to:
• provide the transformer weights of the teacher generative AI model to the student generative AI model by:
• providing first weights of the transformer encoder as initial transformer encoder weights to a student transformer encoder of the student generative AI model;
• and
• providing second weights of the one-shot decoder as initial decoder weights to a student decoder of the student generative AI model;
• and
• train the student generative AI model using the initial transformer encoder weights and the initial decoder weights.