A Novel Framework for Learning Stochastic Representations for Sequence Generation and Recognition
Hwang, Ahmadi
The ability to generate and recognize sequential data is fundamental for autonomous systems operating in dynamic environments. Inspired by the key principles of the brain-predictive coding and the Bayesian brain-we propose a novel stochastic Recurrent Neural Network with Parametric Biases (RNNPB). The proposed model incorporates stochasticity into the latent space using the reparameterization trick used in variational autoencoders. This approach enables the model to learn probabilistic representations of multidimensional sequences, capturing uncertainty and enhancing robustness against overfitting. We tested the proposed model on a robotic motion dataset to assess its performance in generating and recognizing temporal patterns. The experimental results showed that the stochastic RNNPB model outperformed its deterministic counterpart in generating and recognizing motion sequences. The results highlighted the proposed model's capability to quantify and adjust uncertainty during both learning and inference. The stochasticity resulted in a continuous latent space representation, facilitating stable motion generation and enhanced generalization when recognizing novel sequences. Our approach provides a biologically inspired framework for modeling temporal patterns and advances the development of robust and adaptable systems in artificial intelligence and robotics.
본 논문은 수열 생성 및 인식을 위한 새로운 확률적 순환신경망 매개변수 편향(stochastic RNNPB) 프레임워크를 제안한다. 본 모델은 뇌의 예측 부호화 및 베이지안 뇌 가설에서 영감을 받아, 변분 자동인코더의 재매개변수화 기법을 통해 잠재 공간에 확률성을 도입한다. 실험 결과는 확률적 RNNPB 모델이 로봇 운동 수열의 생성 및 인식 작업에서 결정론적 모델을 크게 능가하며, 학습 및 추론 과정에서 불확실성을 정량화하고 조정할 수 있고, 연속적인 잠재 공간 표현을 형성하여 안정적인 운동 생성 및 향상된 일반화 능력을 촉진함을 보여준다.