In the context of neural networks, hidden states are the intermediate layers' activations that carry information from one part of the network to another. When these hidden states are described as "expressive," it means they have the capacity to represent a rich and nuanced range of features and dependencies from the input data. This expressiveness is crucial in tasks such as language modeling, machine translation, and time-series prediction, where the model needs to understand and predict based on intricate and long-range dependencies in the input sequences. The expressiveness of hidden states can be enhanced by architectures like LSTMs (Long Short-Term Memory networks) and GRUs (Gated Recurrent Units), which are designed to maintain and manipulate these hidden states over long sequences.
The concept of hidden states in neural networks dates back to the development of early neural network models in the 1980s. However, the specific notion of "expressive hidden states" gained prominence in the late 1990s and early 2000s with the advent of more sophisticated RNN architectures like LSTMs (1997) and GRUs (2014), which were explicitly designed to address the limitations of standard RNNs in maintaining expressive and long-range dependencies.
Key figures in the development of expressive hidden states include Sepp Hochreiter and Jürgen Schmidhuber, who introduced the LSTM architecture in 1997, and Kyunghyun Cho and colleagues, who proposed the GRU in 2014. These architectures have been instrumental in enhancing the expressiveness of hidden states in neural networks, enabling significant advancements in various AI applications.