Here are a summary of all the current issues which I can gather from literature survey:
Data representation:
How to represent the knowledge domain in a way which is susceptible to computer processing?
How to identify the features which is to characterize the knowledge domain so that data can be collected for processing?
How much data is available?
Or how much data is needed to achieve a good prediction or control?
How does the quality of output correlate with an increasing amount of data?
How to generate realistic data?
How to extract out the data from different parts of the system and correlate them together?
Loss function:
How to construct the loss functions? How to conceptualize or realize the loss function is data is discrete, or categorical by nature?
Supervised vs unsupervised learning:
What are the distinguishing factors between supervised and unsupervised learning?
Learning rate:
How to accelerate the learning process?
Learned data reuse:
How to minimize learning via reusing past learning?
Security and Privacy of learning:
How to protect the learned knowledge from being deduced from a certain malformed input, during the process of learning?
How to ensure that the learned knowledge is not easily modified by certain form of inputs?
Real world adaptability and robustness:
How to ensure the learned knowledge is robust against real world errors?
How to enable adaptability to different variation of the original problem?
Creativity, Imagination, Logical thinking:
How to enable creative variation to the original solution?
How to choose the different distribution to represent the data?
How to enable logical step by step creation of solutions?
Learning algorithm:
How to choose the best model to represent the data to be learned?
Is it possible to integrate the different models into a more complex model?
Or vice versa: to simplify existing models into its most basic form so as to reduce computation and resources consumption?
How deep should the network be?
What is the optimal depth to achieve the target for output?
Mathematical problems:
How to overcome the vanishing gradient problem? Saturating non-linearities? Initialization problems?
How to avoid being too dependent on the structure and characteristics of the training data, and be able to adapt to real world test data?
Is it possible to integrate the different models into a more complex model?
Or vice versa: to simplify existing models into its most basic form so as to reduce computation and resources consumption?
How deep should the network be?
What is the optimal depth to achieve the target for output?
Mathematical problems:
How to overcome the vanishing gradient problem? Saturating non-linearities? Initialization problems?
How to avoid being too dependent on the structure and characteristics of the training data, and be able to adapt to real world test data?
References:
No comments:
Post a Comment