How do we test mental representation theories ? (not as important)
- Psychological plausibility
- Neurological plausibility
- practical applicability
- Representational power
- Computational power
What is meant by Representational power ?
- How much information a mental representation can express.
- What do the mental representations express?
What is meant by Computational power ?
- Basically, what can you do with the mental representation in terms of high lvl thinking
- Can it be used for problem solving, (including planing decison making and explanation), learning and the use of langaug
- how efficient is ur computational procedure acting on the mental representation
What is meant by practical applicability ? (not as important)
- Can the mental representation theory be used in fields of education and artifical intelligence
What is meant by psychological plausibility ? (not as important)
- can the mental representation explain human thinking / undertsanding human cognition
What is meant by neurological plausibility ? (not as important)
- Is the mental representation theory consistent with the results of neuroscientific experiments
How do machine learn ?
- Via an optimization algorithm
What does the optimization algorithm consist of ?
- labeld data (numbers)
2. tunable parameters
How to find the best possible output ?
- Via finding the perfect combination of parameters between all layers
Ho do we find out the best possible parameters combination ?
- Via randomly trying (bad)
- Via the gradient method
- Both try to find the local minimum of the coast function
What does deep learning mean ?
- each layer can use the information extracted in the previous layer to build up a more complex representation of the data
What is part of a parameter ?
- A parameter consist of weights and a Bias
How does the gradient technique work ?
- You take a randome parameter
- Coast function tells u the parameter is incorrect
- the coast function reduces the error via changing the bias and its weights
How does the coast function work?
- The lower the coast function the better the parameter
- The higher the coast function the worst the parameter
- It is the way to say that the network makes an error
How does the coast function now in which direction the weight should shift in order to become a better output ? (2 Dimensional)
- Identfying the slope of the randome used parameter
- shift to the left if slope is positive and shift the input to the right if the slope is negative
- Right = increase the parameter
- left = decrease the parameter
- If you do this over and over again you will approach a local minimum
How does the coast function now in which direction the weight should shift in order to become a better output ? (3 Dimensional)
- identify the gradient of a function
- Make the gradient negative which will show you in which direction you will have to move your parameter
- Do this over and over again to adjust the parameters (gradient descent)
- Identify the local minimum of the coast function
Why do we need to use the negative gradient in order to identfy the local minimum of the coast function ?
- The normal gradient = steepest highest increase
- The negative gradient shows decrease most quick
What is meant by gradient descent ?
- Repeated process of following the negative gradient to adjust parameters
- Also identifies which weights matter more then other regarding the output
What are the different approaches to mental representation ?
- logic, rules, concepts, images, connections (artificial neural networks)
