8: Saturday group stuff Flashcards
What is semantics, in the context of reasoning and deductive logic?
Reasoning in natural language rather than a formal logical system.
What is a causal relationship and a diagnostic relationship?
A causal relationship between two events is one in which the outcome of one event influences the outcome of the other. A diagnostic relationship is in the reverse direction of a causal relationship; it is diagnostic in that knowing the second event happens makes the first one more likely.
What is the minimum number of probabilities given in a joint probability distribution?
The number of all possible probabilities (for outcomes), i.e. the number of cells, -1.
What is a logical model?
A replication of reality containing a group of situations in which certain outcomes are true. Situations are objects that exist and their interpretations (the states they could be in).
What is entailment?
When one condition is true if and only if another is. A entails B if B is true if and only if A is true.
What is the RRT* algorithm?
An extension to and optimisation of the RRT algorithm in which rather than returning the first possible solution (path) a more optimal one is sought and returned instead if possible and found.
What does it mean for a heuristic to be admissible?
The heuristic never overestimates the cost value (always less than or equal to).
What is the ravine problem?
When seeking the lowest point and an AI descends into a valley-shaped (ravine) feature in the 3d weight error space, momentum means you will zig-zag up and down each edge either side instead of following the floor to reach the goal. This reduces optimality and efficiency but can be fixed by adding weight decay.
What is the generalisation problem?
When a neural network becomes too specialised to its training data so has a lower performance for testing/validation (and real-world) data. Fix with early stopping: run on training once and then on training data, find the point in training at which performance on training is best and error is minimum. Then rerun the training but stop early when at the minimum point found.
What is a recurrent neural net?
A neural network in which nodes can be connected to themselves (output of a neuron goes to a part of or all of its input). This facilitates memory and context in the network to some degree.
What is a convolutional neural net?
Neural networks that split up the input and then process the individual parts before recombining to find the final result (divide and conquer). Often used in computer vision applications. Uses ReLu (rectified linear unit) activation function as sigmoid takes too long.
