# Interactive Cognition (9/14/2016)

## Elements of Interactive Cognition (Perception):

1. Environment
2. Sequential Parsing
3. Prediction
4. Interaction with other systems
5. Recurrent (connections)
6. Memory
7. Parallel/Distributed Processing: not clear whether this is for computational advantages or for some other reasons

Formal Models:

1. Environment:
• Labeled examples: a sequence of (X1, l(x1)), … (Xn, l(xn))
• Two models:
(a) PAC (probably approximately correct):
Assume $\exists$ distribution D (unknown) on input examples $x \in \mathbb{X}$. After seeing m labeled examples (x, l(x)) where x ~ D, a $(\epsilon, \delta)$-PAC algorithm finds a hypothesis h with prob. $1-\delta$ such that: $P_{x \sim D}(h(y) \ne l(y)) \leq \epsilon$
(on a new input y ~ D, we should be able to guess l(y))
(b) Mistake-bound:
Given an example x, you predict l(x) right away, and are then told the correct answer. The complexity of an algorithm in the mistake-bound model is m, if for any sequence of examples (of any length), the total number of mistakes made by the algorithm is at most m.
2. Prediction:
1. Max Likelihood Estimators: find a model that most likely generated the data
2. Bayesian Estimation: use priors to make decisions
3. Recurrent (connections):
• Recurrent Neural Network (RNN): contains connections that are not feed-forward:

## Project Example:

Invariant: Parsing a letter

1. Environment and Sequential Parsing: I see a letter as a sequence of strokes
2. Prediction: As someone starts writing a stroke, you can predict which letter it is
3. Interaction with other systems:
• Motor System
• Sound
4. Recurrent: Use feedback from “learned” letters to predict
5. Parallel/Distributed Processing: