Davide Nunes
Defining meaning
Symbol grounding through representation learning
Techniques from computational semantics
Deep learning and neural networks
Generative probabilistic models
Hiperdimensional computing and random projections
To enable machines to understand the world.
Formal definition of names, classes, properties, and relationships between entities in a given domain.
No motivation nor reward to adopt them
They require ontology engineering skills
Merging ontologies requires consensus
They describe very little about the world
Grounded on humans, not on actual data
"Speech has both an individual and a social side, and we cannot conceive of one without the other" — Ferdinand De Saussure
The meaning of a sign is its instantiation in terms of neural activity patterns in response to the stimuli that constitutes this sign, along with the state of the world that makes up its context.
$$ similarity(x,y) = cos(\theta) = \frac{x \cdot y}{\lVert x \rVert \lVert y \rVert} $$
Assumption that latent structure in text can be captured by linear convex combinations of linearly independent vectors
Difficult to integrate new data (no on-line learning)
No extrapolation for missing data (no smoothing)
$$ \forall i,j \\ (1-\varepsilon) \| x_i - x_j \|^2 \le \\ \| f(x_i)-f(x_j)\|^2 \\ \le (1+\varepsilon)\|x_i - x_j\|^2 $$